Search

In response to the suddenness, uncertainty, and complexity of public emergencies, this paper presents an analytical framework for the evolution of public emergencies based on the research idea of "event-relationship-logic". The aim is to comprehensively grasp, comprehend, and predict the development logic and internal mechanisms of public emergencies, assisting emergency departments in promptly understanding the progression of events. Building upon this foundation, this paper proposes a complete process for constructing an Event Logic Graph (ELG) in the field of public emergencies and introduces a multi-structured convolutional neural network model that incorporates word position features to extract causality between events. Experimental findings indicate that the multistructured convolutional neural network model demonstrates higher accuracy and effectiveness compared to the traditional template matching method. Furthermore, through the empirical analysis, we find that the evolutionary path of the causal event chain is relatively short, and the evolutionary motivation of public emergencies is intricate throughout their entire lifecycle. Consequently, emergency departments should focus more on key nodes, central nodes, and intermediate nodes within the causal event chain and adeptly identify these specific nodes to swiftly uncover correlations between events. Additionally, public emergencies easily convert into public crisis, underscoring the necessity for emergency departments to comprehend the evolutionary motivation and inducement factors of public emergencies. [ABSTRACT FROM AUTHOR]

With the continuous development of 3D data acquisition technology, point cloud data, an essential form of 3D data representation is widely used in fields such as autonomous driving, indoor navigation, virtual reality, and augmented reality. As an essential branch of point cloud data processing, cloud classification has important research significance and value. However, this task poses challenges due to the sparsity, irregularity, and unordered nature of point cloud data. Among most of the methods dealing with this problem, there are problems of inadequate extraction of local features, low accuracy of classification, and poor generalization of point cloud data. Therefore, this paper addresses the above issue. This paper presents an improved 3D point cloud classification network by enhancing the Dynamic Graph Convolutional Neural Network (DGCNN). Firstly, this paper combines K-Nearest Neighbors (KNN) and ball radius query for feature extraction to better capture the local structural information in cloud classification data. Secondly, a conventional convolution layer is incorporated between the second and third graph convolution layers to enhance the feature representation in the proposed approach. Finally, this paper uses a global pooling method that combines maximum pooling and average pooling to construct the global structure of the point cloud while preserving the most critical information. Realize 3D point cloud classification. The experimental results demonstrate a 0.9% improvement in accuracy on the ModelNet40 dataset using the proposed improved method. This validates the effectiveness of the enhancement process described in this paper and offers valuable insights for enhancing the performance and application of point cloud classification. [ABSTRACT FROM AUTHOR]

Traffic accident is one of the rarest causes of death in the world. In order to reduce the frequency of traffic accidents, it is of great significance to predict traffic accidents. Due to the dependence of traditional methods, this paper uses deep learning methods to predict traffic accidents, which reduces manual errors. Therefore, this paper first uses Knowledge Graph ( KG ) to analyze the correlation between various factors of traffic accidents, and then proposes a Continuous Wavelet Transform ( CWT ). A model combining three methods of Relation Graph Convolution Neural Networks ( RGCNN ) and Bi-directional Long Short-Term Memory ( BiLSTM ) is used to predict traffic accidents. Through comparative analysis of the model, the accuracy rate, recall rate and F1-score value of the model are higher than other models, and the prediction accuracy rate reaches 99.32 %. Through the parameter analysis of the model, the optimal parameter values are obtained. [ABSTRACT FROM AUTHOR]

Garbage or waste will be a serious problem if they are not managed properly. Sorting garbage is necessary before putting them in the bins. This paper develops an Android application called Lebooh 2 for garbage detection using EfficientDetLite, a Convolutional Neural Network’s architecture for object detection. Lebooh 2 is designed for Indonesian speakers. This paper applies two models: Model 1 and Model 2. Model 1 is trained to recognize 11 types of garbage, i.e., biological, battery, cardboard, metal, glass, plastic, paper, clothes, shoes, electronics, and trash. Model 2 is developed to detect whether the garbage is recyclable or not-recyclable. The experiment using EfficientDet-Lite 3 obtains the mean average precision are around 75% - 77%. User testing involves participants as testers and they collect 400 screenshots of Leboh 2 for solid waste detection. The user testing produces an accuracy of around 78% and 82% for Model 1 and Model 2, respectively. A simple survey of 172 respondents reveals that 22.7% of them have not yet learned how to sort the garbage. [ABSTRACT FROM AUTHOR]

As a clean and efficient renewable energy, wind power is increasingly valued by countries all over the world. However, due to the inherent volatility, intermittency, and uncontrollability of wind power, the secure, stable, and economical operation of the power grid will face unprecedented pressure with wind power integration on a vast scale. Wind power predictions and management are significantly impacted by wind speed. Therefore, it is crucial to propose an accurate, stable, and efficient wind speed prediction method. In this paper, the original wind speed data is denoised using singular spectrum analysis to solve the problem of strong randomness and volatility of wind speed data. Then, the denoised data is processed using wavelet analysis, and the high-frequency and low-frequency sequences are respectively predicted using CNN-BiLSTM and CNN-BP, while the weights are optimized using the PSO algorithm. Finally, the prediction accuracy is further improved by incorporating the error correction module. In the two datasets used for testing, the outcomes of the trial indicate that the wind speed prediction model proposed in this paper performs extremely well in short-term wind speed prediction, superior to other models. The five evaluation indices (MAE, MSE, MAPE, RMSE, and R²) for the first dataset are respectively 0.166, 0.055, 0.05, 0.234, and 0.953. Similar results were obtained for the second dataset. Furthermore, The ablation tests also show that each module suggested in this research is successful in enhancing prediction accuracy. [ABSTRACT FROM AUTHOR]

Lung cancer with the highest mortality rate has attracted public attention. For the difficulty of treating lung cancer increases sharply over time, detecting lung cancer symptoms early on chest computed tomography (CT) is crucial for the subsequent treatment. The number of slices can affect the accuracy of lung cancer examination, so a deep multiple instance learning algorithm was designed and proposed to classify lung cancer effectively. First, feature information is extracted in the patient 3D CT image using the high and low frequency high dimensional features (HLFHD) to balance local detail and global overall information of images. Secondly, to find the decisive features, a sliding recurrent neural network (MSRNN) module is used to take into account the feature variations between slices. The experimental studies in this paper were constructed on two public datasets, namely, CIA and CC-CCII data. Finally, the experimental results show that the proposed algorithm can achieve an ACC of 0.97 and an AUC of 0.99 on the datasets. These results suggest that the proposed algorithm is well suited for lung cancer classification of any number of CT slices, and it can be effectively employed in computer-aided systems to achieve state-of-the-art performance. [ABSTRACT FROM AUTHOR]

