Search

Deep Learning and Machine Learning are becoming more and more popular as their algorithms get progressively better, and their use is expected to have the large effect on improving the health care system. Also, the pandemic was a chance to show how adding AI to healthcare infrastructure could help, since infrastructures around the world are overworked and falling apart. These new technologies can be used to fight COVID-19 because they are flexible and can be changed. Based on these facts, we looked at how the ML and DL-based models can be used to deal with the COVID-19 pandemic problem and what the pros and cons of each are. This paper gives a full look at the different ways to find COVID-19. We looked at the COVID-19 issues in a systematic way and then rated the methods and techniques for finding it based on their availability, ease of use, accuracy, and cost. We have also shown in pictures how well each of the detection techniques works. We did a comparison of different detection models based on the above factors. This helps researchers understand the different methods and the pros and cons of using them as the basis for their research. In the last part, we talk about the open challenges and research questions that come with putting these techniques together with other detection methods. [ABSTRACT FROM AUTHOR]

Academic performance of students in most places around the world is determined by their presence in the classroom. Lecture attendance control at a university is a traditional process in which student engagement in classes is fostered by means of an institutional policy. In a conventional way, professors have employed different techniques to monitor the influx of students in the classroom, as well as their punctuality and history log. Nevertheless, conventional (paper-based) and modern attendance registration approaches (some using worksheets and computational information systems) often lack either promptness, non-impersonation ability, simplicity, reliability, effectiveness, scalability, acceptance, mobility, or affordability. The aim of this paper is to present a new proposal based on a common fingerprint process and in a new speaker recognition configuration by using a Raspberry system for student attendance administration purposes in order to provide most of these advantages. Thus, the propounded scheme uses ordinary fingerprint verification (a binarized 8-bit grayscale image processing) and speech recognition in the task of text-independent closed-set speaker identification using convolutional neural networks by means of previous enrollment and training process using as input a new approach of spectrograms images configuration. The presented system obtained an accuracy of 97.5% and 97.49% for the fingerprint and speech-based task respectively. These results are comparable with the corresponding state-of-the-art environments. The main findings obtained show a flexible structure with growth capacity, dynamism and prompt response due to the set of user assistance reports and the student control techniques. [ABSTRACT FROM AUTHOR]

The traditional way of football training could be subject to some factors such as the field, and the problems and scenes in the training can not be recovered fully. Thus, this paper proposes a football training method based on the improved tiny-yolovs3 model and the virtual reality (VR). Firstly, the paper makes use of the improved tiny-yolov3 model to detect the football in motion. Then, the paper uses the binocular camera to get the coordinates of the football. Finally, the paper reproduces the position of football in the virtual reality environment. The experimental results show that the football training method is feasible. [ABSTRACT FROM AUTHOR]

In order to improve the accuracy of CNN (convolutional neural network) in image classification, an enhanced Inception-ResNet-v2 model based on CNN is designed through the comparative study and analysis of the structure of classification model. This paper proposes to use multi-scale depthwise separable convolution to replace the convolution structure in Inception-ResNet-v2 model, which can reduce the amount of model parameters and extract features under different receptive fields. At the same time, this paper establishes channel filtering module based on global information comparison to filter and join channels, which realizes the effective extraction of features. Finally, through data enhancement, batch normalization and learning rate adjustment, the effect of the model used in this paper is better than most other models in each dataset, and the accuracy rate can reach 94.8%. [ABSTRACT FROM AUTHOR]

YouTube video recommendation algorithm plays an important role in enhancing user engagement and profitability (or monetization), yet it struggles to deal with disturbing visual content. Real-time analysis using deep learning techniques can play a vital role to identify and filter the disturbing content in online videos. In this paper, we propose an end-to-end trainable two-stream deep learning framework that analyzes and classifies the different categories of disturbing content embedded in child-friendly cartoon videos on YouTube and YouTube Kids platforms. At first, the model extracts the static and motion features from videos through an individual pretrained convolutional neural network (CNN) i.e., EfficientNet-B7. In the next phase, the extracted features are processed by spatio-temporal bidirectional long short-term memory (BiLSTM) network to capture the long-term global temporal dependencies of videos. The learned video representations are forwarded to the classifier for multiclass classification into six categories. Three different types of fusion strategies are investigated to combine the spatial and temporal streams. The evaluation of these methods is performed through extensive experiments on a customized large-scale dataset–YouTube cartoon content filtering (YT-C2F) dataset. The spatio-temporal EfficientNet-BiLSTM network with feature-level fusion displays the best results (f1-score = 0.9316) and shows the efficiency of using a two-stream network compared to single-stream baseline methods. This paper makes two-fold contributions. First, using only static information from videos in a deep learning framework is less effective than using both static and motion information. Second, the feature-level fusion of two streams of networks generates better classification results than early and late fusion techniques. The performance comparison of the proposed model with existing state-of-the-art techniques confirmed the competitive or even better results of our framework. [ABSTRACT FROM AUTHOR]

In recent years, there have been frequent jellyfish outbreaks in many offshore areas worldwide, which have severely affected marine fishery production, coastal tourism, coastal industrial cooling water systems, and marine ecology. Achieving the monitoring of jellyfish plays a vital role in solving the problems mentioned above. However, the research on jellyfish is still in the primary stage. Jellyfish detection technology based on deep learning is gradually being applied to jellyfish detection due to its high efficiency and accuracy, but it is not systematic enough and can identify few jellyfish species. So this paper studies a jellyfish detection algorithm based on deep learning. Based on convolution neural network theory and digital image processing technology, 10 species of jellyfish and fish are detected. Because the quality of underwater images affects the detection accuracy, to further improve the accuracy of the detection algorithm, this paper studies the underwater image processing algorithm. Experimental results show that the image quality is better after applying the three algorithms of dark channel prior algorithm, quadratic combining gray world and perfect reflection algorithm, and contrast-limited adaptive histogram equalization algorithm, which is more conducive to detection. Then, deep learning theory is applied to classify jellyfish. By comparing the AlexNet and GoogLeNet backbone networks' classification results, the accuracy of the jellyfish classification task based on the GoogLeNet backbone network is 96.21%, which is better than AlexNet. Finally, the Faster R-CNN algorithm is used to detect jellyfish, and its detection performance is analyzed based on the two backbone networks mentioned above. The results show that the Faster R-CNN algorithm based on GoogLeNet has a higher detection accuracy in the jellyfish detection task, with an average detection accuracy of 74.96%. In addition, we set up a new jellyfish data set, which includes 25,344 images. The images were divided into 11 species, including 10 species of jellyfish and one fish species. The paper's research lays a theoretical and technical foundation for the subsequent construction of a real-time monitoring system for underwater jellyfish optical imaging, plays an important role in the development of jellyfish monitoring technology, and provides valuable information for marine biologists. [ABSTRACT FROM AUTHOR]

