Search

Today, malaria is a dangerous disease caused by Plasmodium parasites and transmitted by the bite of Anopheles mosquitoes. It is estimated that more than 200,000 cases of malaria will occur every day worldwide, and more than 400,000 people will die from malaria. Early diagnosis of malaria is important to mitigate the morbidity and mortality of the disease. In this paper, a fully trained capsule neural network combined with a convolutional neural network model is proposed for malaria detection from blood smear microscopic images. In malaria diagnosis, the hybrid CapsNet model can detect malaria in blood samples. The hybrid model is optimized based on malaria data in microscopic images of infected and uninfected blood. Images are processed and enhanced through rotation before being fed into the hybrid CapsNet model. The CapsNet hybrid model is examined for malaria data with a learning rate of 0.07 and a batch size of 20. Hybrid CapsNet model shows improved results in detecting malaria from microscopic images measured against the currently available standard malaria test. Constraints like malaria parasite detection rate, accuracy, F1 index and blood smear classification is used to gauge the attainment of the hybrid CapsNet model. Detection rate of 99%, accuracy of 99.08% and FAR of 0.97% is attained by the hybrid CapsNet model as contrary to the traditional state-of-the-art model, the DSCN-Net, where the detection rate was 98%, the accuracy was 97.2%, and the FAR was 0.99%. [ABSTRACT FROM AUTHOR]

Facial expressions are a critical form of nonverbal communication, conveying a wide range of emotions. Recent advancements in artificial intelligence and computer vision have led to the development of deep learning methods, particularly convolutional neural networks, which are highly effective in facial emotion recognition. This paper presents EmoNeXt, an advanced deep learning framework for FER that builds upon a modified ConvNeXt architecture and incorporates several key innovations. EmoNeXt integrates spatial transformer networks to enable the model to focus on the most expressive regions of the face, squeeze-and-excitation blocks to enhance channel dependencies, and a self-attention regularization term that encourages the learning of compact and discriminative feature vectors. Initially evaluated on the FER2013 dataset, EmoNeXt is now further validated on two other widely used benchmark datasets, AffectNet and CK+, to demonstrate its robustness and generalizability across various real-world and posed scenarios. Additionally, we conduct an extensive ablation study to analyze and quantify the contribution of each enhancement, confirming their positive impact on model performance. Finally, this paper explores the application of EmoNeXt in emotion recognition for elderly individuals with Alzheimer's disease, highlighting the urgent need for accurate emotion recognition to improve patient care. Our results underscore the potential of EmoNeXt as a valuable tool for enhancing emotional communication in healthcare settings, particularly for patients with neurodegenerative disorders. [ABSTRACT FROM AUTHOR]

One of the most common cancers among women worldwide is breast cancer (BC), and early diagnosis can save lives. Early detection of BC increases the likelihood of a successful outcome by enabling treatment to start sooner. Even in areas without access to a specialist physician, machine learning (ML) aids in early BC detection. The medical imaging community is becoming more interested in using ML, and deep learning (DL) to increase the accuracy of cancer screening. Many disease-related data are sparse. However, for DL models to perform well, a large amount of data is required. Because of this, the DL models that are currently in use on medical images are not as effective as they could be. Convolutional neural network (CNN) models have recently gained popularity in the medical industry, and they perform admirably in terms of high performance and robustness at image classification. The proposed method classifies data using ensemble pre-trained models such as the dense convolutional network (DenseNet)-121 and EfficientNet-B5 feature extractor networks, as well as the support vector machine for classification. Using a modified meta-heuristic optimizer, the selected pre-trained CNN hyperparameters were optimized to improve the performance. The experimental results for the presented model on the INbreast dataset show that the EfficientNet-B5 model is effective for BC classification, with overall accuracy, sensitivity, specificity, precision, and area under the ROC curve (AUC) values of 99.9%, 99.9%, 99.8%, 99.1%, 1.0, respectively. [ABSTRACT FROM AUTHOR]

Breast cancer prognosis prediction is pivotal to improving the survival chances of breast cancer patients. Recently, deep learning models integrated with multi-modal data have been explored extensively to improve the reliability and accuracy of breast cancer prognosis prediction. However, multi-modal data poses challenges to feature extraction due to the varied distribution and dimensionality across the modalities. To address this, we developed a multi-input convolutional neural network model, which extracts features from each modality in the multi-modal dataset simultaneously using separate convolutional layers, then concatenates and passes them to shared dense layers. The proposed model achieved area under the receiver operating characteristic curve values of 0.893 and 0.865 in 5-fold cross-validation and unseen test data respectively (at threshold = 0.2). These outcomes surpassed those of single-input convolutional neural network models and a state-of-the-art method based on multi-modal data. The multi-input convolutional neural network model efficiently handles multi-modal data and is a promising tool for breast cancer prognosis prediction. [ABSTRACT FROM AUTHOR]

Biological vision mechanism plays a very important role in detecting object contour information, so the research of bio-inspired contour detection has attracted widespread attention. In the visual physiological mechanism of contour detection, complex cells in the primary visual cortex (V1) receive the target edges output by simple cells, and subsequently aggregate and integrate the edges into contours. Inspired by the receptive field properties of complex cells in the V1 area, we design a contour detection network with an encoder-decoder structure, called the clustered suppression network. In the encoding network, we propose a feature aggregation encoding framework inspired by the clustered distribution properties of complex cells. At the same time, based on the inhibition mechanism properties of the non-classical receptive field of complex cells and the response properties to motion edges, we propose a suppression module and a spatial motion detection module, which are integrated into the encoding network. In the decoding network, we propose a highly aggregated decoding network inspired by the horizontal connection properties of complex cells. Choosing the BSDS500 natural scene dataset as the experimental object, the F-score is selected as the evaluation index. The average optimal F-score on a single-scale of the proposed method is 0.802. Concurrently, the NYUD-v2 dataset and BIPED dataset are used for further verification and achieved good results. The codes are available at https://github.com/smallsunq/CSNet-Contour-detection. [ABSTRACT FROM AUTHOR]

The cross-border intelligent marketing algorithm based on traditional linear models is relatively single in information feature extraction, making it difficult to effectively handle complex scenarios containing a large amount of implicit information in users and markets, resulting in poor personalized marketing effectiveness. To address this issue, this article proposes a cross-border intelligent marketing model that integrates rating information and user labels using a multi-layer perceptron grey wolf optimization and convolutional neural network (MLP-GWO-CNN). This model extracts implicit high-order information through nonlinear methods and can handle complex and sparse marketing data. Firstly, a dual path deep network structure was designed, in which one path was modeled using a multi-layer perceptron (MLP) to extract user interest features based on historical interaction ratings; Another path utilizes Convolutional Neural Networks (CNN) to extract semantic features from user label information and construct item feature representations. In response to the sensitivity of MLP algorithm to initial values and its tendency to fall into local optima, this paper uses GWO algorithm to optimize MLP. Next, the latent feature vectors generated by MLP and CNN are fused in the output layer to generate the final predictive marketing strategy last. Experiments were conducted using a real cross-border e-commerce dataset, and the results showed that compared with traditional recommendation algorithms, the MLP-GWO-CNN model proposed in this paper performs better in utilizing user tag information, effectively improving the accuracy and personalization of marketing recommendations. The accuracy of the model is over 89%, and the recall rate is over 90%. [ABSTRACT FROM AUTHOR]