Plant diseases and pests represent significant threats to agricultural yields, underscoring the need for precise and prompt identification of pathogens and pests. In this paper, we propose a lightweight convolutional neural network model called Gated Asymmetrical GhostNet (GA-GhostNet), specifically designed for the automatic identification of plant diseases and pests. The model incorporates a Gated Multi-scale Coordinate Attention (GM-CA) module to filter out noise and irrelevant information in the image, while capturing location information of diseases and pests at different scales. The model also utilizes an improved feature extraction module called Asymmetrical Ghost (AG) module to enhance robustness to image flipping, as well as improve feature extraction capabilities. Additionally, the CutMix data augmentation method is employed to improve generalization ability. In extensive experimental evaluations on the IP102 pest dataset, GA-GhostNet achieved impressive results with 63.73% MRec, 67.37% MPre, 65.12% MF1, and 71.90% Acc. Moreover, through transfer learning on the Jute pest, Embrapa disease, and Apple disease datasets, the model outperforms numerous lightweight models, attaining accuracies of 99.89%, 96.97%, and 95.17%, respectively, while having only 3.73 million parameters. GA-GhostNet demonstrates high accuracy and efficiency in identifying plant diseases and pests. [ABSTRACT FROM AUTHOR]

Aiming at the problems of low recognition accuracy and large calculation amount of age estimation, this paper proposes a staged single aggregation backbone network VoVNetv4. In order to improve the feature extraction ability of the network, the Regional One-Shot Aggregation (ROSA) module is proposed in the network for regional encapsulation of the feature module. The classifier adopts the COnsistent RAnk Logits (CORAL) framework, which can solve the problem of inconsistency of classifiers in ordered regression CNNs for multiple binary classification tasks. This paper conducts experimental verification on the Asian Face Age Dataset (AFAD) and compares it with other research methods. The experimental results show that VoVNetv4 can effectively improve the feature extraction and detection ability of the network and improve the detection efficiency of age estimation. [ABSTRACT FROM AUTHOR]

To overcome the challenges of obstacle avoidance for Unmanned Aerial Vehicles (UAVs) in autonomous flights, this paper proposes the Dual Experience Attention Convolution Soft Actor-Critic (DAC-SAC) algorithm. This algorithm integrates a dual experience buffer pool, a self-attention mechanism, and the Soft-Actor-Critic algorithm with a convolutional network. The dual experience buffer pools are used to solve the problem of ineffective UAV training due to the scarcity of successful training data. To overcome the drawbacks of the original Soft Actor-Critic (SAC) algorithm in handling image data, a Convolutional Neural Network (CNN) is applied to reconstruct the actor and critic network, allowing for better image feature extraction and classification. Furthermore, a self-attention mechanism is employed by adding a convolutional self-attention layer to the network. This modification enables dynamic adjustments for the attention weights based on varying input image features, effectively addressing focus-related challenges. Two simulation experiments are performed and the DAC-SAC algorithm achieves a 99.5% success rate in a known environment and an 84.8% success rate when dealing with an unknown environment. These results confirm that the proposed algorithm enables autonomous obstacle avoidance for UAVs even when considering depth images as input. [ABSTRACT FROM AUTHOR]

With the development of deep learning and computer vision, face detection has achieved rapid progress owing. Face detection has several application domains, including identity authentication, security protection, media, and entertainment. Although multi-task cascaded convolutional neural networks (MTCNN) have high accuracy and robustness, the model has the disadvantages of large parameters and computational overhead in the real scene due to the complexity of the real scene and the constraints of hardware facilities. Therefore, the development of an improved network model is crucial. This paper improves the MTCNN model by reducing the number of parameters and the computational overhead and using better model parameters to locate the key points of the face. This model improves the accuracy and robustness of the face age estimation. The WiderFace and CelebA datasets are used for training. The final face detection accuracy reaches 98.7% while simultaneously reducing the number of model parameters to 70% under the same conditions. This model meets the application needs of modern society for face detection and demonstrates the efficiency and accuracy of the improved network model. [ABSTRACT FROM AUTHOR]

Motor imagery (MI) Electroencephalogram (EEG) signal recognition is an important part of brain computer interface (BCI). Nowadays, deep learning has been successfully applied to MI signal classification. However, due to the expensive equipment of EEG acquisition and the limitation of acquisition site that gives rise to the EEG data scarcity, and thus impacts deep learning model classification. Therefore, this paper proposes a novel EEG time-frequency transition method, which uses the continuous wavelet transform adaptive generate EEG time-frequency map. Generative and discriminative adversarial learning framework (GADALF) is proposed to enhance data. The correlation analysis of EEG time-frequency map is carried out by combining event related synchronization (ERS) with event related desynchronization (ERD). Based on the analysis results, the hybrid scale convolutional neural network (HSCNN) is proposed by improving the convolutional neural network. This network combines multi-dimensional convolution with channel attention mechanism to decode the hidden energy information of the a and ß bands for EEG time-frequency map classification. The results showed that the proposed method can achieve higher recognition rates on the laboratory-collected dataset and the dataset BCI competition IV 2b compared with other methods. Finally, this paper designs an intelligent wheelchair control system based on a BCI that verifies the effectiveness of the system by applying the proposedmethod to the system through online experiments. [ABSTRACT FROM AUTHOR]

In recent years, the classification method of motor imagery(MI) electroencephalography(EEG) signals based on deep learning(DL) has become more and more mature in the field of brain-computer interface(BIC). However, most of the studies tend to use a single feature or two associated features when dealing with motor imagery EEG signal classification, while ignoring other features, resulting in poor classification performance. Therefore, this paper proposes a neural network feature fusion algorithm called Multi-feature Fusion Transformer(M-FFT). The network is built based on convolutional neural network (CNN) and Transformer. This method uses CNN and wavelet transform to extract the timefrequency and space-time features, and uses the converter to fuse the three feature domains contained in the two features to establish the information interaction between the three feature domains contained in the two features. Then, the global feature pooling is used to output the feature vector, and finally the softmax function is used to classify the feature vector. In the training process, we use cross entropy as the loss function. Finally, on the brain-computer interface competition IV data set 2a, the average classification accuracy is 85.66%, and the average kappa value is 0.833. The experimental results validate the algorithm performance. [ABSTRACT FROM AUTHOR]