The occurrence of apple diseases has dramatically affected the quality and yield of apples. Disease monitoring is an important measure to ensure the healthy development of the apple industry. Based on a residual network (ResNet50), this paper proposes an MSO-ResNet (multistep optimization ResNet) apple leaf disease recognition model. By decomposing the convolution kernel, updating the identity mapping method, reducing the number of residual modules, and replacing the batch normalization layer, the identification accuracy and speed of the model are improved, and the number of model parameters is reduced. The experimental results show that the average precision, recall, and F1-score of the proposed model for leaf disease identification are 0.957, 0.958, and 0.957, respectively. The parameter memory is 14.77 MB, and the recognition time of each image is only 25.84 ms. The overall performance of the proposed model was better than that of the other models. The proposed model in this paper has high recognition performance and strong robustness and can provide critical technical support for the automatic recognition of apple leaf diseases. [ABSTRACT FROM AUTHOR]

Object detection is one of the most important and challenging branches of computer vision, whose main task is to classify and localize objects in images or videos. The development of object detection technology has been more than 20 years, from the early traditional detection methods to the current deep learning methods, the improvement of object detection accuracy and speed stems from the rapid development of deep learning technology. Traditional object detection techniques have many limitations, and using convolutional neural networks as the main framework for object detection can efficiently extract features and reduce the complexity of manual feature extraction. To comprehensively and deeply understand the development status of object detection, based on the research of domestic and foreign related literature, this paper reviews the research background of object detection, introduces the problems and dilemmas faced by traditional object detection algorithms, and analyzes the current mainstream object detection algorithms. This paper mainly carries out the relevant algorithms from three perspectives: Anchor-based, Anchor-free, and Transformer-based, and summarizes their structure, performance, advantages, and disadvantages in detail. This paper also introduces the commonly used datasets and related performance evaluation indexes for object detection, as well as the applications of object detection in industrial, transportation, medical, and other fields. According to the current research hotspots and the development trend of related technologies, the future research direction of object detection is prospected. [ABSTRACT FROM AUTHOR]

This paper studies a deep learning-based plant detection algorithm. Based on convolution neural network theory and digital image processing technology, 9 species of planes are detected. In order to further improve the accuracy of the detection algorithm, the paper proposes to improve the quality of the images. The experimental results show that the image quality is better after applying the following three algorithms the black channel prioritization algorithm, the quadratic combination algorithm of gray world and perfect reflection, and the contrast-limited adaptive histogram equalization algorithm, which is more conducive to the detection. Then, deep learning theory is applied to classify the plants. Comparing the classification results of AlexNet and GoogLeNet backbone networks, the accuracy of the jellyfish classification task based on GoogLeNet backbone network is 96.21%, which is better than AlexNet. Finally, the Faster R-CNN algorithm is used to detect plants and its detection performance is analyzed based on the two backbone networks mentioned above. The results show that the Faster R-CNN algorithm based on GoogLeNet has better detection accuracy in the jellyfish detection task, with an average detection accuracy of 74.96%. [ABSTRACT FROM AUTHOR]

This paper addresses the diagnosis system of cancer disease using a single setting framework. Most of the radiologists and image specialists are identifying the disease in naked eye. When many conventional systems are used to assess or see a patient's disorder condition, it rarely detects the disease all at once in certain situations. Patients are facing difficulties, when the condition of disease is increasing. Thus, this paper focusses the condition of patient seeing the disease image and developed a single setting framework using a convolutional neural network (CNN) architecture with the help of deep learning approaches. The framework contains several deep learning strategies which are used to determine the patient's relevant illness through affected image, such as mass detection using You-Only-Look-Once (YOLO) approach and the crucial aspect of segmentation by full resolution convolutional networks (FrCN). In last the CNN model is considered for classification. This paper is considered to implement our model using breast cancer disease. The different classifiers and cross-validation tests are taken for evaluating validation matrix items. Comparisons of the existing model with the proposed model are made for improving the diagnosis system. For example, the method Inception V3 for accuracy and AUC are 86.77 and 85.89 on MIAS database whereas proposed model got 99.54 and 98.85 on same evaluation items. Our findings show that the proposed diagnostic model outperforms on conventional detection, segmentation, and classification methods. Thus, our diagnosis process worked much better using deep learning and suggested approaches which will help and facilitate the diagnosis of each contaminated region. In each stage of image processing of the infected region, the suggested diagnostics method could support radiologists. [ABSTRACT FROM AUTHOR]

Gait recognition, as one of the new biometric techniques, mainly judges and identifies a target pedestrian through its walking posture. Gait recognition is effective at long distances, difficult to camouflage and requires no contact or cooperation with the target pedestrian. However, the accuracy of gait recognition is affected by external factors, such as the shooting angle of the video, the clothes and bags worn by the target. In this paper, we solve the above problems based on two aspects. Firstly, a gait recognition method based on MSM Loss is proposed. In this way we are able to extract more discriminative spatio-temporal features; Secondly, we also introduce a new input method, which makes each input sequence more closely related, thus improving the gait recognition rate. Finally, the proposed method is verified on the CASIA-B and OU-MVLP dataset. In CASIA-B, the average recognition rate is obtained under the walking conditions of normal, with bags and with clothes. With rank-1 accuracy under LT, the method proposed in this paper can reach 96.4% under NM, 89.1% under BG and 71.2% under CL. And under the normal walking conditions, our method performs better compared with the best existing gait recognition methods. And in OU-MVLP, we get 87.5% accuracy. [ABSTRACT FROM AUTHOR]

Classification of surgical actions from a real anatomy video sequence is a challenging task due to limited visibility, poorer lighting conditions, lower contrast, and obscured frames of a video sequence. Several deep model-based spatiotemporal and spatial feature learning methods have been presented to classify surgical actions. However, some of the methods have been evaluated on synthetic data due to unavailability of sufficient labeled data. Conversely, some methods have been evaluated on a real anatomy dataset, but produce lower accuracy. Therefore, in this paper, first we analyze the effects of both feature learning methods on surgical action classificaiton from a real anatomy dataset. Thereafter, we propose new methods to enhance surgical action classification. Specific contributions in this paper are as follows. First, two novel deep model-based spatiotemporal feature learning methods are proposed to classify surgical actions. Second, a hypothesis is proposed, stating that the elimination of spatiotemporal features does not affect the performance of the method. Third, to test the proposed hypothesis, a spatial feature learning method comprised of a unique custom Convolutional Neural Network (CNN) is also proposed. Fourth, performance analysis of the proposed spatiotemporal and spatial feature learning methods is presented using a real anatomy Surgical Actions 160 dataset. The experimental results demonstrate that the MobileNetV2-based spatial feature learning method achieves the highest accuracy of 97% in classifying surgical actions. It outperforms the other pre-trained CNN models significantly. Additionally, MobileNetV2 exhibits lower training time. [ABSTRACT FROM AUTHOR]

