Search

Intelligent fault diagnosis techniques play an important role in improving the abilities of automated monitoring, inference, and decision-making for the repair and maintenance of machinery and processes. In this paper, a modified stacked auto-encoder (MSAE) that uses adaptive Morlet wavelet is proposed to automatically diagnose various fault types and severities of rotating machinery. Firstly, the Morlet wavelet activation function is utilized to construct an MSAE to establish an accurate nonlinear mapping between the raw nonstationary vibration data and different fault states. Then, the nonnegative constraint is applied to enhance the cost function to improve sparsity performance and reconstruction quality. Finally, the fruit fly optimization algorithm (FOA) is used to determine the adjustable parameters of the Morlet wavelet to flexibly match the characteristics of the analyzed data. The proposed method is used to analyze the raw vibration data collected from a sun gear unit and a roller bearing unit. Experimental results show that the proposed method is superior to other state-of-the-art methods.

Fast and accurate counting of wheat ears in field conditions is a key element for determining wheat yield. To obtain the number of wheat ears in a field, we propose a new counting algorithm based on computer vision. This algorithm counts wheat ears in remote images through semantic segmentation regression network (SSRNet). SSRNet is a multistage convolutional neural network that we propose to achieve counting problems through regression. In SSRNet, first, the original image is cropped to increase the amount of data. This method effectively solves the small sample dataset. Next, based on the cropping results, we build a fully convolutional neural network (FCNN) to segment wheat ears in field conditions. FCNN increases the accuracy of wheat ears counting by accurately segmenting wheat ears in a complex background. Then, we build a regression convolutional neural network (RCNN) to count wheat ears based on the segmentation results of FCNN. In RCNN, we propose a new activation function positive rectification linear unit (PrLU) to process the last layer of the fully connected layer, so that RCNN can effectively count the number of wheat ears in the image. Finally, a counting strategy is proposed to count the number of wheat ears in the original image. To verify the counting performance of SSRNet, we compare the counting result of SSRNet with the real value of manual statistics. The results show that the average accuracy (Acc), R², and root mean squared error (RMSE) of the SSRNet count results on the test set in this article are 0.980, 0.996, and 9.437, respectively. It can be seen from the results that our proposed method can accurately count wheat ears in field conditions. At the same time, the counting time (0.11 s) shows that SSRNet can quickly estimate the number of wheat ears in field conditions. We concluded that this study can provide important technical support for the high-throughput field wheat ears counting task in large-scale phenotyping work.

Considering that synthetic aperture radar (SAR) images obtained directly after signal processing are in the form of complex matrices, we propose a complex convolutional network for SAR target recognition. In this article, we give a brief introduction to complex convolutional networks and compare them with the real counterpart. A complex activation function is applied to analyze the influence of phase information in complex neural networks. Inspired by the theory of network visualization, a special kind of transfer learning based on the electromagnetic property from the attributed scattering center model is applied in our networks to modulate the first convolutional layer. The experiment shows a better performance in terms of classification accuracy compared to random weight initialization.

Recently, the Extreme Learning Machine (ELM) algorithm has been applied to various fields due to its rapidity and significant generalization performance. Traditionally, deep learning (DL) and wavelet neural networks (WNN) methods reach a high classification accuracy in machine learning applications. As a result, a new structure based on WNN, deep architecture and ELM is proposed in this paper. The proposed method is based on Extreme Learning Machine Auto-Encoder with DL structure and a composite wavelet activation function used in the hidden nodes. To evaluate the performance of our approach, we used standard benchmark data-sets, namely COIL-20, Pima Indian Diabetes (PID), MNIST and EMNIST. Experimental results show that our method offers satisfactory results and performance compared to other approaches.