This paper proposes an object detection method called "Down Up Coordinate Attention Fusion model (DUCAF-Net)" to address the challenges of object density and complex backgrounds in aerial images captured by drones. DUCAF-Net integrates coordinate attention at different resolutions, aiming to learn spatial coordinate information from feature maps at various resolutions, enhancing the expression capability of spatial features, and simultaneously reducing the diffusion of features from small dense objects and the phenomenon of feature coupling. DUCAF-Net introduces deformable convolutions and transposed convolutions, and designs an upsampling module to increase the receptive field of feature maps, better capturing the details of target features, thus improving the object detection performance. The AP score on the VisDrone2019 test set is 22.9. DUCAF-Net demonstrates satisfactory performance in medium-scale object detection and also performs well in small-scale object detection. The experimental results show that DUCAF-Net's performance is delightful. [ABSTRACT FROM AUTHOR]

— In recent years, with the advancement of computer processing power and the rapid development of convolutional neural networks, attention models have been widely used in image classification and object detection, significantly improving the performance of networks. However, most existing methods focus on increasing the depth of networks, which inevitably leads to efficiency issues. To strike a balance between performance and complexity, this paper proposes an optimization based on Triplet Attention, introducing multi-scale attention fusion Triplet Attention and self-attention mechanisms. Firstly, the proposed model is integrated into YOLOv5s and tested on VOC2012, achieving a 3.9% increase in mean average precision (mAP). Then, the model is integrated into ResNet50 and tested on CIFAR-10 and CIFAR-100 datasets. The results show that the proposed model achieves a 1.79% improvement in Top-1 accuracy and a 0.13% improvement in Top-5 accuracy on CIFAR-10,a 3.11% improvement in Top-1 accuracy, and a 1.3% improvement in Top-5 accuracy on CIFAR-100. The experimental results demonstrate that the proposed model significantly enhances network performance. [ABSTRACT FROM AUTHOR]

The problem of low recognition accuracy of emotion recognition models is easily caused by interference such as data redundancy and irrelevant features. In this paper, we propose a speech emotion recognition (SER) method based on an attentional convolutional neural network (CNN) bidirectional gated recurrent unit (Bi-GRU) fusing visual information. First, we pretrained the log-mel spectrograms in a ResNet-based attentional convolutional neural network (RACNN) to extract speech features. Second, the CNN-extracted facial static appearance features are fused with speech features using a deep Bi-GRU to obtain speech appearance features. A series of gated recurrent units with attention mechanisms (AGRUs) are used to extract facial geometric features. Then, the hybrid features are obtained by further combining the integrated speech appearance features with facial geometric features, and kernel linear discriminant analysis (KLDA) is used to discriminate them. Finally, the proposed method in this paper obtained accuracies of 87.92% and 89.65% on the RAVDESS and eNTERFACE'05 emotion databases, respectively. The experimental results demonstrate that the method in this paper effectively improved the accuracy and robustness of SER. [ABSTRACT FROM AUTHOR]

In clinical practice, the X-Ray image classification for the presence or absence of lesions by the medical auxiliary diagnosis system is the basic work, and the classification of lesion images and normal images can be completed by reading the medical X-Ray graphics. The actual process is the convolution and pooling layer of CNN to extract the features of medical images, and the full connection layer of CNN is used to classify the local information. The structure of CNN needs to be adjusted repeatedly. This paper focuses on utilizing Differential Evolution (DE) to automatically search for the optimal architecture of CNN. The research idea of this paper is to use the global optimization ability of Differential Evolution algorithm to regulate the structure of CNN. When the termination condition is satisfied, DE-CNN can automatically adjust the parameters and optimize the structure of the CNN. Among them, for mutation operation, we study a new mutation strategy in this paper, which inherits the vector that may solve the next generation, and appropriately accepts the basic elements of the inferior solution to increase the disturbance of the new species. The optimal individual is found through the optimization process of the DE algorithm, and the CNN model composed of all basic elements in the individual and trained on the medical X-Ray image data set is saved. The new mutation strategy generates mutation solution vector, through experimental verification which is beneficial to the optimization of CNN structure, and the accurate value of DE-CNN algorithm processing medical image classification is better than GoogLenet, ResNet34, VGGNet16, DenseNet, Lenet, AlexNet etc network structure. [ABSTRACT FROM AUTHOR]

As one of the important transportation tools for coal transportation, the operating efficiency of belt conveyor is directly related to the production efficiency of coal. However, belt conveyor failures often occur due to belt deviation. Generally speaking, the transportation distance of belt conveyors is relatively long, and the transportation environment is relatively harsh. The use of manual monitoring of the operation status of belt conveyors is inefficient and increases labor costs. The existing belt deviation detection system can play a good role in some conventional environments. However, under complex working conditions such as heavy dust, insufficient light, and camera deviation, existing detection systems will be affected. Therefore, this paper proposes a belt deviation detection system based on deep learning under complex working conditions. The characteristics of the belt and idler are extracted through the convolutional neural network, so as to position the belt and idler. With the characteristic points on the belt and idler as input, a mathematical model of belt deviation is established to calculate the degree of belt deviation, and alarms according to different degrees of deviation. The system has been tested in the actual field, and has a certain accuracy and real-time, which largely solves the problem of belt deviation detection under complex working conditions, reduces the fault frequency of belt conveyor, and thus improves the working efficiency of belt conveyor. [ABSTRACT FROM AUTHOR]