Lung cancer arises from the abnormal and uncontrolled reproduction of parenchymal cells. Among all cancer cases, lung cancer is one of the prevailing types. Prevalence and death rates of the cancer increases day by day. From this point of view, early diagnosis and treatment of this cancer increases survival times and rates. The main idea in the development of the method presented in this publication is to increase the rate of early diagnosis. Computerized Tomography (CT) is the major screening method when encountered with suspicious symptoms. The cancer can be determined with CT and besides subtyping can be done. Diagnosing the disease with the human eye can sometimes lead to the emergence of deficiencies. This is one of the problems faced today. In this direction, the study in this paper presents a hybrid method to predict and diagnose the lung cancer from CT images to minimize potential human errors. Using the method, feature maps of the CT images of the dataset are obtained using the previously trained DarkNet-53 and DenseNet-201 deep model architectures. DarkNet-53 and DenseNet-201 architectures were chosen because they gave the best feature extraction results among 7 different architectures. The purpose of using the two architectures is to combine two high-performance models to create a hybrid classification method with high accuracy. Feature concatenation is applied to increase the diagnosis accuracy. To optimize the performance and computation cost of the proposed method, the Neighborhood Component Analysis (NCA) optimization method is used in determining and analysis of the features with more information. Therefore, features with less contribution in the accuracy are eliminated. Next, new feature maps are achieved by grading all features upon their weights and applying an elimination using a threshold value. The new feature maps are classified using Classical Machine Learning (CML) classifiers. Classification accuracies on DarkNet-53 architecture were calculated as 69.11% with SoftMax and 96.25% with Ensemble Classifiers and Nearest Neighbor Classifiers respectively. Similarly, accuracies on DenseNet-201 architecture were calculated as 68.29% with SoftMax and 97.39% with Ensemble Classifiers and Nearest Neighbor Classifiers respectively. With the proposed hybrid model, the Ensemble Classifiers reached the accuracy of 98.69% and the highest accuracy is achieved by using k-Nearest Neighbor Classifier (kNN) with the value of 98.86%. The results are supported with detailed illustrations. [ABSTRACT FROM AUTHOR]

Nowadays capturing video through mobile phones, digital cameras and uploading it in social media is a trend. These videos do not have semantic tags. Searching these kinds of videos is difficult to web users. Content Based Video Retrieval (CBVR) helps to identify the most relevant videos for a given video query. The objective of the paper is retrieve most relevant videos for a given query video in reduced time. To meet the objective, this paper proposes an efficient video retrieval system using salient object detection and keyframe extraction methods to reduce the high dimensionality of video data. The spatio-temporal features are extracted using two-stream Convolutional Neural Network (CNN) and stored in a feature dataset. The salient objects are used to search the exact subject that is given as query. The relevant videos are identified through similarity matching of feature dataset that are created using the input dataset with the feature of query video. To reduce the complexity of similarity matching, the proposed method replaces feature dataset with classification score dataset. Experiments are conducted on TRECVID and CC_Web_Video datasets and evaluated using precision, recall, specificity, accuracy and f-score. The proposed method is compared with recent methods and proved its efficiency with approximately 99.68% precision rate on TRECVID dataset and 88.9% precision rate on CC_Web_Video dataset. The proposed outperforms most recent methods by 0.001 increase in mean Average Precision (mAP) on CC_Web_Video dataset and 4% increase in precision rate on TRECVID dataset. The computation time is reduced by 100 min on TRECVID and 200 min on CC_Web_Video datasets. [ABSTRACT FROM AUTHOR]

With the development of the era of internet, an increasing number of images flow into people's daily life, and it's really a challenge to quickly search interesting images in such a huge image database. The most advanced method is using a deep neural network to get hash code of images to achieve fast image retrieval at present. However, people usually use the single pooling method to screen the image pixels when designing the neural network, and some effective information of the image will be gradually lost due to the single pooling method with the number of network deepening. Aiming at this problem, this paper proposes a convolutional neural network combining multiple pooling methods to preserve the effective information of the image as much as possible. To verify the effectiveness of the proposed method, many experiments are carried out on the CIFRA-10, NUS-WIDE and MNIST datasets. The experimental results show that the proposed method is better than most existing hash-based image retrieval methods. [ABSTRACT FROM AUTHOR]

Nowadays more and more elderly people are suffering from Alzheimer's disease (AD). Finely recognizing mild cognitive impairment (MCI) in early stage of the symptom is vital for AD therapy. However, brain image samples are relatively scarce, meanwhile have multiple modalities, which makes finely classifying brain images by computers extremely difficult. This paper proposes a fine-grained brain image classification approach for diagnosing Alzheimer's disease, with re-transfer learning and multi-modal learning. First of all, an end-to-end deep neural network classifier CNN4AD is designed to finely classify diffusion tensor image (DTI) into four categories. And according to the characteristics of multi-modal brain image dataset, the re-transfer learning method is proposed based on transfer learning and multi-modal learning theories. Experimental results show that the proposed approach obtain higher accuracy with less labeled training samples. This could help doctors diagnose Alzheimer's disease more timely and accurately. [ABSTRACT FROM AUTHOR]

The factors that concern the current AI medical models are the lack of generalizing capability when they are subjected to clinical data and also the scarcity of labeled medical data from which they can learn. This paper studies the role of transfer learning by fine-tuning the network when different fractions of medical data are available at the downstream task of diabetic retinopathy (DR) severity detection. The experimental results signify that supervised pre-training on ImageNet, followed by fine-tuning on labeled domain-specific fundus images significantly improves the efficacy of the medical image classifier when trained on full training data thereby suggesting transfer learning works. But what is less known is how the fine-tuning performance is affected when subjected to different fractions of data and if the learning is label efficient. Hence, we investigate the performance of the model under different fractions of labeled data (20 %, 40 %, 60 %, and 80 % of the entire data) on DR classification task, the results suggest that supervised fine-tuning underperforms when model is trained under low data regime. The proposed model achieves test accuracy of 0.8010, AUC of 0.86, F1 score of 0.6477, and cohen kappa score of 0.7007 when trained on full training data but underperforms when subjected to low data regime. Thereby suggesting the limits of supervised learning when the model is trained using limited annotated data. Hence our work opens door to further research in achieving good performance at low data regimes. [ABSTRACT FROM AUTHOR]

Recent advancements in activity recognition from sports videos have attracted wide scientific interest of the Computer Vision community. However, the activity recognition problem from cricket video sequences is largely under-represented in the literature. This paper aims to devise a convolutional neural network (CNN) based model for sports activity recognition. The model is trained on the pre-trained VGG16, VGG19, ResNet50, and Inception V3 Models and tested on the clustered cricket videos frames extracted from the data set especially prepared for this research. The clustering of the frames is done by using K-Mean clustering algorithm. K-Fold cross validation is done which gave an accuracy of 99% on clustered data and 91% on un-clustered data. The accuracy and time complexity of the proposed method is better as compared to the state of the art methods used for activity recognition from videos. [ABSTRACT FROM AUTHOR]

The mango crop production and quality are affected by several factors. Plant disease is one of the prime factors that impact crop yield. The disease impacts all parts of plants such as leaves, roots, and stems. Farmers waste most of their energy, time, and effort dealing with it manually and lead to heavy loss in the yield of the crops. Mango production is affected by several types of diseases. Prior recognition of disease would ease and reduce the diagnostic process that enhances the production of quality crops. A huge investigation has been carried out in the field of identifying and classifying leaf disease. Hence, an appropriate artificial intelligence (AI) based solution is needed to support farmers. The objective of the work is to provide an effective and efficient AI-based solution to detect and classify leaf disease earlier. Earlier disease detection in crops is the most prominent way to prevent loss of money and time. The plant leaf image is mainly used as a source to detect disease in the plant. This paper employs an ensemble stacked deep learning model to resolve the problem of automatic identification of mango-leaf diseases. In the proposed approach, initially, the images are segmented for the region of interest and input to a stack of various deep neural networks. The outcome of the deep neural network is aggregated with a machine learning model to identify leaf disease. This model is used to identify various mango leaf diseases such as Powdery mildew, Anthracnose, etc. In the experiment, the deep learning models are stacked and aggregated with machine learning to identify mango leaf diseases. The proposed model is compared with state-of-the-art models and outperforms with 98.57% accuracy. [ABSTRACT FROM AUTHOR]