Coronavirus (COVID-19) infection was initially acknowledged as a global pandemic in Wuhan in China. World Health Organization (WHO) stated that the COVID-19 is an epidemic that causes a 3.4% death rate. Chest X-Ray (CXR) and Computerized Tomography (CT) screening of infected persons are essential in diagnosis applications. There are numerous ways to identify positive COVID-19 cases. One of the fundamental ways is radiology imaging through CXR, or CT images. The comparison of CT and CXR scans revealed that CT scans are more effective in the diagnosis process due to their high quality. Hence, automated classification techniques are required to facilitate the diagnosis process. Deep Learning (DL) is an effective tool that can be utilized for detection and classification this type of medical images. The deep Convolutional Neural Networks (CNNs) can learn and extract essential features from different medical image datasets. In this paper, a CNN architecture for automated COVID-19 detection from CXR and CT images is offered. Three activation functions as well as three optimizers are tested and compared for this task. The proposed architecture is built from scratch and the COVID-19 image datasets are directly fed to train it. The performance is tested and investigated on the CT and CXR datasets. Three activation functions: Tanh, Sigmoid, and ReLU are compared using a constant learning rate and different batch sizes. Different optimizers are studied with different batch sizes and a constant learning rate. Finally, a comparison between different combinations of activation functions and optimizers is presented, and the optimal configuration is determined. Hence, the main objective is to improve the detection accuracy of COVID-19 from CXR and CT images using DL by employing CNNs to classify medical COVID-19 images in an early stage. The proposed model achieves a classification accuracy of 91.67% on CXR image dataset, and a classification accuracy of 100% on CT dataset with training times of 58 min and 46 min on CXR and CT datasets, respectively. The best results are obtained using the ReLU activation function combined with the SGDM optimizer at a learning rate of 10-5 and a minibatch size of 16. © 2022 Tech Science Press. All rights reserved.

In recent years, plenty of metrics have been proposed to identify networks that are free of gradient explosion and vanishing. However, due to the diversity of network components and complex serial-parallel hybrid connections in modern DNNs, the evaluation of existing metrics usually requires strong assumptions, complex statistical analysis, or has limited application fields, which constraints their spread in the community. In this paper, inspired by the Gradient Norm Equality and dynamical isometry, we first propose a novel metric called Block Dynamical Isometry, which measures the change of gradient norm in individual block. Because our Block Dynamical Isometry is norm-based, its evaluation needs weaker assumptions compared with the original dynamical isometry. To mitigate the challenging derivation, we propose a highly modularized statistical framework based on free probability. Our framework includes several key theorems to handle complex serial-parallel hybrid connections and a library to cover the diversity of network components. Besides, several sufficient prerequisites are provided. Powered by our metric and framework, we analyze extensive initialization, normalization, and network structures. We find that Gradient Norm Equality is a universal philosophy behind them. Then, we improve some existing methods based on our analysis, including an activation function selection strategy for initialization techniques, a new configuration for weight normalization, and a depth-aware way to derive coefficients in SeLU. Moreover, we propose a novel normalization technique named second moment normalization, which is theoretically 30% faster than batch normalization without accuracy loss. Last but not least, our conclusions and methods are evidenced by extensive experiments on multiple models over CIFAR10 and ImageNet., Comment: Under review as a regular paper in TPAMI

Lately, there has been a rapid increase in the use of software-based deep learning neural networks (S-DNN) for the analysis of unstructured data consumption. For implementation of the S-DNN, synapse-device-based hardware DNN (H-DNN) has been proposed as an alternative to typical Von-Neumann structural computing systems. In the H-DNN, various numerical values such as the synaptic weight, activation function, and etc., have to be realized through electrical device or circuit. Among them, the synaptic weight that should have both positive and negative numerical values needs to be implemented in a simpler way. Because the synaptic weight has been expressed by conductance value of the synapse device, it always has a positive value. Therefore, typically, a pair of synapse devices is required to realize the negative weight values, which leads to additional hardware resources such as more devices, higher power consumption, larger area, and increased circuit complexity. Herein, we propose an alternative simpler method to realize the negative weight (named weight shifter) and its hardware implementation. To demonstrate the weight shifter, we investigated its theoretical, numerical, and circuit-related aspects, following which the H-DNN circuit was successfully implemented on a printed circuit board.

The aim of this study is to present the numerical solutions of the higher order singular nonlinear differential equations using an advanced intelligent computational approach by manipulating the Morlet wavelet (MW) neural networks (NNs), global approach as genetic algorithm (GA) and quick local search approach as interior-point method (IPM), i.e., GA-IPM. MWNNs is applied to discretize the higher order singular nonlinear differential equations to express the activation function using the mean square error. The performance of the designed MWNNs using the GA-IPM is observed to solve three different variants based on the higher order singular nonlinear differential model to check the significance, efficacy and consistency of the designed MWNNs using the GA-IPM. Furthermore, statistical performances are provided to check the precision, accuracy and convergence of the present approach.