The precise recognition of the sources of voltage sag is the basis and key to formulating a voltage sag governance plan and clarifying responsibility for an accident. Due to the complexity of grid devices and the identification of the mode of power consumption, the conventional approach to identifying voltage sags faces a new challenge. Due to the dependence on accurate voltage sag models, the traditional methods are inadequate for complex problems with multiple uncertainty factors. The random matrix theory based on a data-driven method analyzes data correlation using single characteristic statistics, which is not suitable for the dimensional change of the matrix. It is obvious that this method is not applicable to the position of the source of the voltage sag due to random sags. Therefore, this paper proposes a voltage sag source recognition method based on the combination of random matrix theory (RMT) and a convolutional neural network (CNN). In this method, the CNN optimizes the characteristic statistics of RMT and constructs new characteristic statistics such as the correlation analysis index of the voltage sag source recognition model to avoid the error caused by a single characteristic statistic. First, random matrix theory is used to extract characteristics from historical data, and characteristic statistics under different faults and different data dimension matrices are obtained and then, are used as the input characteristic sets of the CNN. Then, through training extraction, the optimized characteristic statistics that can be applied to a variety of conditions are obtained. Furthermore, a voltage sag source recognition model based on random matrix theory is constructed by using optimized characteristic statistics. The correlation analysis index is obtained to recognize voltage sag sources under different data conditions. Finally, examples are given to verify the feasibility and effectiveness of the proposed method. [ABSTRACT FROM AUTHOR]

Hyperspectral image (HSI) has been widely applied in many fields because of its rich ground target information. Classification is one of the effective means to explore the information in HSIs. However, some HSIs contain shadow areas due to lighting, occlusion and other reasons, usually treated as a separate category or discarded directly. Most image enhancement methods remove the noise in the shadow regions with the loss of many valuable signals. Dynamic stochastic resonance (DSR) can enhance the signal by using noise to generate resonance. Therefore, a double dynamic stochastic resonance (D-DSR) method of shadow enhancement is proposed in this paper by twice successive DSRs in the spatial dimension of HSI. Meanwhile, a parameter-optimized convolution neural network (CNN) model is presented and applied to HSI classification. Firstly, through the horizontal and vertical sampling in two different spatial directions, the shadow area can be enhanced by D-DSR while maintaining the consistency of the enhancement degree of the horizontal and vertical textures. Secondly, the enhanced HSI can be obtained by fusing the shadow area processed by D-DSR with the original image. Thirdly, with parameter optimization, CNN can be used to classify the enhanced HSI. Finally, to evaluate the performance of the proposed method, the comparative experiments are conducted on a real-world HSI. The experimental results show that the proposed approach can help improve the classification and has potential prospects in HSI processing. [ABSTRACT FROM AUTHOR]

As one of the key components in many fields, rolling bearing fault detection is very important. Rolling bearing is in complex and changeable working conditions, so it is challenging to detect its fault. Because the traditional method has weak adaptability in complex and changeable situations, it needs to rely on the opinions of experts more often. Deep learning methods can make up for the shortcomings of traditional methods. Therefore, this paper proposes a method combining continuous wavelet transform (CWT), dual-channel convolutional neural network(DCCNN), and long short-term memory network (LSTM), mainly for fault detection of vibration signals of rolling bearings. Firstly, the vibration signal is denoised by CWT, then the feature of the vibration signal is extracted by DCCNN, and finally, the time series of the vibration signal is extracted by LSTM. Compared with CNN, CWT-CNN, CNN-LSTM, and CWT-CNN-LSTM four models, and analyzed the parameters of the model. The results show that the accuracy of CWT-DCCNN-LSTM model detection is better than other models, and the accuracy rate reaches 99.98 %. [ABSTRACT FROM AUTHOR]

Deep neural networks have gained significant attention in text-independent speaker recognition tasks. However, due to the fixed parameters of traditional static convolutional neural networks, they cannot flexibly capture the variation in phonemes that are integral to speech sentences. To address this limitation, this paper proposes a channel-space attention-based dynamic convolutional speaker recognition method. This method employs dual-attention mechanisms to generate dynamic convolutional kernels, which improves the capture of phoneme variation information between different inputs in the speech signal. We conducted experiments using the TIMIT dataset to evaluate the proposed method's effectiveness in various network frameworks. Our results show that the best performance can be achieved when dynamic convolution is generated using four static convolutional kernels. Specifically, in the ResNet-34 framework, the Equal Error Rate (EER%) of the proposed method is improved by 31.1% over the static convolutional method CNN and by 20.3% over the single-attention dynamic convolutional method (DynamicConv). Additionally, the performance of the proposed method is enhanced in all other network frameworks. These findings demonstrate the effectiveness of the proposed method and the importance of considering phoneme variations in speaker recognition systems. [ABSTRACT FROM AUTHOR]

The process of designing Convolutional Neural Networks (CNN) architecture has been automated using Neural Architecture Search (NAS) algorithms, however, these algorithms pose a significant computational demand due to the large search space of possible architectures. This paper proposes a two-stage algorithm to address the computational resource demand of NAS by generating architectures directly from the dataset. In the first stage, the complexity tree is generated from the dataset based on the complexity of individual images. The image complexity is determined by color variation and changes in color intensity. This complexity tree is then utilized in the second stage for designing the CNN architecture. As the complexity tree solely necessitates one-time generation for the dataset, our proposed model does not have significant computational demands. In the experiment that was performed utilizing standard benchmark datasets, the generated (CNN) models attained an accuracy of 97.35% on CIFAR-10 and 72.57% on CIFAR-100 datasets. Our proposed model was compared with state-of-the-art models and yielded a notable improvement in accuracy. Furthermore, it outperformed all other models in terms of computational resource requirement. [ABSTRACT FROM AUTHOR]

This paper proposes a framework for creating trained models for object detection in digital documents using convolutional neural networks (CNNs). CNNs are artificial networks that are capable of learning at different levels of abstraction and are structured to resemble neurons in the primary visual cortex of a biological brain. The X101-FPN model, which is based on pre-trained neural networks exposed within the collections offered by the Detectron 2 library, is selected for this research. The model contains a hierarchical structure of 101 CNN layers, a residual neural network, and a pyramidal network of functions. This research presents a case study for the detection of QR codes using Artificial Intelligence (AI) techniques with convolutional neural networks, based on successful case studies using different Deep Learning techniques. The pre-trained model is trained on different types of preprocessed images and with various approaches to measuring correct accuracy and performance against proposed scenarios. Finally, various tests are performed to conclude the conditions and best versions of training for QR code detection. [ABSTRACT FROM AUTHOR]