Wool fabric is an important material in the weaving sector despite its disadvantages, such as the difficulty of classifying and storing pattern information and identifying when this information needs to be reused. With the deep learning method, these problems can be ingeniously solved. Therefore, the use of convolutional neural networks (CNNs) in artificial intelligence to extract texture features is a helpful and vital approach to resolving fabric pattern classification problems in fabric traceability and management. In this paper, we combined the unique advantages of Inception and ResNet to improve the feature extractor. After training the new CNN with fabric images, the texture features are well extracted and the fabric categories are properly classified through optimized classifiers. Different data are employed to confirm the generalization and robustness of the model, including training with different image sizes and small training sets. The proposed network model outperforms the classic deep learning classification algorithms in both accuracy and speed. [ABSTRACT FROM AUTHOR]

The completely paralyzed and quadriplegic patients cannot communicate with others. However, the imagined thoughts of these patients can be used to drive assistive devices by brain-computer interfacing (BCI), the success of which relies on better classification accuracies. In this paper, we have performed an experiment for the classification of imagined words, which can provide an alternative neural path of speech communication for deprived people. A 32-channel industry-standard physiological signal system is used to measure imagined electroencephalogram (EEG) signals of five words (sos, stop, medicine, washroom, comehere) from 13 subjects. We have used the Hilbert transform to calculate time and joint time–frequency features from the imagined EEG signals. The above features are extracted individually in electrodes corresponding to nine brain regions. Each region of the brain is further analyzed in seven EEG frequency bands. The imagined speech features from each of the 63 combinations of brain region and frequency band are classified by the proposed deep architectures like long short term memory (LSTM), gated recurrent unit, and convolutional neural network (CNN). Some combinations are also classified by six traditional machine learning classifiers for performance comparison. In a five-class classification framework, we achieved the average and maximum accuracy of 71.75% and 94.29%. CNN gave high accuracy, but LSTM gave less network prediction time. Our results show that the alpha band can classify imagined speech better than other frequency bands. We have implemented subject-independent BCI, and the results are better than the state-of-the-art methods present in the literature. [ABSTRACT FROM AUTHOR]

With the rapid and unprecedented growth of textual data in recent years, there is a remarkable need for automatic text summarization models to retrieve useful information from these large numbers of textual documents without human intervention within a reasonable time. Text summarization is commonly performed based on extractive and abstractive paradigms. Although different machine learning and deep learning based methods have been proposed for the task of text summarization during the last decades, they are still in their early steps of development and their potential has yet to be fully explored. Accordingly, a new summarization model is proposed in this paper which takes advantage of both extractive and abstractive text summarization models as a single unified model based on the strategy gradient of reinforcement learning. The proposed model also employs the combination of convolutional neural network and gated recurrent unit in both extraction and abstraction modules besides attention mechanism. Moreover, language models, namely Word2Vec and BERT, are used as the backbone of the proposed model to better express sentence semantics as a word vector. We conducted our experiments on widely-studied text summarization datasets (CNN\Daily Mail and DUC-2004) and according to the empirical results, not only the proposed model achieved higher accuracy compared to both extractive and abstractive summarization models in terms of ROUGE metric but also its generated summaries presented higher saliency and readability based on human evaluation. [ABSTRACT FROM AUTHOR]

Aiming at the essential feature of the time-continuity of birdsong in nature, this paper proposed a birdsong classification model composed of two feature channels, which combines the features of time domain and time frequency domain. In order to make better use of the features, we used the improved average threshold method to denoise the original time-domain waveform features to reduce the influence of noise features. The most suitable feature extractor and the best fusion method of these two features are discussed. In this paper, the 3D convolutional neural network (3DCNN) and 2D convolutional neural network (2DCNN) were respectively applied as feature extractors of log_mel spectrum and waveform images. Then the advanced feature, which was extracted from these two feature channels, was fused in the middle stage, and the output enhanced feature was used as the input of double gated recurrent unit (d-GRU) network. In the work, birdsongs of four species from Xeno-Canto were selected for testing. The results showed that these three methods had improved the classification effect: feature fusion method in time domain and time-frequency domain, weighted average threshold noise reduction method and the method of extracting birdsong features via different types of feature extractors. The method of this paper had achieved mean average precision (MAP) of 95.9% in the classification comparison experiments, which was an inspiring outcome. [ABSTRACT FROM AUTHOR]

Cardiovascular is nowadays a common and threatening disease for humans. Computer-aided diagnostic (CAD) can diagnose cardiovascular by finding anomalies in an electrocardiogram (ECG). However, this conventional diagnostic approach is inefficient and needs extensive analysis and medical knowledge to diagnose accurately. Deep learning can help in the timely detection of anomalies. ECG being a multi-tone signal, CNN-based temporal features are not sufficient for classification. A classification can be improved by integrating multi-spectral information with temporal features. Furthermore, the one-dimension (1D) ECG signal needs complex preprocessing, including signal cleaning and RR peak detection, whereas deep CNN is more efficient for 2D signals. This paper proposes an automated CAD system for ECG classification using a 2D deep-convolution network (CNN). The proposed CNN uses multi-spectral information by integrating wavelet-based spectral features with CNN's temporal features. The 1D ECG is reshaped to a 2D image, and a wavelet-encoded 2D CNN is proposed to classify these 2D images into four classes. The proposed model is evaluated using MIT-BIH datasets containing two-channel ambulatory ECG signals. A comparative study of both 1D and 2D approaches using deep CNN is presented in the paper initially. Later, the proposed network is validated by evaluating various quantitative parameters using the MIT-BIH dataset, and a comparative analysis is shown. The proposed CNN outperformed other models achieving the highest accuracy of 99.52% and 95.64% F1 score. The study reveals that the proposed model avoids the complex preprocessing of the ECG signals of cleaning, RR point detection, and waveform cropping. Further integration of wavelet-based multi-spectral features in CNN has improved the classification accuracy. [ABSTRACT FROM AUTHOR]

In recent decades, many people have died because of cancer. Like many diseases, early detection of the disease can greatly help the healing process of the sick people. One of the most common types of cancer is the skin cancer. Skin lesions can lead to a kind of malignant cancer. Hence, it is important to diagnose the type of the skin lesion. Mostly, the skin lesion diagnosis needs to be done by an expert dermatologist. But a dermatologist's examination can be time consuming and inaccurate. Therefore, using modern computer aided diagnosis methods can help to increase the speed and accuracy of the diagnosis of this disease. This paper proposes a new approach for skin lesion recognition using a new Convolutional Neural Network (CNN) with 69 layers. The proposed method has been tested on three benchmark datasets of skin lesions including PH2, ISIC 2016, and ISIC 2017. The experimental results illustrate that the proposed method performs better compared to the state-of-the-art in the field. Our method reached to 97.2%, 96.3%, and 99.4% of accuracy on PH2, ISIC2016, and ISIC2017 datasets, respectively. [ABSTRACT FROM AUTHOR]