In recent decades, brain tumors have been regarded as a severe illness that causes significant damage to the health of the individual, and finally it results to death. Hence, the Brain Tumor Segmentation and Classification (BTSC) has gained more attention among researcher communities. BTSC is the process of finding brain tumor tissues and classifying the tissues based on the tumor types. Manual tumor segmentation from is prone to error and a time-consuming task. A precise and fast BTSC model is developed in this manuscript based on a transfer learning-based Convolutional Neural Networks (CNN) model. The utilization of a variant of CNN is because of its superiority in distinct tasks. In the initial phase, the Magnetic Resonance Imaging (MRI) brain images are acquired from the Brain Tumor Image Segmentation Challenge (BRATS) 2019, 2020 and 2021 databases. Then the image augmentation is performed on the gathered images by using zoom-in, rotation, zoom-out, flipping, scaling, and shifting methods that effectively reduce overfitting issues in the classification model. The augmented images are segmented using the layers of the Visual-Geometry-Group (VGG-19) model. Then feature extraction using An Attribute Aware Attention (AWA) methodology is carried out on the segmented images following the segmentation block in the VGG-19 model. The crucial features are then selected using the attribute category reciprocal attention phase. These features are inputted to the Model Agnostic Concept Extractor (MACE) to generate the relevance score between the features for assisting in the final classification process. The obtained relevance scores from the MACE are provided to the max-pooling layer of the VGG-19 model. Then, the final classified output is obtained from the modified VGG-19 architecture. The implemented Relevance score with the AWA-based VGG-19 model is used to classify the tumor as the whole tumor, enhanced tumor, and tumor core. In the classification section, the proposed Relevance score with the AWA-based VGG-19 model has achieved 96.20%, 98.20%, and 99.40% enhanced accuracy on the BRATS 2019, 2020, and 2021 databases. In the segmentation section, the Relevance score with the AWA-based VGG-19 model has achieved 91.06%, 91.40%, and 94.86% higher Dice Similarity Coefficient (DSC) value, where the attained results are superior to the traditional CNN model. Validation regarding accuracy, specificity, and sensitivity is carried out to verify the classification performance, and the analysis in terms of DSC and HD is done to validate the segmentation performance. The performance of the implemented BTSC model is validated by comparing it against methods such as MAC-ARUNet3+, BTSC-TNAS, Mask R-CNN Resnet 101, CNN, SwinBTS, Residual mobile U-net, and Ensemble deep learning. [ABSTRACT FROM AUTHOR]

This research investigates the design of a nanofiltration (NF) system based on a nanocomposite membrane containing graphene oxide (GO) for the demineralization of sweet whey and the modeling the NF process. The effects of various process variables, including, transmembrane pressure (TMP), Reynolds number, feed pH, and temperature, on the rejection of the minerals were surveyed. Consequently, the recovery of whey proteins from industrial whey using fabricated membranes in a cross-flow membrane module was also investigated. Among the input variables, the pH of the whey solution has the greatest effect on membrane flux and salt rejection performance. In the dead-end filtration system, the highest flux was achieved for the GO-modified membrane under laboratory conditions with a pressure of 10 bar and a pH of 6. The dynamic flux behavior of whey output and salt rejection from whey was modeled using convolutional neural network (CNN) machine learning tools. Linear and non-linear correlations demonstrated that the CNN model correlates well with experimental data on dynamic flux (R2–1.00). Overall, this study on dynamic flux and rejection of minerals from whey using CNN modeling can improve optimal conditions for whey demineralization and reduce laboratory testing costs by predicting the results of untested experimental variables. [ABSTRACT FROM AUTHOR]

This study explored the potential of building an image-based quality control system for particleboard manufacturing. Single-layer particleboards were manufactured under 27 operating conditions and their modulus of elasticity (MOE) and the modulus of rupture (MOR) were determined. Subsequently, images of the upper surface, lower surface, and cross-section of each specimen were collected. Two types of convolutional neural networks (CNNs) were designed: a single-input CNN processing one image and a multi-input CNN capable of analyzing multiple images simultaneously. Their prediction accuracies were then compared. Among the single-input CNNs, the cross-sectional image yielded the best prediction accuracy for both the MOE and MOR. For multi-input CNNs, the combination of the upper surface and cross-sectional images produced the highest scores when the model merged the information from each image at early stage, outperforming single-input CNNs. Adding density information to multi-input CNNs significantly improved prediction accuracy for both MOE and MOR, achieving optimal results. Regression activation maps were constructed to visualize the image features that were strongly correlated with the predicted results. For MOE prediction, the precise location of phenol formaldehyde (PF) resin and particle alignment were crucial. For MOR prediction, the interface between particles and PF resin was the key. [ABSTRACT FROM AUTHOR]

With the increase of Chinese emphasis on calligraphy education, more and more students choose to learn calligraphy courses in colleges. However, the present process of calligraphy education lacks enough content of appreciation and learning of stele characters of ancient calligraphers. To reduce the difficulty of carrying out the content of stele calligraphy education in colleges and universities, this study constructed a recognition model of ancient Chinese characters in stele. The recognition model will filter and denoise the input image data, open and close calculation, binarization, data amplification and skeletonization. It feeds into two designed convolutional neural networks to extract graphic features. The test results of recognition models showed that the recognition accuracy and recall rates of SCF, Gabor, Fast R-CNN and S-ICNN models on the test set were 79.13%, 74.52%, 82.66%, 93.87% and 76.94%, 72.41%, 88.25%, 94.09%, respectively. The data showed that the recognition accuracy of the ancient Chinese character recognition model designed by this research is significantly higher than the common methods in the market. It has certain application potential, but further research is needed to realize the purpose of applying it in college calligraphy education. [ABSTRACT FROM AUTHOR]

In recent years, numerous countries have initiated geochemical survey projects, highlighting the importance of identifying geochemical anomalies for the discovery of potential mineral deposits. In addition, anthropogenic activity, missing or inaccurate data, and overburden can lead to local enrichment or deficiency of elements, resulting in false or weak geochemical anomalies. Simultaneously considering spatial and spectrum information in the data can eliminate spectrum differences caused by the data inaccuracy and enhance weak mineralization anomalies. Therefore, introducing spatial-spectrum models is beneficial for leveraging the strengths of both approaches. This study proposes a two-branch fusion network for extracting spatial-spectrum features from a geochemical survey data cube. The spectrum branch consists of a one-dimensional convolutional neural network (1DCNN) that can be utilized to extract geochemical spectrum information within a single pixel, covering major and trace geochemical elements and accounting for both positive and negative geochemical anomalies. The spatial branch is a superpixel graph convolutional network (SGCN), which is composed of internal and external graph convolutions. The SGCN not only can extract spatial relationships between neighboring pixels and even pixels at a long distance, but also takes into account the anisotropy of mineralization. Furthermore, spatial information can smooth out false geochemical anomalies caused by inaccurate or missing data. A case study was conducted to identify mineralization-related geochemical anomalies and validate the proposed hybrid deep learning model in northwestern Hubei Province, China. Experiments have shown that (1) superpixel segmentation is an effective tool for geochemical anomalies identification, (2) the incorporation of spectrum- and spatial-based methods contributes to the model's ability to discern anomalies and backgrounds within the geochemical data cube, improving its accuracy in anomaly detection, and (3) the identified anomalous areas provide clues for future mineralization searches. [ABSTRACT FROM AUTHOR]

Mineral identification and discrimination play a significant role in geological study. Intelligent mineral discrimination based on deep learning has the advantages of automation, low cost, less time consuming and low error rate. In this article, characteristics of quartz, biotite and K-feldspar from granite thin sections under cross-polarized light were studied for mineral images intelligent classification by Inception-v3 deep learning convolutional neural network (CNN), and transfer learning method. Dynamic images from multi-angles were employed to enhance the accuracy and reproducibility in the process of mineral discrimination. Test results show that the average discrimination accuracies of quartz, biotite and K-feldspar are 100.00%, 96.88% and 90.63%. Results of this study prove the feasibility and reliability of the application of convolution neural network in mineral images classification. This study could have a significant impact in explorations of complicated mineral intelligent discrimination using deep learning methods and it will provide a new perspective for the development of more professional and practical mineral intelligent discrimination tools. [ABSTRACT FROM AUTHOR]