It is important to diagnose lung nodules as early as possible in order to improve the cure rate of lung cancer patients. The CT technology is currently the most prevalent method of diagnosis and treatment in clinical medicine. However, the amount of data generated by CT diagnosis and treatment is increasing exponentially. In order to segment lung nodules from medical images, it is crucial to employ computer and artificial intelligence technology. The deep learning model proposed in this paper is MRBU-Net-WD model, a more effective and enhanced version of the U-Net. MRBU-Net-WD is distinguished by combining residual 3D convolution modules with multiscale densely connected modules. Moreover, In addition, in order to avoid the phenomenon of gradient disappearance when the network depth of the model is increasing. The Bi-FPN is therefore introduced to enhance the feature maps of the network at each depth, as well as to provide an effective fusion of features across depths. The weighted Dice loss function has a significant improvement for the pixel imbalance between lung nodules and background images in lung CT images. To test the proposed model, the LUNA-16 dataset has been extensively trained and evaluated. The performance of the comparison methods generally outperforms that of the original U-Net model and several other newly proposed models. [ABSTRACT FROM AUTHOR]

Although methods based on deep learning have achieved remarkable success in the field of point cloud analysis, the recognition accuracy of point cloud analysis is still far from practical applications. In this paper, we propose a novel end-toend attentive enhanced convolutional neural network for point cloud analysis, named AECNN, which can analyze point clouds with high accuracy. The key component in our method is the attentive enhanced convolution (AEConv) module. It designs an attention mechanism to enhance the features of each group through the interaction of information between groups, so that the feature expression is more sufficient and richer. Based on AEConv, we further design an attentive enhanced convolutional neural network (AECNN) for 3D model classification and segmentation. AECNN continuously abstracts the entire point cloud by stacking downsampling and AEConv to obtain feature descriptors that can represent the entire point cloud, and finally uses a classification head and a segmentation head to complete the task of point cloud classification and segmentation, respectively. Experiments are conducted on the point cloud recognition and large-scale point cloud outdoor segmentation datasets, i.e., ModelNet40 and vKITTI, respectively. Extensive experiments show that our AECNN can achieve high performance for classification and segmentation tasks, which exceeds previous point cloud recognition and segmentation algorithms. [ABSTRACT FROM AUTHOR]

Action recognition in real-world videos is difficult because of factors such as scenery muddle, scale alter, dynamic standpoint, and sharp motion. This paper proposes a novel two-stream architecture fusing static and dynamic features to recognize human actions in videos. Firstly, the original image (single frame and optical flow fields) is extracted by Convolutional Neural Network (CNN) to obtain feature maps. Secondly, we extract the obtained feature maps via a 3x3 convolution over all the neighbor features, leading to a static representations of features. Then we concatenate these static features with input feature maps to produce the dynamic attention matrix through two 1x1 convolutions. All of the generated feature maps are then aggregated using the learnt attention matrix, producing a dynamic representation. Thirdly, we take the interaction of the static and dynamic presentations as final outputs. Finally, we utilize Long Short-Term Memory (LSTM) to catch time sequence information among dense optical flow. The experimental results on the three hard datasets UCF101, HMDB51, and Kinetics400 have shown that the method works better than other state-of-the-art methods. [ABSTRACT FROM AUTHOR]

Scanning Electron Microscopy (SEM) image plays a significant role in industrial, medical, and research fields. However, image defects, including existing noise will degrade the quality of the original image. In order to denoise the image, noise estimation is crucial. In the past, several noise estimation methods are designed, but they are not effective in estimating the noise level of the SEM images. Therefore, this paper presents a novel method to effectively classify noise variance of SEM images through a deep learning algorithm. The images are classified based on their respective noise variances. Before the classification process begins, the original SEM images are added with 4 categories of White Gaussian noises. To categorize the noise level of the image, a complex Gaussian-Noise Convolutional Neural Network (GN-CNN) method with encoder layer, convolutional layer, attention layer, decoder layer, and decision layer are adopted. In this study, a total of 1200 image data from different variety and complexity (720 images for training, 240 images for validation and testing respectively) are used. The experimental study shows the capability and reliability of the developed system in classifying noise variances of SEM images with the F1 score of 93.97% and testing accuracy of 93.8%. It has outperformed other baseline deep-learning models, including Long Short-Term Memory (LSTM), Residual Network 18 (ResNet18), and Residual Network 34 (ResNet34), and pure Convolutional Neural Network (CNN). Thus, the designed network can impressively surpass human-eye performance in noise variance categorization. [ABSTRACT FROM AUTHOR]

Parkinson's disease (PD) is a long-term disease that mainly influences the central nervous system and thus affects movement, such as inability to move rigidity, and tremors. So, analysis of patients' movements under control, especially handwriting, is a helpful way to diagnose Parkinson's disease. Diagnosis, as the first step in medical practice, is critical to clinical decision-making. This paper uses multiple computational intelligence classification techniques such as Decision Tree, Naive Bayes, Support Vector Machine, and Random Forest to investigate the existence of the PD. Also, Convolutional Neural Network (CNN) and the Best First (BF) strategy are used as feature extractors. These techniques are applied over both Meander and Spiral data and some selected traits derived from the patient's handwriting during the handwritten exam. The available HandPD dataset has been used with both its images and selected attributes. The CNN is used for the feature extraction process across the images of the used dataset. Whereas, the BF search strategy is used to extract features based on the changes between the handwritten trace and the exam template features combined with instances resampling. Compared with other well-known diagnostic systems, the proposed one has the highest recognition rate. [ABSTRACT FROM AUTHOR]