Blind image deblurring is the process of recovering the original image from a degraded image under unknown point spread function, and it is the solution to an ill-posed inverse problem. In this paper, the blurry image is firstly divided into skeleton image and blur kernel, aiming to achieve accurate blur kernel estimation. Then the advantages of model-based optimization method and discriminative learning method are integrated through variable splitting technique. Finally, a trained convolutional neural network (CNN) is used as a module to be inserted into a model-based optimization method to solve the problem of blind image deblurring more effectively. By comparing visual and quantitative experimental data, the network proposed in this paper can provide powerful prior information for blind image deblurring and the restoration effects can approximate or exceed those of some representative algorithms. [ABSTRACT FROM AUTHOR]

To solve the low detection accuracy of SSD for the small size object, this paper proposed an improved algorithm of SSD object detection based on the feature pyramid (FP-SSD). In the deep convolutional neural network, the high-level features contain well semantic information but are not sensitive to the translations. The low-level features have high resolutions but could not represent the features well. The feature pyramid structure contains multi-scale features. To combine the high and low-level features of the pyramid, the algorithm of this paper applied the deconvolution network to the high-level features of the feature pyramid to get the semantic information, dilated convolution network to learn the position information of the low-level features and used convolution for the middle level features to reduce the feature channels, then used convolution to fuse the features. After using the algorithm, a multi-scale detection structure is constructed. FP-SSD achieves a mean accuracy of 79% on PASCAL VOC2007, and 47% on MSCOCO, which has a great improve compared with SSD. We compared the detection accuracy and results with all kinds of scales by experiments, compared with SSD, the accuracy of FP-SSD is higher, which has more accurate location and higher recognition confidence. [ABSTRACT FROM AUTHOR]

The mortality of skin pigmented malignant lesions is very high, especially melanoma. Due to the limitation of marking means, the large-scale annotation data of skin lesions are generally more difficult to obtain. When the deep learning model is trained on a small dataset, its generalization performance is limited. Using prior knowledge to expand small sample data is a general model method of learning classification, which is difficult to deal with complex skin problems. On the basis of a small amount of labeled skin lesion image data, this paper uses the improved Relational Network for measurement learning to realize the classification of skin disease. This method uses relative position network (RPN) and relative mapping network (RMN), in which RPN captures and extracts feature information by attention mechanism, and RMN obtains the similarity of image classification by weighted sum of attention mapping distance. The average accuracy of classification is 85% on the public ISIC melanoma dataset, and the results show the effectiveness and applicability of the method. [ABSTRACT FROM AUTHOR]

Gaming video streaming services are growing rapidly due to new services such as passive video streaming of gaming content, e.g. Twitch.tv, as well as cloud gaming, e.g. Nvidia GeForce NOW and Google Stadia. In contrast to traditional video content, gaming content has special characteristics such as extremely high and special motion patterns, synthetic content and repetitive content, which poses new opportunities for the design of machine learning-based models to outperform the state-of-the-art video and image quality approaches for this special computer generated content. In this paper, we train a Convolutional Neural Network (CNN) based on an objective quality model, VMAF, as ground truth and fine-tuned it based on subjective image quality ratings. In addition, we propose a new temporal pooling method to predict gaming video quality based on frame-level predictions. Finally, the paper also describes how an appropriate CNN architecture can be chosen and how well the model performs on different contents. Our result shows that among four popular network architectures that we investigated, DenseNet performs best for image quality assessment based on the training dataset. By training the last 57 convolutional layers of DenseNet based on VMAF values, we obtained a high performance model to predict VMAF of distorted frames of video games with a Spearman's Rank correlation (SRCC) of 0.945 and Root Mean Score Error (RMSE) of 7.07 on the image level, while achieving a higher performance on the video level leading to a SRCC of 0.967 and RMSE of 5.47 for the KUGVD dataset. Furthermore, we fine-tuned the model based on subjective quality ratings of images from gaming content which resulted in a SRCC of 0.93 and RMSE of 0.46 using one-hold-out cross validation. Finally, on the video level, using the proposed pooling method, the model achieves a very good performance indicated by a SRCC of 0.968 and RMSE of 0.30 for the used gaming video dataset. [ABSTRACT FROM AUTHOR]

In this paper, a unique facial expression recognition system has been proposed. The objective of this paper is to identify the type of human facial expression and to improve the performance by incorporating different variant patterns present in facial images. In literature, the use of different patterns of the human face has not yet been employed in its full swing. In our proposed method, this gap has been overcome by fusing a data augmentation in the preprocessing step and an optimized weight in the deep CNN model at the feature extraction step. To make this proposed model noise-free, we focus on the feature extraction method rather than designing a complex model to decompose a set of feature vectors into expression-specific feature vectors. Here, to analyze the performance of FER system, we use cultural difference datasets in both laboratory controlled and wild environments. To develop a precise facial expression recognition system, we propose an invariant deep convolutional neural network (DCNN) model that learns from all image variants to improve performance. Our model also induces a new pipeline strategy to correlate a preprocessing and feature extraction step in an optimized way. Experiments on laboratory and wild-controlled datasets show that our FER system attains a more significant performance than other state-of-the-art models due to its ability to utilize all the different variant patterns present in the facial image. [ABSTRACT FROM AUTHOR]

U-NET is a fully convolutional network (FCN) architecture designed to research the segmentation of biomedical images. The depth of the U-NET is one of the major constraints of this model while computing the performances. The larger depth of the U-NET means that its computational complexity is high as well. In certain cases, this large depth, as in the original model, is not justified for biomedical imaging modalities. In this paper, we have done an efficient analysis of U-NET architecture's depth variation, i.e., after removing different layers. For the analysis, the datasets BraTS-2017 and BraTS-2019, which consist of High-Grade Glioma (HGG) and Low-Grade Glioma (LGG) MR Scans, have been used for tumor segmentation. We have achieved a dice coefficient of at least 0.8866 and as high as 0.8887 on the discovery cohort, and at least 0.8895 and as high as 0.8911 cross-validation replication cohort. The results show that there are the least significant changes occurring in the performance parameters while moving from the higher to the lower depth of the model. Hence, in this paper, we presented that the large depth of U-NET, which costs more in terms of computational complexity, is not always required. Moreover, the U-NET models with depth reduction, which decreases the computational complexity, can achieve nearly the same results as in the case of the original U-NET. [ABSTRACT FROM AUTHOR]