Edge and contour detection plays critical roles in computer vision and image processing, with extensive applications in advanced tasks including object recognition, shape matching, visual saliency, image segmentation, and inpainting. In recent decades, this field has attracted significant attention, leading to the development of numerous sophisticated methods that approximate human visual performance. Despite these advances, notable gaps remain. This review offers a comprehensive analysis of representative techniques, categorizing them into traditional and learning-based approaches, and examines their strengths and limitations to identify the underlying reasons for these gaps. Traditional methods are further divided into four sub-categories: local pattern, edge grouping, active contour, and bio-inspired techniques, with a specific emphasis on the promising potential of bio-inspired methods. Learning-based approaches, on the other hand, are classified into two types: classical learning, which typically relies on handcrafted features designed from empirical knowledge, and deep learning, which autonomously extracts features from large-scale datasets without human intervention. Additionally, benchmarks and evaluation metrics related to edge and contour detection are discussed, with potential issues identified within these frameworks. A quantitative assessment of representative methods is conducted across three popular benchmarks. Lastly, challenges and future prospects in edge and contour detection are explored, focusing on five key aspects: model architecture, learning strategies, feature extraction and fusion, method integration, and cross-domain applications. These considerations aim to bridge the gap with human visual perception. Overall, this work is expected to benefit researchers and advance progress in the field. [ABSTRACT FROM AUTHOR]

Voice activity detection (VAD) is an important component of signal processing that is critical for various applications, including speech recognition, speaker recognition, and speaker identification for example to eliminate different background noise signals. With the increasing use of deep learning techniques in speech-based applications, VAD has become more accurate and efficient. In this paper, a novel VAD method based on deep neural networks, specifically, a residual neural network (ResNet), a convolutional neural network with multilayer perceptron, and recurrent neural network components are proposed. The results of the proposed VAD are compared with other state-of-the-art methods, and its effectiveness is demonstrated in basic speech enhancement in the presence of various types of noise. The proposed VAD utilizes a combination of time–frequency features, such as log mel spectrogram representations, and the ResNet-32 as a flexible classifier method to detect speech/no-speech activities. The advantages of the proposed VAD method include its ability to adapt to different types of background noise, such as stationary, non-stationary, and periodic noises, and its flexibility in terms of selecting appropriate deep learning models for different applications. The results of the proposed VAD method show significant improvements compared to other methods, which demonstrates its effectiveness in speech enhancement in noisy environments. It is shown that the best performance is achieved by a hybrid of time–frequency features as log mel spectrogram representations and ResNet-32 as a flexible classifier method to detect speech/no-speech activities correctly. [ABSTRACT FROM AUTHOR]

Organ and tool detection and segmentation in real time during surgery have been significant challenges in the development of robotic surgery. Most existing detection methods are unsuitable for the surgical environment, where the lighting conditions, occlusions, and anatomical structures can vary significantly. This study presents an organ and surgical tool segmentation and detection algorithm using a manually annotated dataset based on YOLOv8 (You Only Look Once), a state-of-the-art object detection framework. The YOLOv8 deep learning neural network is trained to detect and segment organs and tools during laparoscopic cholecystectomy using a manually annotated dataset of frames taken from actual surgeries. After four experiments using combinations of small and extra-large model sizes and the original and a modified dataset, the resulting algorithm is evaluated and tested in real time on a new surgical video. The method shows it can provide real-time feedback to the surgeon by accurately locating and segmenting the target organs displayed in the surgical video. The method outperforms the baseline methods, with a "bounding box" mean average precision (mAP50) and precession (P) of (50.2%, 51.6%), (52.8%, 76.9%), (83.2%, 81.1%), and (86.3%, 85.7%) for the first, second, third, and fourth experiments, respectively, and a "masking segment" of mAP50 and precession of (50.5%, 51.8%), (54.3%, 76.1%), (82.6%, 80.4%), (86.0%, 85.4%) for the first, second, third, and fourth experiments, respectively. The best-performing model has a speed of around 13.1 ms per frame. This novel application could be a stepping stone in future work, such as developing an algorithm to display the results to the surgeon in a heads-up-display (HUD) to help navigate the scenes or even be implemented in robotic surgeries. [ABSTRACT FROM AUTHOR]

The image rectangling task aims to solve the problem of boundary irregularity in the stitched image without reducing the wide-field-of-view content information of the stitched image. Existing image rectangling methods are either limited by application scenarios, or have a little incomplete phenomenon in the rectangular boundary. To this end, we propose a stepwise deep rectangling model based on the idea of stepwise regression for the general image rectangling task. Considering the influence of factors such as luminosity differences in various regions of the image, we introduce a shallow feature encoder to eliminate the influence of such factors on mesh prediction. At the same time, we embed the mask information into the encoded image to constrain the network to learn a rectangular image with complete boundary. Subsequently, we perform mesh cumulative regression prediction based on the multi-level features extracted by the feature extractor. Experimental results show that the proposed method performs well in a variety of stitched image rectangling scenarios, exhibiting state-of-the-art performance in both qualitative and quantitative comparisons. [ABSTRACT FROM AUTHOR]

Objectives: The Alberta Stroke Program Early CT Score (ASPECTS), a systematic method for assessing ischemic changes in acute ischemic stroke using non-contrast computed tomography (NCCT), is often interpreted relying on expert experience and can vary between readers. This study aimed to develop a clinically applicable automatic ASPECTS system employing deep learning (DL). Methods: This study enrolled 1987 NCCT scans that were retrospectively collected from four centers between January 2017 and October 2021. A DL-based system for automated ASPECTS assessment was trained on a development cohort (N = 1767) and validated on an independent test cohort (N = 220). The consensus of experienced physicians was regarded as a reference standard. The validity and reliability of the proposed system were assessed against physicians' readings. A real-world prospective application study with 13,399 patients was used for system validation in clinical contexts. Results: The DL-based system achieved an area under the receiver operating characteristic curve (AUC) of 84.97% and an intraclass correlation coefficient (ICC) of 0.84 for overall-level analysis on the test cohort. The system's diagnostic sensitivity was 94.61% for patients with dichotomized ASPECTS at a threshold of ≥ 6, with substantial agreement (ICC = 0.65) with expert ratings. Combining the system with physicians improved AUC from 67.43 to 89.76%, reducing diagnosis time from 130.6 ± 66.3 s to 33.3 ± 8.3 s (p < 0.001). During the application in clinical contexts, 94.0% (12,591) of scans successfully processed by the system were utilized by clinicians, and 96% of physicians acknowledged significant improvement in work efficiency. Conclusion: The proposed DL-based system could accurately and rapidly determine ASPECTS, which might facilitate clinical workflow for early intervention. Clinical relevance statement: The deep learning-based automated ASPECTS evaluation system can accurately and rapidly determine ASPECTS for early intervention in clinical workflows, reducing processing time for physicians by 74.8%, but still requires validation by physicians when in clinical applications. Key Points: The deep learning-based system for ASPECTS quantification has been shown to be non-inferior to expert-rated ASPECTS. This system improved the consistency of ASPECTS evaluation and reduced processing time to 33.3 seconds per scan. 94.0% of scans successfully processed by the system were utilized by clinicians during the prospective clinical application. [ABSTRACT FROM AUTHOR]

Restoring high dynamic range images with a single low dynamic range image is a challenging task due to the limited brightness range captured by most consumer camera sensors, leading to missing details in overexposed or underexposed areas of the image. The current methods suffer from weak reconstruction ability for inconspicuous complex features in the extreme exposure area, resulting in artifacts still appearing in the extreme exposure area after reconstruction. This paper proposes an end-to-end network based on dual attention to solve this problem. Specifically, this paper designs a pixel-level spatial attention mechanism with luminance mask to refine the spatial attention to the pixel level, and can focus on inconspicuous feature areas in the overexposed and underexposed regions. This paper also designs a concentrated channel attention mechanism to maximize the use of image context information, accurately assigning the channel weight and effectively extracting the complex features in the image. Qualitative and quantitative evaluations show the superiority of our approach over state-of-the-art methods. [ABSTRACT FROM AUTHOR]