In recent years, the Corona Virus Disease 2019 (Covid-19) epidemic has raged around the world, with more than 500 million people diagnosed. Relevant medical research and analysis results on Covid-19 indicate that wearing masks is an effective method to prevent and restrain virus transmission. Mask detection stations have been set up in hospitals, railway stations, schools, where there is large crowd flow, but results are not as good as expected. In order to ameliorate pandemic preventing and control measures, a mask wearing detection algorithm YOLOv3-M3 was designed and proposed in this paper. The algorithm can effectively detect people without mask, while consequently reminding them. Firstly, we substituted the feature extraction network of YOLOv3 with MobileNetv3, a lightweight convolutional neural network. Secondly, we utilized K-Means++ to substitute the original ground truth clustering algorithm to improve prediction precision. In addition, the bounding box regression loss function was revised as CIoU loss function. This loss function solves the issues of overlapping between the ground truth and the anchor box, which has increased the training speed. After experiments, the precision of YOLOv3 algorithm on mAP 0.5 and mAP 0.75 is 93.5% and 71.9%, respectively. Elevating 3.1% and 2.6%, respectively, higher than that of YOLOv3 algorithm, and it was superior to SSD, SSD Lite, YOLOv3-Tiny and other one-stage object detection algorithms. The detection speed can reach 13.6 frame/s, which has met the requirements of pandemic prevention and control in most places and can be deployed on terminal devices for object detection. [ABSTRACT FROM AUTHOR]

Electronic medical records (EMR) record the whole process of patients' diagnosis and treatment in the hospital, which contains a lot of valuable information. Through this information, medical services can be provided to patients in more timely and convenient manner. Accurately identifying patients with similar diseases based on EMR is the key to personalized healthcare. Most patient similarity studies mainly utilize discrete medical entities (drugs, procedures) embedded as patient feature representation. But these structured data could be either incomplete or erroneous which has a significant impact on the final patient representation. And the previous studies rarely considered the structural and semantic information existing between medical entities. Therefore, we propose a patient similarity framework based on a medical attributed heterogeneous graph convolution network, named AHGCN-PS. Firstly, the framework leverages the patients' medical entity and incorporates the patients' medical text as the attributes of patients to obtain more integral patient information. Then, we construct a medical attributed heterogeneous information network from EMR, capturing the structural information in the network and the hidden semantic information between different nodes by selecting different meta-paths. Then, we adopt a graph convolutional neural network and a semantic attention mechanism to aggregate node neighbor information and meta-path semantic information. Finally, this paper uses the obtained patient node feature representation for patient similarity calculation. We use the real-world ICU patient dataset MIMIC-III to evaluate the experimental performance of AHGCN-PS, the experimental results demonstrate the effectiveness and feasibility of the patient similarity framework. [ABSTRACT FROM AUTHOR]

Physical disability or arm paralysis is a common symptom for the post-stroke survivor. The upper limb rehabilitation is introduced to improve the motor ability of the upper limb and recovery from stroke. However, the recovery rate of the motor ability upper limb is based on physical condition and therapy performance of a patient (subject). The rehabilitation may require more manpower at a center and is time consuming for a physiotherapist to monitor a patient during rehabilitation without the use of technology. The purpose of research in this paper is to evaluate the condition of subjects using a deep learning model with biosignal devices after virtual reality (VR) upper limb assessment. Fifteen control persons and fifteen post-stroke patients have performed two games under VR upper limb assessment, namely, Touch the Ball, and Stack the Cube. The patients were equipped with an electroencephalogram (EEG), electromyographic (EMG), and electrocardiographic (ECG). The measurements were taken before, during, and after the assessment. A common practice in data handling is that all EEG, EMG and ECG signals are pre-processed to remove noises or to condition data. In this work, all three raw biosignals were collectively represented as images. They were used to train deep learning models (namely, convolutional neural network and long-short term memory) of which the models were used to evaluate the condition of a subject. The classification performance of the deep neural network in classifying the biosignals is highly accurate and precise. [ABSTRACT FROM AUTHOR]

This paper studies the problem of visual odometry based on a deep recurrent convolutional neural network. A new visual odometry algorithm based on dual-stream convolutional neural networks with long short-term memory is proposed. The color stream of the convolutional neural network acquires the color features in the RGB image. The depth stream acquires the contour features in the depth image, generates fusion features through the feature fusion unit, and finally predicts the pose at the current moment through autonomous sequential modeling using recurrent neural networks. Experimental validation on the TUM dataset showed that the method introduces contour features into the system through a dual-stream architecture of neural networks, which provides higher accuracy and robustness compared to other convolutional neural network-based visual odometry systems, especially in the presence of motion blur and poor lighting. [ABSTRACT FROM AUTHOR]

The analysis of patients' Electronic Medical Records (EMR) data through deep learning is helpful for disease diagnosis, disease risk factor assessment, disease risk prediction, etc. The medical texts of patients in EMR contain rich information since the medical text is usually stored in the form of documents, some redundant and irrelevant content may be generated, which is unimportant to patients diagnosed with diseases, and they hurt the overall training efficiency and effectiveness of the model. At present, in the research on automatic disease diagnosis, the structured and unstructured medical text data of patients are rarely considered simultaneously, and they cannot be effectively fused. This paper effectively uses the structured and unstructured medical text data of patients in EMR and proposes a novel disease diagnosis deep framework DiseaseNet via fusing medical record summarization. Firstly, the framework generates the medical record summarization of the patient's medical text, to obtain fine-grained and more relevant patient information. The structured data of patients and medical record summarization use the BERT model for fusion and embedding, and then the deep features of patients were extracted by Bidirectional Long Short-Term Memory (BiLSTM) and Convolutional Neural Network (CNN), respectively. BiLSTM integrates the output of the self-attention layer with the feature representation of the CNN output. Many experimental results on the real intensive care patient dataset MIMIC-III showed that DiseaseNet can significantly improve the performance of disease diagnosis by fusing the features of patients' structured data with the medical record summarization generated by unstructured medical text. [ABSTRACT FROM AUTHOR]

Malaria is a deadly disease spread by the bite of a female anopheles mosquito infected by the plasmodium parasite. Despite thorough research in the medical field, the pervasiveness of this deadly disease is increasing globally. More than 1.5 billion cases have been averted in the last two decades, with 7.6 million deaths. The traditional microscopic method of detecting whether a person is infected or not is time-consuming. Its accuracy depends on smear quality and individual expertise in counting and classifying parasitized and uninfected cells. Computer-aided diagnostic methods based on image processing and Machine learning (ML) use hand-generated features and need expertise in texture, morphological aspects, and analysis of a region of interest. Convolution Neural Network (CNN) based methods are superior to the traditional ML-based approach. They are highly scalable and give the best results with end-toend feature extraction and classification. Hence, this research intends to design and develop a reliable framework for automatic malaria parasite detection using Deep learning (DL), which could serve as the best effective aid. This research paper evaluates the performance of pre-trained CNN-based DL models like AlexNet, ResNet50, and VGG19 as feature extractors for analyzing infected and non-infected cells. Statistical results show pre-trained CNN models serve as the best feature extractor tool for this purpose. Features extracted using VGG19 are proven to be more efficient than the ResNet50 and AlexNet to detect the presence of malaria parasites with a training accuracy of 95.28% and a testing accuracy of 93.89%. This transfer learning and CNN model are integrated further with a web application for accurate and automatic malaria parasite detection. [ABSTRACT FROM AUTHOR]