Recently, anchor-based methods have achieved great progress in face detection. They adopt standard anchor matching strategy to sample positive anchors according to predefined IoU threshold. However, the max IoUs of extreme aspect ratio faces are still lower than fixed positive threshold, leading to the sampling failure from these faces. To construct a more robust detection model, more positive anchors from extreme aspect ratio faces need to be sampled and participate in the training phase. The goal of the present research is to improve the detection performance by reasonably extending sampling range of face aspect ratio. In this paper, we firstly explore the factors that affect the max IoU of each face in theory. Then, anchor matching simulation is performed to evaluate the sampling range of face aspect ratio. Finally, we propose a Wide Aspect Ratio Matching (WARM) strategy to collect more representative positive anchors from ground-truth faces across a wide range of aspect ratios. Besides, we present a novel feature enhancement module, named Receptive Field Diversity (RFD) module, to provide diverse receptive field corresponding to different aspect ratios. Extensive experiments have been conducted on popular benchmarks to show the effectiveness of our method, which can help detectors better capture extreme aspect ratio faces. Our method achieves promising APs on WIDER FACE validation dataset (easy: 0.965, medium: 0.955, hard: 0.904) and impressive generalization capability on FDDB dataset. [ABSTRACT FROM AUTHOR]

This paper proposes an efficient deep convolutional neural network with features fusion for recognizing radar signal, which mainly includes data pre-processing, features extraction, multi-features fusion, and classification. Radar signals are first transformed into time-frequency images by using choi-williams distribution and smooth pseudo-wigner-ville distribution, and the image pre-processing methods are used to resize and normalize the time-frequency images. Then, two constructed deep convolutional neural network models are aimed to extract more effective features. Furthermore, a multi-features fusion model is proposed to integrate features extracted from two deep convolutional neural network models, which makes full use of the relationship among different features and further improves the recognition performance. Experimental results shown that the average recognition accuracy of the proposed method is up to 84.38% when the signal to noise ratio is at −12 dB, and even reach to 94.31% at −10 dB, which achieved the superior recognition performance than others, especially at the lower signal to noise ratio. Moreover, the recognition performance of various radar signals can be largely improved, especially for 2FSK, 4FSK and SFM. This work provides a sound experimental foundation for further improving radar signal recognition in modern electronic warfare systems. [ABSTRACT FROM AUTHOR]

In recent years, deep learning has evolved as a rapidly growing and stimulating field of machine learning and has redefined state-of-the-art performances in a variety of applications. There are multiple deep learning models that have distinct architectures and capabilities. Up to the present, a large number of novel variants of these baseline deep learning models is proposed to address the shortcomings of the existing baseline models. This paper provides a comprehensive review of one hundred seven novel variants of six baseline deep learning models viz. Convolutional Neural Network, Recurrent Neural Network, Long Short Term Memory, Generative Adversarial Network, Autoencoder and Transformer Neural Network. The current review thoroughly examines the novel variants of each of the six baseline models to identify the advancements adopted by them to address one or more limitations of the respective baseline model. It is achieved by critically reviewing the novel variants based on their improved approach. It further provides the merits and demerits of incorporating the advancements in novel variants compared to the baseline deep learning model. Additionally, it reports the domain, datasets and performance measures exploited by the novel variants to make an overall judgment in terms of the improvements. This is because the performance of the deep learning models are subject to the application domain, type of datasets and may also vary on different performance measures. The critical findings of the review would facilitate the researchers and practitioners with the most recent progressions and advancements in the baseline deep learning models and guide them in selecting an appropriate novel variant of the baseline to solve deep learning based tasks in a similar setting. [ABSTRACT FROM AUTHOR]

Brain cancer is quite possibly the most common cause of death in recent years. Appropriate diagnosis of the cancer type empowers the specialists to make the right choice of treatment, decision, and to save the patient's life. It goes without saying the importance of a computer-aided diagnosis system with image processing that can classify the tumor types correctly. In this paper, an enhanced approach has been proposed that can classify brain tumor types from magnetic resonance images (MRI) using deep learning and an ensemble of machine learning (ML) algorithms. The system named BCM-VEMT can classify among four different classes that consist of three categories of brain cancers (Glioma, Meningioma, and Pituitary) and a non-cancerous class, which means normal type. A convolutional neural network was developed to extract deep features from the MRI images. These extracted deep features are fed into ML classifiers to classify among these cancer types. Finally, a weighted average ensemble of classifiers is used to achieve better performance by combining the results of each ML classifier. The dataset of the system has a total of 3787 MRI images of four classes. BCM-VEMT has achieved better performance with 97.90% accuracy for the Glioma class, 98.94% accuracy for Meningioma, 98.92% accuracy for Pituitary and 98.00% accuracy for the Normal class. BCM-VEMT can have great significance for medical sectors in classifying brain cancer types. [ABSTRACT FROM AUTHOR]

The feature extraction ability of some existing super-resolution networks is relatively weak. And these networks do not further process the extracted features. These problems make the networks often show limited performance, resulting in blurred details and unclear edges of the reconstructed images. Therefore, further research is needed to resolve these problems. In this paper, we propose a novel dual-branch feature learning super resolution network (DBSR). The core of DBSR is the dual-branch feature learning (DB) block. In order to enhance the ability of feature extraction, the block adopts a multi-level and dual-branch structure. At the same time, some components for further processing features are introduced in each branch to maximize the learning ability of the block. The reconstructed images of DBSR are clearer than other networks in line and contour, and better results are obtained in peak signal-to-noise ratio (PSNR) and structural similarity (SSIM). For example, when the scaling factor is 2, the PSNR/SSIM on each test dataset is 38.25dB/0.9614, 34.11dB/0.9218, 32.35dB/0.9018, 32.94dB/0.9354 and 39.46dB/0.9783 respectively. The experimental results demonstrate that DBSR achieves better accuracy and visually pleasing than the current excellent methods. [ABSTRACT FROM AUTHOR]

In the past few years, sentiment analysis (SA) of online content has gained more attention in the research area due to the enormous increase of online content from various sources like websites, social blogs, etc. Many organizations use SA techniques to determine the opinion of users and to ensure their satisfaction. Numerous techniques are suggested by many researchers to identify the sentiments of online content. Among them, hybrid of deep learning and lexicon-based SA techniques are gaining more attention due to their outstanding performance than other approaches. Though the lexicon-based SA approaches integrated with deep learning SA approaches possess more advantages they suffer from lack of accuracy and scalability issues due to the high-dimensional features. To eliminate this issue, a hybrid SA approach is proposed in this paper with a bio-inspired feature selection technique. The Valence Aware Dictionary for Sentiment Reasoning (VADER) approach is integrated with the hybrid deep learning approach of attention-based bidirectional long short-term memory and variable pooling convolutional neural network (VPCNN-ABiLSTM) for SA. The optimal features are selected to minimize the scalability issue by integrating the chimp optimization algorithm with the opposition-based learning technique. The performance of the proposed approach is evaluated for four types of benchmark datasets in terms of precision, accuracy, recall, and F1 score. The proposed approach with OBL-CHOA based feature selection technique achieved higher accuracy of 97.1% with the reduction of 13.6% features. The accuracy of the proposed approach with the feature selection technique is 6.9% higher than the existing BiLSTM-CNN based SA approach. [ABSTRACT FROM AUTHOR]