In Lie group convolutional neural networks (LG-CNNs), the calculation and storage of Lie group distances have quadratic space complexity. In order to improve the memory utilization efficiency of LG-CNNs, a novel Lie group convolutional neural network called LieCConv is proposed. LieCConv utilizes an innovative sampling algorithm and a linear space complexity calculation and storage approach for Lie group distances, substantially enhancing network memory efficiency. Firstly, LieCConv employs a novel sampling algorithm called array-neighborhood sampling (ANS) in the downsampling stage. ANS only requires neighborhood information to obtain an excellent sample set with a low threshold of use. The sample set generated by ANS reflects the distribution of the original set. Then, LieCConv adopts a batch calculation and storage scheme for Lie group distances, which effectively declines the space complexity of calculating and storing Lie group distances from quadratic complexity to linear complexity, reducing the memory consumption during training. Finally, the contrast between ANS and farthest point sampling was presented, demonstrating that ANS better captures the distribution characteristics of the original dataset. The memory usage of LieCConv and LieConv was compared, revealing that LieCConv reduces the memory usage for calculating and storing Lie group distances to less than 500 MB. And the performance of LieCConv was evaluated on RotMNIST, RotFashionMNIST and TT100K, validating that LieCConv is universal and effective. [ABSTRACT FROM AUTHOR]

Purpose: Catheter-based radiofrequency ablation for pulmonary vein isolation has become the first line of treatment for atrial fibrillation in recent years. This requires a rather accurate map of the left atrial sub-endocardial surface including the ostia of the pulmonary veins, which requires dense sampling of the surface and currently takes more than 10 min. The focus of this work is to provide left atrial visualization early in the procedure to ease procedure complexity and enable further workflows, such as using catheters that have difficulty sampling the surface. Methods: We propose a dense encoder–decoder network with a novel regularization term to reconstruct the shape of the left atrium from partial data which is derived from simple catheter maneuvers. To train the network, we acquire a large dataset of 3D atria shapes and generate corresponding catheter trajectories, from which traversed point clouds are obtained. Once trained, we show that the suggested network can sufficiently approximate the atrium shape based on a given trajectory. Results: We compare several network solutions for the 3D atrium reconstruction. We demonstrate that the solution proposed produces realistic visualization using partial acquisition within a 3-min time interval using human clinical cases. [ABSTRACT FROM AUTHOR]

Purpose: This work introduces a black box intelligent fault classification algorithm designed to distinguish between classes that result from minimal changes caused by varying numbers of faults present in the twin rotor multi-input–multi-output system (TRMS). Methods: Initially, faults present in steady state vibration signals are classified using features that are extracted from classical signal processing tools. Subsequently, results are improved by combining features extracted through the proposed algorithm. The methodology employs a long short term memory (LSTM) network to extract features for stationary signals, and it combines with convolutional neural network (CNN) to extract features from the transform of both stationary and nonstationary signals. Results: The proposed fault detection algorithm has been tested using vibration signals collected from steady and unsteady state TRMS encompassing healthy and faulty scenarios. Maximum accuracy of 0.96 and 0.77 have been achieved in the case of steady state and unsteady state, respectively. Conclusion: The study offers novel feature extraction technique using LSTM for steady state, and LSTM with CNN for unsteady state from classical signal processed data to improve the accuracy of fault classification. [ABSTRACT FROM AUTHOR]

The unmanned aerial vehicles are used with LIDAR technology and the CNN method to detect damages to roads, buildings, and bridges. The Light detection and ranging (LIDAR) is used for mapping and capturing the damage to roads and buildings, and it is a 3D mapping. The convolutional neural network (CNN) method and deep learning method are used to properly research the damaged areas and depend on low- to high-level pattern detection. It is used in visual detection and shows consistently superior accuracy for spectrogram classifications. It collects the data from damaged areas and gives the information to the device. Here, the instructions are designed in Python. We use multisensory to detect the cracks and pits, and the damaged places will be detected using sensors and sent as a pronouncement. The images are captured by the LIDAR and processed according to the instructions given by the build programming language. It is used to reduce work time and make it highly efficient. It can detect the damages automatically on high buildings, bridges, and roads. It is mostly used in civil departments. The experimental results shows that the proposed model attained the maximum accuracy of 95.88%. [ABSTRACT FROM AUTHOR]

An immense volume of user-generated content exists online due to the exponential growth of internet usage among individuals. However, this abundance presents substantial challenges for recommendation systems and online customer services. The diverse data, encompassing consumer emails, reviews, and comments, contain a broad spectrum of information, making it challenging to decipher the underlying emotions and nuances. In social media analytics, sentiment analysis has emerged as a pivotal tool to unveil the emotional context embedded within textual data, offering deeper insights into user attitudes, opinions, and sentiments. This study introduces a novel strategy to strengthen the performance and precision of sentiment analysis-based book recommendation systems through ensemble learning on hybrid deep learning models. These book recommendation models leverage customer ratings and reviews from a vast Amazon books dataset as input. Initially, we used TextBlob to assess the polarity of customer reviews, categorizing them into neutral, negative, and positive sentiments. Subsequently, the input data underwent preprocessing, tokenization, and word embedding using bidirectional encoder representations from transformers (BERT). To effectively analyze and filter processed review comments and ratings, the proposed ensemble model integrates a diverse array of hybrid deep learning architectures, including long short-term memory (LSTM), bidirectional LSTM (BiLSTM), gated recurrent unit (GRU), and convolutional neural network (CNN). Extensive experimentation validated the superiority of the proposed ensemble model, achieving an impressive accuracy and F1-score of 98.21%. The significance of the approach lies in its ability to provide more accurate and contextually relevant book recommendations by considering the nuanced emotions expressed in customer reviews. This contributes to enhancing user satisfaction and engagement with recommendation systems, ultimately improving the overall quality of personalized book suggestions. Evaluation metrics further validate the efficacy of the proposed model, underscoring its practical utility in real-world applications of sentiment-based book recommendation systems. [ABSTRACT FROM AUTHOR]

Approximate computing is accepted for error-tolerant tasks including image and signal processing applications, where execution time is more critical than accuracy. Approximation in the multiplier design is considered highly beneficial due to reduced hardware requirements and yet achieves acceptable inference from the rendered output. A probabilistic approach to build approximate compressors specifically for Radix-4 modified booth multiplier is designed to provide power and footprint savings, and at the same time offer performance benefits with lesser errors. The inexact compressor design when applied on the booth encoded reduced partial products (PP) offers a significant improvement over the existing approximate multiplier design. Five variants of inexact multiplier designs were derived by placing approximate compressors across the columns of the Booth multiplier generated partial product matrix (PPM). Two additional design variants including truncation on the lower significant part of the PPM in the multiplier design were also investigated. All the proposed multiplier design variants were synthesized and characterized for hardware parameters, and error metrics. Two design configurations: Design-1 with the conventional approach of sign extending each PP rows till the (2 n - 1) th bit in an n × n multiplier, and Design-2 without sign extension with a modified PPM, are considered to evaluate all the seven proposed multiplier variants in each case. The 9 × 9 Design-1 achieved the maximum benefit of 46.27%, 61.18%, 16.79%, and 66.18% in footprint, power, delay, and overall product of delay and power (PDP), while Design-2 achieved the maximum of 61.33%, 77.54%, 41.45%, and 86.8% footprint, power, delay, and PDP savings respectively over their exact versions. 16 × 16 Design-2 variant exhibited the maximum of 11.29%, 32.01%, and 32.42% area, power and PDP improvements, respectively, among the evaluated design variants in comparison with the exact one. The proposed approximate multipliers also showcased superior Mean Relative Error Distance (MRED) and P RED (Probability of getting an RED smaller than 2%) when compared with various existing works. The proposed design variants in this work were evaluated for the applications: image smoothing using the Gaussian filter, image segmentation using k-means clustering which is an unsupervised learning algorithm popularly used in computer vision systems, μ - l a w algorithm which is a standard companding algorithm mainly used in telecommunication systems, and Convolutional Neural Network (CNN) to present error-tolerant results. [ABSTRACT FROM AUTHOR]