The existing large scale point cloud recognition and segmentation methods based on deep learning only focus on recognition accuracy. They always ignore the memory and computational complexity limitations in practical applications (eg autonomous driving applications). A large-scale 3D model recognition and segmentation method based on lightweight channel-aware dynamic convolutional neural network is proposed in this paper. Firstly, a channel-aware module is constructed to adaptively predict the most important input feature channels of point clouds. It reduces the computational costs while ensuring the integrity of the original network structure and improves the recognition ability of the model. Secondly, by constructing a feature channel attention predictor, we design a lightweight channel-aware dynamic convolution. It can dynamically prune some unimportant feature channels of point cloud and reduce redundant calculation costs. Finally, a channel-aware dynamic dynamic convolutional neural network is constructed to efficiently recognize and segment large-scale point clouds with low complexity. Experiments are carried out on the 3D model recognition dataset ModelNet40 and the large-scale point cloud indoor segmentation dataset S3DIS and the outdoor segmentation dataset vKITTI, respectively. Experiments show that our method not only has higher recognition and segmentation accuracy than existing methods, but also has lower computational and memory complexity. [ABSTRACT FROM AUTHOR]

Computer-Aided Diagnosis (CAD) has become a requisite and fundamental part in medical detection and diagnosis. As an indispensable component of CAD, the lung fields segmentation is critical for further analysis. Nowadays, it is well known that the methods of deep convolutional neural networks (DCNNs) have achieved outstanding performance in medical image segmentation. Especially the U-Net and its extensions have obtained promising accuracy on medical image segmentation. However, due to the superimposed regions in lung fields and varied shapes among different individuals, it is difficult to detect and segment the boundaries precisely. Besides, insufficient training dataset may results in poor generalization ability of the networks. To address these problems, this paper uses the standard U-Net as its backbone, and optimizes and improves it effectively. The proposed method can segment the lung fields in chest X-ray images automatically, which integrates U-Net, Bi-directional ConvLSTM (BConvLSTM), Squeeze and Excitation (SE), and fully connected CRF into one framework. In the proposed architecture, a single U-Net network is employed as the backbone. Then, the BConvLSTM is employed to concatenate the features extracted from the encoder and the corresponding decoder. In this way, the skip connection of the U-Net structure is replaced by BConvLSTM. Furthermore, the SE modules are embedded in the decoder to recalibrate the channel-wise features. Besides, the fully connected CRF is used to further refine the initial segmentation contours, which can fully consider the mutual information between pixels in the original image. Compared with diverse lung fields segmentation algorithms on JSRT and MC datasets, the superiority, effectiveness and robustness of the proposed method are verified. [ABSTRACT FROM AUTHOR]

Plant diseases remain a threat to global food supply, as it causes unimaginable loss of food and revenue. Intelligent mobile plant disease diagnostic system has however become valuable due to its usefulness for the early diagnosis and detection of plant diseases using leave images even when there are no availability of competent and adequate experts in such locations. The objective of this paper is to develop an intelligent mobile plant disease diagnostic system that runs on a smartphone. The diagnostic system is based on NASNet-Mobile, a lightweight convolutional neural network (CNN) architecture using the images of the plant leaves for plant disease diagnosis. A mobile application is developed for both android and iOS smartphones to capture the plant leaf images. The system runs on a web service that gets diagnosis from the CNN model. The plant leave images captured using the developed mobile application are sent through the web service and the recognition of the plant disease is achieved using NASNet-Mobile CNN model. The proposed NASNet-Mobile CNN model plant disease diagnostic system achieved an accuracy rate of 99.31%. [ABSTRACT FROM AUTHOR]

In semiconductor manufacturing, surface defects on wafers must be classified accurately for better yield management. To manage the increasing chip demand in speed and scale, automatic defect classification (ADC) system has been introduced. Most existing ADC systems utilize machine learning-based algorithms that require manual feature extractions and manual intervention such as human-based classification for accuracy and consistency. These methods are labour-intensive, unreliable, and highly prone to human error. Therefore, by leveraging on deep learning technologies, this paper proposes DLADC - an ADC system using a deep convolutional neural network (CNN) architecture for detecting and classifying semiconductor wafer surface defects. The proposed system takes Scanning Electron Microscope (SEM) images as input and outputs the defect's class and location. The proposed system also sub-classifies particle-type defects into various sizing groups. Identification of defect types that occurred on wafer surfaces allows for better defect root cause analysis, and the additional information of defect size further serves as an essential indication of the origin of machine failure. The proposed DLADC promotes 2x time saving while achieving an improved accuracy of 93.69% based on experimental results with a real semiconductor defect dataset. Not only does DLADC outperforms the 70% classification performance of trained operators, but it also surpasses the 90% classification performance of industrially pragmatic defect classification. [ABSTRACT FROM AUTHOR]

Bearings are to machinery like joints to humans, which means bearings are an integral part of the machinery. Recently, with the rapid development of sensors and artificial intelligence, a data-driven machine learning-based model for bearing remaining useful life (RUL) has become a powerful tool in academia and industry. However, there are also some existing problems, for example, the pre-set state-related threshold need to be manually determined, the bearing object being studied is under a single operating condition, RUL estimation in the entire life span in many studies is ignored and only a few samples as observed data to predicted and tested, which are all not conducive to the achievements transformation in actual engineering projects. In this paper, a novel generalized framework based on neural networks is proposed to solve these obstacles. Firstly, convolutional neural networks (CNN) and deep encoder (DE) are respectively used to extract features and reduce dimensions, obtaining dense and low dimensions health indicators (HIs). Secondly, Gated recurrent units (GRUs) are used to investigate timeseries information from dense and low dimensional HIs. Thirdly, the proposed method was verified on two platforms and experimental results show that the proposed method is effective and general, and superior to other baselines. [ABSTRACT FROM AUTHOR]