A forged image is a major source of counterfeit news and is mostly used in a spiteful manner such as exciting targeted mob, identity theft, defaming individual, or misleading law bodies. Therefore, a technique is required to detect the tampered regions in a forged image. Deep learning is surpassing technology for prediction or classification tasks in images. Challenges in this technology are a variety of datasets to train the model and specific architecture for a specific application. In this paper, a deep learning model is extended for the localization of tampered regions in a forged image. This is an extension of the well-known U-Net segmentation model. In the proposed model, batch normalization layers and identity-blocks are placed at suitable places of the U-Net model to overcome the challenges such as overfitting and loss of information during max-pooling. To overcome the challenge of the dataset five different publicly available datasets are taken to train, validate and test the model. The trained model is also tested on four created forged images (not belong to the dataset) whose acquisition sources may different i.e. medical image, identity document, natural image, and scanned report. The result of the proposed model is compared with state-of-the-art techniques which show that the method works better than others. [ABSTRACT FROM AUTHOR]

The grading of mangoes is an essential aspect of providing quality fruits to consumers and control the needs of the fruit processing industry. Manual visual inspection leads to inconsistencies, and it is human labour intensive. This paper is focused on improving the accuracy of the automatic mango grading system by doing multi-level grading using Deep Learning, Computer Vision and Image processing techniques. The proposed system is based on the mango maturity ripening stage, shape, texture features, colour and defects to identify the mango variety and classify based on quality. The maturity ripening stage of the mango is extracted using the Convolutional Neural Network (CNN). Computer Vision and Image processing techniques are used to extract shape, texture features and defects. The extracted features are input to the Random Forest classifier to identify the mango variety and grade the mango quality into three classes Notfit, Average and Good. The system has been validated on the dataset created for this study across three different varieties, Banganapalli, Neelam and Rumani, the most popular in Tamil Nadu. The proposed system using features extracted from CNN enhanced the system's efficiency with an accuracy of 93.23% for variety recognition and 95.11% for quality grading. Hence the proposed system is fully automated, commercially viable and has improved accuracy in variety recognition and quality grading of mangoes across different varieties. [ABSTRACT FROM AUTHOR]

With the dramatic growth of various social media platforms, sentiment analysis (SA) of and emotion detection (ED) in various social network posts, blogs, and conversations are very useful and effective for mining the true opinions on different issues, entities, or aspects. During the last decade, many statistical and probabilistic models based on lexical and machine learning approaches have been employed for these tasks. Majority of the relevant literature has emphasized on improving the contemporary SA determination and emotion extraction techniques. With the recent advancements in deep neural networks, various deep learning models have been heavily used to enhance the accuracy of SA. Convolutional neural networks (CNN), a deep neural network model formerly adopted for visual data processing only, has recently gained acceptance for textual inputs as well. As the inputs for SA may be textual, visual, or any combination of these, CNN seems to be a powerful tool. Capturing spatial and contextual information in an incremental fashion respectively from visual and textual inputs proves CNN as an effective model for SA. In this paper, we present an extensive survey that covers the applicability, challenges, and issues for textual, visual, and multimodal SA using CNNs. A detailed discussion and analysis for SA using a CNN model is summarized. For both of the unimodal inputs i.e., textual and visual, we present an optimized algorithmic approach for SA determination using CNN. [ABSTRACT FROM AUTHOR]

Various Human Activities are classified through time-series data generated by the sensors of wearable devices. Many real-time scenarios such as Healthcare Surveillance, Smart Cities and Intelligent surveillance etc. are based upon Human Activity Recognition. Despite the popularity of local features-based approaches and machine learning approaches, it fails to capture adequate temporal information. In this paper, the deep convolutional neural model has been proposed by combining external features, i.e. orientation invariant (|| v ||) and consecutive point trajectory information (|| Δ v ||) with tri-axis data of the accelerometer. The proposed external features based approach experimented on three different deep learning architecture, namely Long-Short Term Memory (LSTM), Convolutional Neural Networks (CNN) and Convolution Long-Short Term Memory (ConvLSTM). Accuracy of the algorithms radically improve with the additional input feature || v || and || Δ v || along with tri-axis data of accelerometer. The results show that the performance of all three LSTM, CNN and ConvLSTM models is better to compare with the state of art methods on WISDOM dataset and Activity dataset also the performance of ConvLSTM is 98.41% for WISDOM dataset and 98.04 for activity dataset, which is higher than that of CNN and LSTM model used in this paper. [ABSTRACT FROM AUTHOR]

Considering the noise impact in the bearing operating environment and the time-consuming and non-universal design of traditional diagnostic algorithms, this paper proposes a new model for rolling bearing fault diagnosis, which uses convolutional pooling group (CPG) to extract features from data, At the same time, expanding the dual convolutional kernel to obtain a larger receptive field obtained the WCPGCNN (A CPG Convolutional Neural Network with Wide Convolutional Kernel as the First Lay) model based on the CPG network architecture. Firstly, the fault features of the input signal are automatically extracted through four convolutional pooling groups; Next, fault features are further extracted using the fully connected layer, and finally input into the Softmax layer for fault identification. By utilizing algorithms such as Adam, dropout, and batch normalization, the model performs well in terms of accuracy, noise resistance, and timeliness, while also possessing good cross dataset high generalization ability. This article uses the rolling bearing fault standard data from Case Western Reserve University (CWRU) and the American Society for Mechanical Fault Prevention Technology (MFPT), and verifies through multiple controlled experiments that the model established in this article has high accuracy and good generalization characteristics. [ABSTRACT FROM AUTHOR]

Head pose estimation methods can be generally classified into two categories: model-based and appearance-based methods. The model-based approach relies on facial landmarks for three-dimensional reconstruction, aiming to achieve high-precision results. However, this method is heavily dependent on the accuracy of these landmarks. The appearance-based approach utilizes images as input and employs feature extraction and calculations to generate outcomes. While the appearance-based method boasts greater robustness, its accuracy falls short of the former. In this paper, a new and effective hybrid method is proposed. This hybrid approach combines the strengths of both methods. Unlike the conventional model-based methods, the proposed method regards the facial landmarks in 2D images as a sequence of neural network inputs and then obtains the head pose estimation results for users by neural network regression. The proposed method solves the fuzzy rotation labeling problem by using a rotation matrix representation, introducing a 6D rotation matrix representation as an intermediate state of the rotation matrix to achieve effective direct regression. Introducing face processing enhances the robustness of the model in cross-dataset scenarios. The proposed method achieves remarkable results based on imprecise face recognition and a simplistic model. The proposed method can be divided into three parts. First, the proposed method applies face processing on the input image; second, the method detects facial landmarks; and third, it converts these facial landmarks into sequences and obtains the 6D rotation representation of the head pose by regression. Extensive experiments on the publicly available BIWI, PRIMA, and DrivFace datasets show that this method is functional and performs better than other state-of-the-art methods. Compared to other methods, this approach demonstrates an average performance improvement of at least 10% across the dataset. [ABSTRACT FROM AUTHOR]