In human–human interactions, detecting emotions is often easy as it can be perceived through facial expressions, body gestures, or speech. However, in human–machine interactions, detecting human emotion can be a challenge. To improve this interaction, Speech Emotion Recognition (SER) has emerged, with the goal of recognizing emotions solely through vocal intonation. In this work, we propose a SER system based on deep learning approaches and two efficient data augmentation techniques such as noise addition and spectrogram shifting. To evaluate the proposed system, we used three different datasets: TESS, EmoDB, and RAVDESS. We employe several algorithms such as Mel Frequency Cepstral Coefficients (MFCC), Zero Crossing Rate (ZCR), Mel spectrograms, Root Mean Square Value (RMS), and chroma to select the most appropriate vocal features that represent speech emotions. Three different deep learning models were imployed, including MultiLayer Perceptron (MLP), Convolutional Neural Network (CNN), and a hybrid model that combines CNN with Bidirectional Long-Short Term Memory (Bi-LSTM). By exploring these different approaches, we were able to identify the most effective model for accurately identifying emotional states from speech signals in real-time situation. Overall, our work demonstrates the effectiveness of the proposed deep learning model, specifically based on CNN+BiLSTM enhanced with data augmentation for the proposed real-time speech emotion recognition. [ABSTRACT FROM AUTHOR]

In this paper, we propose a Three-stream Inception Former-based Action Recognition Network, called TIFAR-Net to recognize actions in Infrared (IR) videos. It consists of two major stages. First, fine-tuning and feature extraction using a Inception Transformer (IFormer) network, and second, feature fusion and classification using Multi-Head Self-Attention (MHSA) network. Specifically, input IR videos are converted into compact yet effective representations referred to as Optical Flow Motion Images, Optical Flow Dynamic Images and Infrared Motion Images. The first two types of images are constructed using optical flow fields, while the third type of image is computed directly from raw IR frames. These image-based representations enable us to fine-tune a three-stream IFormer network, which is a hybrid Convolution Neural Network and Vision Transformer model. Features extracted from each stream are concatenated and passed through MHSA module to prune unimportant information and capture key global information from input images. Finally, classification is performed using fully connected and Softmax layers. Through extensive ablation experiments, we verified that our proposed TIFAR-Net improves the performance of IR action recognition and achieves state-of-the-art results on InfAR dataset (88.5%), IITR-IAR dataset (77.93%) and UNISV dataset(97.08%). [ABSTRACT FROM AUTHOR]

Information technology brings us not only marvelous convenience and productivity, but also potential insecure factor, which may pose threats to our properties, data or even reputation. Malicious software is exactly an accomplice of such attacks. Fundamentally, the key step to deal with malicious software is to accurately identify and classify it. Although traditional static and dynamic analysis approaches could accomplish this task to some extent, they have intrinsic defects in terms of variant feature exaction, vulnerability to code obfuscation and encryption, or excessive resource consumption. Recently, CNN-based malware classification methods, which employ CNN models to classify visualized malware images, provide a promising way to accomplish malware classification tasks. However, most mainstream CNN models require inputs with a fixed size, while various sizes of original malware samples frequently lead to various sizes of malware visualization images. Simply resizing these images causes losses of malware features, resulting in drops of classification accuracy. In this paper, we propose a malware visualization method based on transition probabilities of malware operation codes to generate proper images with a uniform size as inputs for CNN models. As a result, the conventional resizing operations could be avoided. The proposed method is compatible with most mainstream CNN models. Moreover, the proposed method could address problems concerning insufficient or imbalanced datasets, which may challenge the classification abilities of CNN models. Experimental results demonstrate the excellent compatibility and classification performance of the proposed method in terms of accuracy, precision, recall and F1-score. For reproducible research, the source codes and training models of the proposed method are available at . [ABSTRACT FROM AUTHOR]

Artificial Intelligence (AI) has emerged as a promising solution for real-time monitoring of the quality of additively manufactured (AM) metallic parts. This study focuses on the Laser-based Directed Energy Deposition (L-DED) process and utilizes embedded vision systems to capture critical melt pool characteristics for continuous monitoring. Two self-learning frameworks based on Convolutional Neural Networks and Transformer architecture are applied to process zone images from different DED process regimes, enabling in-situ monitoring without ground truth information. The evaluation is based on a dataset of process zone images obtained during the deposition of titanium powder (Cp-Ti, grade 1), forming a cube geometry using four laser regimes. By training and evaluating the Deep Learning (DL) algorithms using a co-axially mounted Charged Couple Device (CCD) camera within the process zone, the down-sampled representations of process zone images are effectively used with conventional classifiers for L-DED process monitoring. The high classification accuracies achieved validate the feasibility and efficacy of self-learning strategies in real-time quality assessment of AM. This study highlights the potential of AI-based monitoring systems and self-learning algorithms in quantifying the quality of AM metallic parts during fabrication. The integration of embedded vision systems and self-learning algorithms presents a novel contribution, particularly in the context of the L-DED process. The findings open avenues for further research and development in AM process monitoring, emphasizing the importance of self-supervised in situ monitoring techniques in ensuring part quality during fabrication. [ABSTRACT FROM AUTHOR]

The cervical cell classification technique can determine the degree of cellular abnormality and pathological condition, which can help doctors to detect the risk of cervical cancer at an early stage and improve the cure and survival rates of cervical cancer patients. Addressing the issue of low accuracy in cervical cell classification, a deep convolutional neural network A2SDNet121 is proposed. A2SDNet121 takes DenseNet121 as the backbone network. Firstly, the SE module is embedded in DenseNet121 to increase the model's focus on the nucleus region, which contains important diagnostic information, and reduce the focus on redundant information. Secondly, the sizes of the convolutional kernel and pooling window of the Stem layer are adjusted to adapt to the characteristics of the cervical cell images, so that the model can extract the local detailed information more effectively. Finally, the Atrous Dense Block (ADB) is constructed, and four ADB modules are integrated into DenseNet121 to enable the model to acquire global and local salient feature information. The accuracy of A2SDNet121 for two and seven-classification tasks on the Herlev dataset is 99.75% and 99.14%, respectively. The accuracy for two, three, and five-classification tasks on the SIPaKMeD dataset reaches 99.55%, 99.75% and 99.22%, respectively. Compared with other state-of-the-art algorithms, the A2SDNet121 model performs better in the multi-classification task of cervical cells, which can significantly improve the accuracy and efficiency of cervical cancer screening. [ABSTRACT FROM AUTHOR]

The proliferation of cooking videos on the internet these days necessitates the conversion of these lengthy video contents into concise text recipes. Many online platforms now have a large number of cooking videos, in which, there is a challenge for viewers to extract comprehensive recipes from lengthy visual content. Effective summary is necessary in order to translate the abundance of culinary knowledge found in videos into text recipes that are easy to read and follow. This will make the cooking process easier for individuals who are searching for precise step by step cooking instructions. Such a system satisfies the needs of a broad spectrum of learners while also improving accessibility and user simplicity. As there is a growing need for easy-to-follow recipes made from cooking videos, researchers are looking on the process of automated summarization using advanced techniques. One such approach is presented in our work, which combines simple image-based models, audio processing, and GPT-based models to create a system that makes it easier to turn long culinary videos into in-depth recipe texts. A systematic workflow is adopted in order to achieve the objective. Initially, Focus is given for frame summary generation which employs a combination of two convolutional neural networks and a GPT-based model. A pre-trained CNN model called Inception-V3 is fine-tuned with food image dataset for dish recognition and another custom-made CNN is built with ingredient images for ingredient recognition. Then a GPT based model is used to combine the results produced by the two CNN models which will give us the frame summary in the desired format. Subsequently, Audio summary generation is tackled by performing Speech-to-text functionality in python. A GPT-based model is then used to generate a summary of the resulting textual representation of audio in our desired format. Finally, to refine the summaries obtained from visual and auditory content, Another GPT-based model is used which combines the output of the frame summary and audio summary modules and give the final enhanced summary. By minimizing the complications involved with traditional and sophisticated methodologies, this research helps with the development of a straightforward but efficient cooking video summarization system. The results achieved in the work are on par with the existing work in the respective field which demonstrates comparable performance and efficacy in converting cooking videos into detailed recipe texts. [ABSTRACT FROM AUTHOR]