Accurate identification of table grapes is crucial to the harvesting process of the grape picking robot. This paper proposes an efficient grape detection model, YOLO-Grape, to solve the problem of unrecognition or decreased recognition accuracy caused by the complicated growth environment, shadows of branches and leaves, and overlapping grapes. To improve the network recognition accuracy, a down-sampling fusion structure is integrated into the network, and the Mish activation function is used. Meanwhile, an attention mechanism is added to the network, and the non-maximum suppression (NMS) function is replaced with the soft non-maximum suppression (Soft-NMS) function, thereby reducing the missing of predicted boxes due to overlapping grapes. Besides, the depthwise separable convolution is introduced to improve the detection speed of the network. In addition, transfer learning is used in the training process to improve the detection accuracy and generalization ability of the model. On the test data set of 700 grape images, the experimental results show that YOLO-Grape achieves an F1-score of 90.47%, a mAP of 91.08% and a detection speed of 81 frames per second. Compared with Faster-RCNN(Resnet50), SSD300, YOLOv4, and YOLOv4-tiny, the mAP of the YOLO-Grape model is increased by 1.67%, 2.28%, 0.84%, and 6.69%, respectively. The average recognition speeds of the YOLO-Grape model were 31.15, 3.38 and 6.45 times of Faster-RCNN(Resnet50), SSD300, and YOLOv4 respectively. Through four sets of comparative experiments, it is found that the proposed YOLO-Grape model achieves high recognition accuracy for occluded grapes, meeting the requirements of grape picking robots for real-time detection of multiple varieties of table grapes in complex situations.

In the modern world, it is necessary to monitor a user's mobility in everyday life to provide advanced mobile services. Mostly, the location-based services (LBS) depend on both the present and future location of the user, which shows the increasing demand on predicting the user's future location. The prediction of trajectories among two different locations is very important as it aids to optimize travel paths among the locations. Moreover, existing techniques rely on high-quality data to deliver optimal results. In this research, we proposed a new optimized deep learning architecture for mobility tracing. The prediction model is based on Optimized DCNN (Deep Conventional Neural Network) that is already trained with the traced location of the user and from the user’s trained movement the model can predict its next location. For precise prediction, the weight and activation function of DCNN is optimally selected with the aid of HS-EH (Hybrid SeaLion -Elephant Herding) algorithm, which is a hybridization of Sea lion Optimization (SLnO) and Elephant Herding Optimization (EHO). Eventually, the proposed method performance is executed in MATLAB 2019a and compared with state-of-art methods with certain performance metrics. Especially, the accuracy of the proposed DCNN + HS-EH algorithm at training rate 10 is 82.96%, 48.88%, 79.27%, and 61.67% better than existing methods.

As the main research content of artificial intelligence, the artificial neural network has been widely concerned because of its excellent performance in the fields such as computer vision and natural language processing since it was proposed in the 1940s. The neuron model of the traditional neural network was proposed by McCulloch and Pitts in 1943 (MP neurons), But MP neurons is too simple to representing biological neurons. Based on this, this paper studies the attention mechanism and proposes vectorized neuron and its activation function. Firstly, we propose vectorized neurons, then use the attention mechanism to dynamically generate connection weights between vectorized neurons. Nextly, we construct a new type of neural network with vectorized neurons, which we called neural functional group (NFG). Finally, we tested the proposed neural functional group model on two tasks: image classifcation and few-shot learning. The vectorized neuron can be conditionally activated through its activation function. Besides, the vectorized neuron has the potential of representing complex biological neurons, which is difficult for MP neuron. The experimental results show that it can achieve higher accuracy with fewer parameters than convolutional neural networks (CNN) and capsule networks in image classication task; it also competitive to CNN based feature extractor in few-shot learning task.