In this paper, we study the highway risk classification and accidents prediction using machine learning models. First, highway facilities data are collected in the form of numbers and text. The facilities information is then processed by encoding and digitalization. Second, stochastic optimization algorithm is used to select features for modeling the risk. Supervised information is provided by human expert. Third, two types of customized convolutional neural networks are introduced to highway road risk modeling. Different feature combinations are evaluated. Finally, experiments are carried out for risk classification, accidents number regression and feature set evaluation. The results show that the proposed highway risk model is effective and may contribute considerably to road risk management. [ABSTRACT FROM AUTHOR]

Wearing face masks in public spaces has become an essential step to prevent the spread of COVID-19. This step poses some challenges to conventional face recognition due to several reasons: 1) the absence of large real-world masked face recognition dataset, and 2) the loss of some visual cues due to the occlusion by the face masks. To address these challenges, this paper presents a real-world masked face recognition dataset that consists of 80500 masked face images of 161 subjects, referred to as MFRD-80K dataset. Every subject contributes 500 masked face images, which are then partitioned into 60:20:20 for train, validation and test. Subsequently, we conduct some benchmark studies to evaluate the performance of the existing face recognition and classification methods on the MFRD-80K dataset. The methods include k-Nearest Neighbour, Multinomial Logistic Regression, Support Vector Machines, Random Forest, Multilayer Perceptron and Convolutional Neural Networks. Since the parameter settings affect the performance of each method, a grid search is performed to determine the optimal parameter settings. The empirical results demonstrate that Convolutional Neural Network achieves the highest test accuracy of 97.16% on MFRD-80K dataset. [ABSTRACT FROM AUTHOR]

In the world, lung cancer is the third most dreadful cancer. Thus, detection of lung cancer cells at early stage is a challenge. The symptoms of lung cancer do not appear in earlier stages which causes high death rates when compared with other types of cancer. In lung cancer detection, image processing algorithms have shown great performance in various high-end tasks. In this paper, different classification methodologies used for the prediction of lung cancer in its early stage are explained. Machine learning techniques are used to identify whether lung tumors are malignant or benign. Machine learning approaches such as: Convolutional neural network (CNN), Support vector machine (SVM), Artificial neural network (ANN), Multi-Layer Perceptron (MLP), K-Nearest Neighbor (KNN), Entropy degradation method (EDM) and Random Forest (RF) are discussed in detail and their performance is evaluated in terms of accuracy, sensitivity and specificity. In this analysis, CNN approach using small dataset shows best result with 96% accuracy compared to other methodologies and EDM shows the worst accuracy of 77.8%. [ABSTRACT FROM AUTHOR]

Hand gesture is a communication tool that allows messages to be conveyed, actions to be performed through hand gestures. Hence, it has the ability to simplify communication and enhance human computer interaction. This paper proposed Wide Residual Network for static hand gesture recognition. WRN improves feature propagation and gradient flows by utilizing shortcut connection in residual block. Wide residual block further improves upon residual block by increasing the width of the network and improving feature reuse, and thereby allowing the depth of the network to be trimmed and fewer trainable parameters to be learned. The network is experimented on three public datasets and compared with existing convolutional neural network (CNN) variants proposed for static hand gesture recognition. Experimental results show Wide Residual Network outperforms the existing CNN variants proposed for hand gesture recognition. [ABSTRACT FROM AUTHOR]

Road traffic monitoring is important for intelligent transportation, and researchers have begun to focus on the detection of traffic events using acoustic information. In this paper, we apply model fusion to traffic acoustic event classification. First, an improved, two-channel convolutional neural network (CNN) model is proposed as the weak classifier for constructing the fusion model. The mel-cepstral feature and its first-order and second-order difference are selected as the input features. Six different input features are constructed after signal preprocessing and segmentation. Second, after training six different CNN models, the voting method and support vector machine (SVM) stacking method are used to construct the final fusion model. Experimental results demonstrate that the detection rate of traffic acoustic events reaches 95.1%, which is higher than that of traditional traffic detection algorithms. [ABSTRACT FROM AUTHOR]

Facial expression recognition(FER) is an significant branch of artificial intelligence and plays an indispensable role in optimizing and improving the experience of HCI. Although much progress has been made in this field, there is a need to optimize the accuracy rate of facial expression recognition and the network’s generalization ability. This Paper brings forward a new end-to-end region attention and multitask learning network (RAMN) for FER. The method learns the importance of the facial features and combines different numbers of regional features obtained from the neural network for FER. A region loss function (R-loss) and collaborative loss function (C-loss) are used to jointly optimize RAMN. The R-loss can learn the most discriminative region in the facial images, and the C-loss substantially reduces the intra-class differences and increases the intra-class differences in the facial features. Finally, experiments on two uncontrolled datasets (FER2013, RAF-DB) and two controlled datasets (CK +, Oulu-CASIA) to verify the superior performance ability of the proposed method. A contrast with existing methods demonstrates the superiority of the RAMN method for FER. [ABSTRACT FROM AUTHOR]

To improve the classification accuracy of the massive amount of image data, we propose a novel extreme learning machine classification model for image classification which based on convolutional neural networks. Based on the framework of Alex Net network, the hash function is constructed as the hidden layer between image representation and classification output in convolutional neural network. At the same time, the extreme learning machine is introduced at the end of the network layer, which saves the classification time, improves the classification efficiency and further improves the feature expression ability of the network. Through comparative experiments in the standard databases MNIST and CIFAR-10, the effects of various improved methods under different situations are analyzed. The experimental results show that the proposed convolutional neural network image classification method based on the extreme learning machine improves the average precision by 3%-31% compared with other image classification methods in this paper. [ABSTRACT FROM AUTHOR]

Convolutional neural networks have received extensive attention and been applied to regular data like image processing recently. This paper applied the convolutional neural network to calculate the ontology similarity, and proposed an ontology learning algorithm based on CNN. Meanwhile, a sports ontology covering most sports concepts was constructed, and the detection of the new ontology learning algorithm is proved to be effective for P.E. ontology. [ABSTRACT FROM AUTHOR]