In recent years, the Deep Neural Networks (DNN) have managed tremendous growth and widespread applications. However, their applicability in embedded and mobile devices has been constrained by its complex arrangement, maximal computation, and the storage requirements. This paper designed an Evolutionary Multi-Objective Optimization (EMO) through a Dynamic and Global Filter Pruning (DG-FP) approach (EMO-DGFP) for an effective Convolutional Neural Network (CNN). Initially, Train the CNN in the Python system, then use the DG-FP approach to remove unnecessary filters. Furthermore, the CNN filters were adjusted using an EMO technique. The filter pruning issue has been resolved using an EMO based on non-dominated sorting. Then fine-tuning process was performed to regain its accuracy. The developed model efficiency has been tested against other widely used models, and the results showed 99.96% accuracy, 2% mistake rate, 2.6% compression ratio, and 80.67% pruning rate. [ABSTRACT FROM AUTHOR]

Multivariate Time Series Forecasting (MTSF) is challenging due to the difficulty of extracting complex periodic patterns from temporal data. Currently, many Transformer-based models and MLP-based models achieve success by using Fourier transform (FT) to extract periodic information from time series data. However, Transformer models perform well but use many resources. MLP models are lightweight but too simple to capture complex features. Recent studies show that CNN-based models offer a better balance between efficiency and performance, but they struggle to capture both coarse and fine-grained temporal features. In addition, the single perspective of FT limits its ability to represent complex periodic features. To address these issues, we propose the Multi-Scale Fractional Fourier Transform Convolutional Neural Network (MFFCNN) for MTSF. The fractional Fourier transform (FrFT) is an extension of the FT. It extracts periodic features from different angles by adjusting the rotation. MFFCNN uses these multi-scale features to generate adaptive patches for different datasets. Then, 2D convolution is applied to model relationships within and between patches, capturing both inter- and intra-series features from a multi-scale perspective. Results on ten benchmark datasets show that MFFCNN outperforms seven state-of-the-art models. Our code is stored in https://github.com/immortalityC?tab=repositories. [ABSTRACT FROM AUTHOR]

Artificial neural networks have been one of the science’s most influential and essential branches in the past decades. Neural networks have found applications in various fields including medical and pharmaceutical services, voice and speech recognition, computer vision, natural language processing, and video and image processing. Neural networks have many layers and consume much energy. Approximate computing is a promising way to reduce energy consumption in applications that can tolerate a degree of accuracy reduction. This paper proposes an effective method to prevent accuracy reduction after using approximate computing methods in the CNNs. The method exploits the k-means clustering algorithm to label pixels in the first convolutional layer. Then, using one of the existing pruning methods, different pruning amounts have been applied to all layers. The experimental results on three CNNs and four different datasets show that the accuracy of the proposed method has significantly improved (by 17%) compared to the baseline network. [ABSTRACT FROM AUTHOR]

Accurately extracting organs from medical images provides radiologist with more comprehensive evidences to clinical diagnose, which offers up a higher accuracy and efficiency. However, the key to achieving accurate segmentation lies in abundant clues for contour distinction, which has a high demand for the network architecture design and its practical training status. To this end, we design auxiliary and refined constraints to optimize the energy function by supplying additional guidance in training procedure, thus promoting model’s ability to capture information. Specifically, for the auxiliary constraint, a set of convolutional structures are involved into a conventional network to act as a discriminator, then adversarial network is established. Based on the obtained architecture, we further build adversarial mechanism by introducing a second discriminator into segmentor for refinement. The involvement of refined constraint contributes to ameliorate training situation, optimize model performance, and boost its ability of collecting information for segmentation. We evaluate the proposed framework on two public databases (NIH Pancreas-CT and MICCAI Sliver07). Experimental results show that the proposed network achieves comparable performance to current pancreas segmentation algorithms and outperforms most state-of-the-art liver segmentation methods. The obtained results on public datasets sufficiently demonstrate the effectiveness of the proposed model for organ segmentation. [ABSTRACT FROM AUTHOR]

The application of artificial neural networks (ANNs) can be found in numerous fields, including image and speech recognition, natural language processing, and autonomous vehicles. As well, intrusion detection, the subject of this paper, relies heavily on it. Different intrusion detection models have been constructed using ANNs. While ANNs are relatively mature to construct intrusion detection models, some challenges remain. Among the most notorious of these are the bloated models caused by the large number of parameters, and the non-interpretability of the models. Our paper presents Convolutional Kolmogorov-Arnold Networks (CKANs), which are designed to overcome these difficulties and provide an interpretable and accurate intrusion detection model. Kolmogorov-Arnold Networks (KANs) are developed from the Kolmogorov-Arnold representation theorem. Meanwhile, CKAN incorporates a convolutional computational mechanism based on KAN. The model proposed in this paper is constructed by incorporating attention mechanisms into CKAN's computational logic. The datasets CICIoT2023 and CICIoMT2024 were used for model training and validation. From the results of evaluating the performance indicators of the experiments, the intrusion detection model constructed based on CKANs has an attractive application prospect. As compared with other methods, the model can predict a much higher level of accuracy with significantly fewer parameters. However, it is not superior in terms of memory usage, execution speed and energy consumption. [ABSTRACT FROM AUTHOR]

The scenarios when heavy rainfall and high tides occur in succession or simultaneously can lead to compound flooding. Compound floods exhibit greater destructiveness than floods caused by one driver in coastal cities. Prediction for compound floods with real-time and high accuracy can contribute to mitigating the losses caused by floods. However, existing rapid forecasting studies neglect the compound impact of rainfall and tides in coastal floods. In this study, the information on rainfall and tides is utilized as input features to capture the drivers of compound flooding. To reduce the risk of overfitting, the light gradient boosting machine (LightGBM) is employed for feature selection. The one-dimensional convolutional neural network (CNN) is then trained on the reduced-dimensionality data. Hence, we construct LightGBM-CNN to predict flood distribution in coastal cities. The model is applied on Haidian Island, Hainan Province, China. The results indicate that incorporating rainfall and tides as input features significantly reduced the mean absolute error (MAE) from 0.179 to 0.044 and the root mean square error (RMSE) from 0.223 to 0.101, compared to using rainfall as input features. Compared to the CNN without feature selection using LightGBM, the performance of LightGBM-CNN has shown a significant improvement. The results suggest that the LightGBM-CNN offers a foundational reference for compound flood forecasting in coastal cities. [ABSTRACT FROM AUTHOR]

Sequential patterns and the order of items in sequences are particularly important for sequential recommendation (SR), which decide what item will interact with the user. However, there are still some problems for the existing methods: (1) They treat all the features of each item equally, we believe that those features the user pays more attention to of each item play a key role to predict next item. (2) Many methods not only ignore sequential patterns and the order of sequences, but also cannot highlight more important features, they only focus on whether the features exist. To address these issues, we propose the Sequential Patterns Enhanced Capsule Network (SPECN). SPECN leverages a self-attention mechanism, using user information as a guide to highlight the most relevant features for each item, then concatenates these features with the original item features in the sequence.SPECN applies horizontal and vertical capsule networks which package neurons into vectors to extract sequential patterns features and the order of sequences. The horizontal capsule network enhances sequential pattern features by learning both the original and user-focused features of individual or adjacent items, containing original features and those features that the user pays more attention to of single item or adjacent items’ features (features of the previous item that the users pay more attention to and features of the current item) to enhance the sequential patterns features. The vertical capsule network captures finer-grained feature representations for each item, improving the recommendation quality. We conduct several experiments on three real-world datasets to demonstrate the superiority of SPECN, outperforming existing methods in terms of accuracy and robustness. [ABSTRACT FROM AUTHOR]