Online handwritten Chinese text recognition (OHCTR) has been a challenging problem due to the large character set, diverse writing styles and variable text line length. The existing convolutional recurrent network (CNN) architectures have achieved greatly success in OHCTR, but they need to convert chronological sequence coordinates into image-like representations or vectors, which will lead to information loss and increase time consumption. To avoid the conversion process, we propose a novel end-to-end attention convolutional recurrent network (EACRN) for OHCTR in this paper. Specifically, The EACRN directly extract local contextual features from raw chronological sequence coordinates using end-to-end CNN. After that, bidirectional long short-term memory (BiLSTM) is employed to capture long-term dependencies of local contextual features. Then multi-head attention is utilized to weight local contextual features. Finally, focal connectionist temporal classification (CTC) objective function is introduced into OHCTR by us to increase attention to low-frequency characters in text and makes predictions. Experiments on two publicly datasets, standard benchmarks dataset CASIA-OLHWDB2.0-2.2 and in-air handwritten Chinese text dataset IAHCT-UCAS2018, demonstrate that our method obtains higher recognition accuracy with faster computation speed and more compact model compared with previous CNN architectures. [ABSTRACT FROM AUTHOR]

Training complex deep neural networks can result in overfitting when the networks are trained from random weight initialization on small datasets. Augmentation helps to reduce the negative effects of overfitting. The findings in computer vision and audio recognition research reveals that the performance of machine learning classifiers is significantly improved when augmentation is used. In the context of ecology, researchers conduct field surveys whereby microphones are placed in some location and audio data is recorded over a period of time. There is however no guarantee that the particular species of interest in the field survey will vocalize frequently near the microphone. Thus, the amount of data captured for the species of interest might be limited, and it may then be the source of overfitting. The main contribution of this paper is in performing experiments with time and frequency masking, and noise addition augmentation techniques in training a visual convolutional neural networks (CNN) repurposed for pattern recognition in acoustic spectrograms. These techniques increased the audio examples for the pin-tailed whydah and the Cape robin-chat to create a robust audio vocalization classifiers. To evaluate the performance of the augmentation techniques we conducted a comparison between experiments run with and without augmentation. We chose to use CNN as our classifier given that they are state-of-the-art in audio recognition tasks and they have revealed good performance. In the used augmentation techniques; time masking achieved 90.2% as the highest testing accuracy while pink noise is the most successful best classifier. [ABSTRACT FROM AUTHOR]

With the rapid increase in user-generated multimedia content, extensive outreach over social media, and their potential in critical applications such as law enforcement, sourcey identification from re-compressed and noisy multimedia are of great importance. This paper proposes a system for speaker-independent cell-phone identification from recorded audio. This system is capable of dealing with test audio with different speech content and a different speaker compared to the training audio. Each recorded audio has the device fingerprint implicitly embedded in it, which encourages us to design a CNN-based system for learning the device-specific signatures directly from the magnitude of discrete Fourier transform of the audio. This paper also addresses the scenario where the recorded audio is re-compressed due to efficient storage and network transmission requirements, which is a common phenomenon in this age of social media. The scenario of the cell-phone classification from the audio recordings in the presence of additive white Gaussian noise is addressed as well. We show that our proposed system performs as well as the state-of-art systems for the speaker-dependent case with clean audio recordings and exhibits much higher robustness in the speaker-independent case with clean, re-compressed, and noisy audio recordings. [ABSTRACT FROM AUTHOR]

Activity recognition represents the task of classifying data derived from different sensor types into one of predefined activity classes. The most popular and beneficial sensors in the area of action recognition are inertial sensors such as accelerometer and gyroscope. Convolutional neural network (CNN) as one of the best deep learning methods has recently attracted much attention to the problem of activity recognition, where 1D kernels capture local dependency over time in a series of observations measured at inertial sensors (3-axis accelerometers and gyroscopes) while in 2D kernels apart from time dependency, dependency between signals from different axes of same sensor and also over different sensors will be considered. Most convolutional neural networks used for recognition task are built using convolution and pooling layers followed by a few number of fully connected layers but large and deep neural networks have high computational costs. In this paper, we propose a new architecture that consists solely of convolutional layers and find that with removing the pooling layers and instead adding strides to convolution layers, the computational time will decrease notably while the model performance will not change or in some cases will even improve. Also both 1D and 2D convolutional neural networks with and without pooling layer will be investigated and their performance will be compared with each other and also with some other hand-crafted feature based methods. The third point that will be discussed in this paper is the impact of applying fast fourier transform (FFT) to inputs before training learning algorithm. It will be shown that this preprocessing will enhance the model performance. Experiments on benchmark datasets demonstrate the high performance of proposed 2D CNN model with no pooling layers. [ABSTRACT FROM AUTHOR]

Due to the similar appearances among many retail products, it is a big challenge to identify the product with high accuracy and low computational cost in smart retail scenes. In this paper, we proposed a lightweight retail product identification and localization method based on an improved convolutional neural network. First, we use group convolution and deep separable convolution to optimize the structure of the backbone network and reduce the amount of calculation. Second, the multiscale structure was adjusted to optimal scales. We further use the k-means clustering algorithm to re-cluster six anchors with different sizes. Third, we introduced spatial pyramid pooling (SPP) to replace pooling by convolution to effectively improve the robustness against image distortion, such as cropping and scaling. Finally, we use mosaic data enhancement method to improved the robustness of the network. Experiments on the RPC dataset show that, compared with YOLOv5, the number of parameters is reduced by 1/6.4 times, and FLOPs is reduced by 1/9 times. Experiments on the DeepBlue Retail Dataset show that compared with YOLOv5, the number of parameters is reduced by 1/7.8 times, and FLOPs is reduced by 1/9.3 times. Realtime evaluation under the same hardware show that the FPS of the proposed model is 123 in the forward inference test, while the FPS of the YOLOv5 model under the same conditions is 58. [ABSTRACT FROM AUTHOR]

In this digital world, digitized documents can be considered original or a piece of evidence; checking the authenticity of any suspicious image has become an unavoidable concern to preserve the trust in its legitimacy. However, identifying the source of a digital image without any prior embedded information is a very challenging task. This paper proposes a novel one-dimensional convolutional neural network (1D-CNN) model to solve the source scanner identification (SSI) problem blindly. Unlike traditional methods based on handcrafted features, the proposed framework can dynamically learn and extract scanner device-specific features. This work, comprised of the 1D-CNN and a support vector machine (SVM) as a classifier, was trained on nine scanners of different brands and models. The experimental result shows that our model achieves 98.15% accuracy on full images and overall accuracy of 93.13% on segments from test images, outperforming other state-of-art approaches. Our model also proves to be able to distinguish between scanners of the same model. Furthermore, the SVM classifier improved the 1D-CNN accuracy by approximately 3% compared to its original configuration. [ABSTRACT FROM AUTHOR]

Recently, convolutional neural networks have achieved great success in image classification. However, the traditional convolutional neural network lacks the ability to distinguish image features, especially for the low resolution images with less feature information. In the vehicle recognition task, it is inevitable to lose some feature information by convolution during the process of the low-level feature is abstracted into the high-level semantic feature. In this paper, an improved convolutional neural network model with higher robustness is proposed, we call it feature fusion convolutional neural network (FFCNN), which can not only produce more discriminative features, but also can avoid interference caused by environmental factors to some extent. Firstly, the strategy of feature fusion is used to fuse the different low-level features in the convolution neural network. Secondly, in order to prevent overfitting, we combine with the network model of sparse and data augmentation to optimize the structure of the network model. The results of the experiment show that the model proposed in this paper has higher recognition accuracy compared with the traditional vehicle recognition methods and the original convolutional neural network models. [ABSTRACT FROM AUTHOR]