Accurately segmenting organs, or diseased regions of varying sizes, is a challenge in medical image segmentation tasks with limited datasets. Although Transformer-based methods have self-attention mechanisms, excellent global modelling abilities and can effectively focus on crucial areas in medical images, they still face the intractable issues of computational complexity and excessive reliance on data. The Swin-Transformer has partly addressed the problem of computational complexity, however the method still requires large amounts of training data due to its lack of inductive bias capabilities. When applied to medical image segmentation tasks with small datasets, this leads to suboptimal performances. In contrast, the CNN-based methods can compensate for this limitation. To address this issue further, this paper proposes Swin-IBNet, which combines the Swin-Transformer with a CNN in a novel manner to imbue it with inductive bias capabilities, reducing its reliance on data. During the encoding process of Swin-IBNet, two novel and crucial modules, the feature fusion block (FFB) and the multiscale feature aggregation block (MSFA), are designed. The FFB is responsible for propagating the inductive bias capability to the Swin-Transformer encoder. Different from the previous use of multiscale features, MSFA efficiently leverages multiscale information from different layers through self-learning. This paper not only attempts to analyse the interpretability of the proposed Swin-IBNet but also performs more verifications on the public Synapse, ISIC 2018 and ACDC datasets. The experimental results show that Swin-IBNet is superior to the baseline method, Swin-Unet, and several state-of-the-art methods. Especially on the Synapse dataset, the DSC of Swin-IBNet surpasses that of Swin-Unet by 3.45%. [ABSTRACT FROM AUTHOR]

Multi-focus image fusion can obtain high-quality images by overcoming the limited depth of field of optical lenses. Benefiting from deep learning, we design a local-global joint attention module and propose a novel multi-focus image fusion network. The module essentially is an attention module. Local and global features are extracted respectively through point-wise convolution and spatial pyramid pooling. A joint attention map is produced by reducing the dimension and fusing these two features. The proposed network is mainly composed of a feature fusion module and two weight-shared dense feature extraction modules, each connected to six consecutive attention modules. Such design has two benefits: adequate extraction of initial features and capturing of local and global features. Subjective visual evaluation demonstrates that the proposed network can preserve the authenticity of fusion results. And it also reduces the appearance of artifacts and detail losses between the focus and defocus regions. Objective metric evaluation shows that the proposed network outperforms most of the existing models, such as SwinFusion, GACN, and UFA-FUSE, in Lytro, MFI-WHU, and MFFW datasets. Ablation experiments demonstrate that the design of attention and the overall framework of the network is reasonable. Overall, the proposed model can finish the multi-focus image fusion task with high quality. [ABSTRACT FROM AUTHOR]

Underwater object segmentation presents significant challenges due to the degradation of image quality and the complexity of underwater environments. In recent years, deep learning has provided an effective approach for object segmentation. However, DeepLabV3+, as a classical model for general scenes, shows limitations in achieving accurate and real-time segmentation in complex underwater conditions. To address this issue, we propose a DeepLab-FusionNet, an extended version of DeepLabV3+, specifically designed for underwater object segmentation. The model utilizes a multi-resolution parallel branch structure to extract multi-scale information and employs an improved inverted residual structure as the basic feature extraction module in the encoding network. Structural reparameterization technique is introduced to optimize inference speed and memory access costs during the inference stage. Additionally, a module for linking deep and shallow level information is constructed to reduce the loss of detail and spatial information during downsampling and convolution. Evaluation on the SUIM dataset shows a 3.3% increase in mean Intersection over Union (mIoU) and a speed improvement of 34 frames per second (FPS) compared to the baseline model DeepLabV3+. Further comparisons with other classic lightweight models and Transformer-based models on the UIIS and TrashCan datasets demonstrate that our model achieves good accuracy and balanced computational efficiency in challenging underwater environments. Although there is room for improvement due to overfitting and fixed convolution kernel limitations, future integration with Transformer methods is planned. Our model offers an effective solution for real-time target segmentation for underwater robots, with broad applications in human exploration and development of marine resources. Our codes are available at: [ABSTRACT FROM AUTHOR]

Cervical cancer is the fourth most prevalent cancer among women worldwide and a major contributor to cancer-related mortality in females. Manually classifying cytopathology screening slides remains one of the most important and commonly used methods for diagnosing cervical cancer. However, this method requires the participation of medical experts and is highly labor intensive. Consequently, in regions with limited medical resources, prompt cervical cancer diagnosis is challenging. To address this issue, the BiNext-Cervix model, a new deep learning framework, has been proposed to rapidly and accurately diagnose cervical cancer via Pap smear images. BiNext-Cervix employs Tokenlearner in the initial stage to facilitate interaction between two pixels within the image, enabling the subsequent network to better understand the image features. Additionally, the BiNext-Cervix integrates the recently introduced ConvNext and BiFormer models, allowing for deep exploration of image information from both local and global perspectives. A fully connected layer is used to fuse the extracted features and perform the classification. The experimental results demonstrate that combining ConvNext and BiFormer achieves higher accuracy than using either model individually. Furthermore, the proposed BiNext-Cervix outperforms other commonly used deep learning models, showing superior performance. [ABSTRACT FROM AUTHOR]

As web attacks have recently increased in number and sophistication, traditional machine learning methods have struggled to defend against well-designed attacks. Therefore, deep learning methods have been widely used in web attack detection, leveraging their ability to discern intricate features within the original payload for precise identification of web application threats. In this study, we propose a novel hybrid neural network model for web attack detection, named hybrid convolutional long short-term memory (HCLSTM). Specifically, the HCLSTM model utilizes two branches to extract features from Hypertext Transfer Protocol (HTTP) request packet: a Deep Feedforward Neural Network (DFNN) branch for extracting word features from Uniform Resource Locator (URL), and a Convolutional Neural Network (CNN) branch for capturing combinatorial and local relationships within payloads. Then, the extracted features from both branches are concatenated and subsequently fed into a Bidirectional Long Short-Term Memory (Bi-LSTM) network integrated with a self-attention mechanism, designed to capture intricate link relationships between URL and payloads. The final classification layer produces the detection results. To evaluate the proposed model, we conducted experiments on CSIC 2010 HTTP dataset. The experimental results reveal that HCLSTM can accurately detect web attacks with a high accuracy of 99.46% and a low false positive rate of 0.02%. [ABSTRACT FROM AUTHOR]

Introduction: A large number of middle-aged and elderly patients have an insufficient understanding of osteoporosis and its harm. This study aimed to establish and validate a convolutional neural network (CNN) model based on unenhanced chest computed tomography (CT) images of the vertebral body and skeletal muscle for opportunistic screening in patients with osteoporosis. Materials and methods: Our team retrospectively collected clinical information from participants who underwent unenhanced chest CT and dual-energy X-ray absorptiometry (DXA) examinations between January 1, 2022, and December 31, 2022, at four hospitals. These participants were divided into a training set (n = 581), an external test set 1 (n = 229), an external test set 2 (n = 198) and an external test set 3 (n = 118). Five CNN models were constructed based on chest CT images to screen patients with osteoporosis and compared with the SMI model to predict the performance of osteoporosis patients. Results: All CNN models have good performance in predicting osteoporosis patients. The average F1 score of Densenet121 in the three external test sets was 0.865. The area under the curve (AUC) of Desenet121 in external test set 1, external test set 2, and external test set 3 were 0.827, 0.859, and 0.865, respectively. Furthermore, the Densenet121 model demonstrated a notably superior performance compared to the SMI model in predicting osteoporosis patients. Conclusions: The CNN model based on unenhanced chest CT vertebral and skeletal muscle images can opportunistically screen patients with osteoporosis. Clinicians can use the CNN model to intervene in patients with osteoporosis and promptly avoid fragility fractures. Critical relevance statement: The CNN model based on unenhanced chest CT vertebral and skeletal muscle images can opportunistically screen patients with osteoporosis. Clinicians can use the CNN model to intervene in patients with osteoporosis and promptly avoid fragility fractures. Key Points: The application of unenhanced chest CT is increasing. Most people do not consciously use DXA to screen themselves for osteoporosis. A deep learning model was constructed based on CT images from four institutions. [ABSTRACT FROM AUTHOR]

Ethiopia, known as the birthplace of coffee, relies on coffee exports as a major source of foreign currency. This research paper focuses on developing a hybrid feature mining technique to automatically classify Ethiopian coffee beans based on their provenance: Harrar, Jimma, Limu, Sidama, and Wellega, which correspond to their botanical origins. A dataset of coffee bean images is collected from various regions through the Ethiopian Commodity Exchange (ECX) in Addis Ababa. The proposed system incorporates preprocessing phases including image resizing, filtering, contrast enhancement, noise removal, grayscale conversion, and segmentation using a combined thresholding and K-means approach for grayscale and RGB images, respectively. Classification is performed using a radial basis function (RBF) kernel function of support vector machine (SVM). To address the color-feature similarity challenge, the study explores merging color and texture features using the histogram of oriented gradients (HOG) local feature descriptor. Performance evaluation is conducted for HOG feature extraction, CNN feature extraction, and a hybrid feature vector (HOG-CNN) using a multi-class SVM classifier, achieving accuracies of 74.17%, 85.83%, and 97.5%, respectively. The deep-shallow-based feature (CNN-HOG) combination demonstrates the highest accuracy of 97.5% in this study. The findings highlight the effectiveness of the proposed hybrid feature mining approach in automatically classifying Ethiopian coffee bean varieties, with potential applications in quality control and traceability within the coffee industry. [ABSTRACT FROM AUTHOR]

Sentiment analysis has become a difficult and important task in the current world. Because of several features of data, including abbreviations, length of tweet, and spelling error, there should be some other non-conventional methods to achieve the accurate results and overcome the current issue. In other words, because of those issues, conventional approaches cannot perform well and accomplish results with high efficiency. Emotional feelings, such as fear, anxiety, or traumas, often stem from many psychological issues experienced during childhood that can persist throughout life. In addition, people discuss and share their ideas on social media, often unconsciously representing their hidden emotions in the comments. This study is about sentiment analysis of tweets shared by several people. In fact, sentiment analysis can determine whether the shared comments and tweets are positive or negative. The paper introduces the use of a Convolutional Neural Network (CNN), a kind of neural network, optimized by the Enhanced Gorilla Troops Optimization Algorithm (CNN-EGTO). Two datasets provided by the SemEval-2016 are used to evaluate the system, while the polarity of tweets were manually determined. It was determined by the findings of the present study that the suggested model could approximately achieve the values of 98%, 95%, 98%, and 96.47% for accuracy, precision, recall, and F1-score, respectively, for positive polarity. In addition, the suggested model could gain the values of 97, 96, 98, and 97.49 for precision, recall, accuracy, and F1-score, respectively, for negative polarity. Consequently, it was found that the suggested model could outperform the other models by considering their performance and efficiency. These values of performance metrics represent that the suggested model could determine the polarity of sentence, positive or negative, with great efficiency. [ABSTRACT FROM AUTHOR]

Mobile Ad Hoc Networks (MANETs) are increasingly replacing conventional communication systems due to their decentralized and dynamic nature. However, their wireless architecture makes them highly vulnerable to flooding attacks, which can disrupt communication, deplete energy resources, and degrade network performance. This study presents a novel hybrid deep learning approach integrating Convolutional Neural Networks (CNN) with Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) architectures to effectively detect and mitigate flooding attacks in MANETs. To optimize the model's efficiency, a unique DECEHGS algorithm combining Differential Evolution and Evolutionary Population Dynamics techniques is employed, enhancing both convergence and performance. The proposed model demonstrates significant improvements over existing methods, achieving an accuracy of 95%, a 12% increase in packet delivery ratio, and a 20% reduction in routing overhead compared to traditional techniques. These advancements underline the model's superiority in detecting malicious nodes, conserving energy, and ensuring reliable network performance. The comprehensive evaluation using MATLAB R2023a validates the proposed approach as an effective and energy-efficient solution for enhancing MANET security. [ABSTRACT FROM AUTHOR]

Some hardware is similar in color and shape between different classes, and some hardware varies within a class, thereby decreasing the accuracy of classification models. The baseline of image classification depends on network architectures. To solve this problem, this paper proposes an improved residual network architecture that achieves high accuracy in hardware classification. This paper proposes a hardware dataset that fills the blank of hardware images. The improved residual network architectures are based on depthwise over-parameterized convolution and pyramidal convolution. This model could obtain information from different levels and channels. Moreover, compared with the novel architectures, our model does not require increasing the computation. Experiments on the hardware dataset show that our neural network architectures achieved better accuracy compared with notable architectures such as ResNet50 and DenseNet121. The classification accuracy of the proposed network is 94.6% and the f1 score is 90.26%. To verify the proposed network architecture, we applied our model in fine-grained datasets; our model achieved the best performance in the same training setting. [ABSTRACT FROM AUTHOR]

Cross-modal retrieval can break through the limitations of modalities and carry out information retrieval across data of different modalities to meet the needs of users in obtaining multi-modal correlation retrieval. Cloud computing has the advantages of high efficiency and low cost, but data security hinders its development. While cloud computing offers high efficiency and cost-effectiveness, concerns surrounding data security impede its full potential. Privacy-preserving cross-modal retrieval emerges as a viable solution, catering to users’ demands for efficient retrieval while safeguarding data confidentiality. However, a major challenge still exists in this field: how to bridge the inherent semantic gap within heterogeneous and chaotic information. To address this challenge, this paper proposes dual-branch networks for privacy-preserving cross-modal retrieval in cloud computing. Firstly, a dual-branch feature extraction network of encrypted image-text is constructed, enhancing the extraction of meaningful features from encrypted data. Secondly, a cross-modal alignment method is designed to eliminate the heterogeneous gap between different modalities through the alignment within and between modalities. Finally, to fully exploit the storage and computing advantages of cloud computing, both encrypted data and the cross-modal feature extractor are deployed to the cloud. Leveraging the dynamic update capabilities of cloud-stored encrypted data enables continuous model refinement, enhancing retrieval accuracy while reducing the storage and computational burdens on data owners. Extensive experiments conducted on the publicly available benchmark image-text dataset Wikipedia indicate that, compared to existing methods, our approach achieves improvements of 5.4%, 1%, 1.6%, and 20.1% in the four metrics of image-to-text (i2t), text-to-image (t2i), image-to-all (i2all), and text-to-all (t2all), respectively. [ABSTRACT FROM AUTHOR]

Metal surface defect detection has played a crucial role in the manufacturing of high-quality metal products. However, defect detection is challenging due to inter-class similarities and intra-class variances. Previous studies have primarily focused on detecting surface defects in individual metal categories, with limited exploration of models capable of detecting defects across multiple metal types. Aiming at the above issues, we proposed an improved model named DEFECT-YOLO, based on the YOLOv8s model. Concretely, a lightweight Effective Multi-scale Convolution Module (EMCM) is designed to enhance accuracy and reduce parameter count. Secondly, a multi-scale Atrous Spatial Pyramid Pooling with Efficient Multi-scale Attention (ASPPE) module is employed to enhance the ability of the model to capture local information. Subsequently, the multi-branch coordinate attention (MCA) mechanism is implemented in each prediction branch to further enhance the feature extraction capability for defects. Finally, the Normalized Wasserstein Distance (NWD) loss function is applied to jointly evaluate object localization loss with the complete intersection over union (CIoU) loss function, thereby improving the accuracy of the model for small-scale defects. Comprehensive experiments were conducted on a dataset containing various metal surface defects. The results indicate that the DEFECT-YOLO model achieves a 3.6% improvement in mean average precision (mAP@0.5) compared to the baseline model, with only a modest increase of 1.64 M parameters. Additionally, the model maintains real-time detection capabilities at 89 frames per second (FPS), which is essential for industrial applications. The code is available at . [ABSTRACT FROM AUTHOR]

The growth of cloud computing has encouraged resource-constrained data owners to upload skeleton data to the cloud for action identification, but this practice increases the risk of privacy breaches. Although the traditional anonymized privacy protection method protects the user privacy, it sacrifices the performance of action recognition. To solve the above problems, a Convolutional Neural Network (CNN) architecture compatible with Fully Homomorphic Encryption (FHE) is proposed in this paper, which can achieve secure action recognition without sacrificing the accuracy of action recognition. In addition, to solve the problem of low computational efficiency of the Residue Number System (RNS) variant of CKKS (RNS-CKKS) applied to CNN networks, a parallel fully homomorphic convolution method is designed to improve computational efficiency. To reduce the overhead of rotating key generation and transmission, a multi-layer key generation system is constructed. Finally, the superiority of the proposed model is verified on real data sets. [ABSTRACT FROM AUTHOR]