Problems such as a vanishing gradient and overfitting will occur when a recurrent neural network (RNN) is exploited to detect abnormal network traffic. In addition, some network traffic is unbalanced, which leads to low detection accuracy. Therefore, an unbalanced abnormal traffic detection method has been proposed. It is composed of the improved bidirectional residual gated recurrent unit (Res-BIGRU) and integrated dynamic extreme learning machine (IDELM). First, the candidate hidden state activation function of the GRU is changed into an unsaturated activation function. The residual connection is used to avoid the vanishing gradient. The purpose of alleviating network degradation is achieved, and the traffic features extracted are better. Second, an IDELM is proposed to solve the unbalanced classification. The minority samples are generated by the IDELM model. The set model in game theory is used to compute the combined weight, which improves the fitting effect. Third, two IDELMs are used to update the final classification results. Fourth, four network datasets and IoT datasets are used to verify the performance. The average accuracy on four network datasets is 91.11% when samples are unbalanced. Furthermore, it can be concluded that the improved Res-BIGRU and IDELM strategy is effective. Better classification results can be achieved when network traffic is unbalanced. In particular, the performance is better in unbalanced NSL-KDD datasets. The index values obtained are the best compared with other methods. It is also suitable for intrusion detection of the Internet of Things, which has good performance. The further advantage lies in that the robustness is better when there are other sample interferences.

Recently, enormous datasets have made power dissipation and area usage lie at the heart of designs for artificial neural networks (ANNs). Considering the significant role of activation functions in neurons and the growth of hardware-based neural networks like memristive neural networks, this work proposes a novel design for a hyperbolic tangent activation function (Tanh) to be used in memristive-based neuromorphic architectures. The purpose of implementing a CMOS-based design for Tanh is to decrease power dissipation and area usage. This design also increases the overall speed of computation in ANNs, while keeping the accuracy in an acceptable range. The proposed design is one of the first analog designs for the hyperbolic tangent and its performance is analyzed by using two well-known datsets, including the Modified National Institute of Standards and Technology (MNIST) and Fashion-MNIST. The direct implementation of the proposed design for Tanh is proposed and investigated via software and hardware modeling.

Over the years, numerous evidences have demonstrated that microbes living in the human body are closely related to human life activities and human diseases. However, traditional biological experiments are time-consuming and expensive, so it has become a research topic in bioinformatics to predict potential microbe-disease associations by adopting computational methods. In this study, a novel calculative method called BPNNHMDA is proposed to identify potential microbe-disease associations. In BPNNHMDA, a novel neural network model is first designed to infer potential microbe-disease associations, its input signal is a matrix of known microbe-disease associations, and its output signal is matrix of potential microbe-disease associations probabilities. And moreover, in the novel neural network model, a new activation function is designed to activate the hidden layer and the output layer based on the hyperbolic tangent function, and its initial connection weights are optimized by adopting Gaussian Interaction Profile kernel (GIP) similarity for microbes, which can improve the training speed of BPNNHMDA efficiently. Finally, in order to verify the performance of our prediction model, different frameworks such as the Leave-One-Out Cross Validation (LOOCV) and k-Fold Cross Validation ( k-Fold CV) are implemented on BPNNHMDA respectively. Simulation results illustrate that BPNNHMDA can achieve reliable AUCs of 0.9242, 0.9127 ± 0.0009 and 0.8955 ± 0.0018 in LOOCV, 5-Fold CV and 2-Fold CV separately, which are superior to previous state-of-the-art methods. Furthermore, case studies of inflammatory bowel disease (IBD), asthma and obesity demonstrate that BPNNHMDA has excellent prediction ability in practical applications as well.

Manusia sebagai makhluk sosial yang selalu ingin berhubungan dengan manusia lainnya memaksa manusia untuk saling berkomunikasi. Di sinilah peran bahasa menjadi amat penting, karena dengan adanya bahasa, maka akan dengan mudah mengerti apa yang ingin disampaikan oleh orang lain. Untuk itu, perlu adanya media yang dapat membantu memahami berbagai bahasa di dunia, salah satunya adalah mesin penerjemah. Salah satu metode yang dapat digunakan untuk membuat mesin penerjemah adalah Neural Machine Translation (NMT). NMT yang sekarang sudah ada masih memiliki berbagai kekurangan dan perlu dilakukan pengembangan lebih jauh. Diantaranya pada masalah overfitting yang membuat modelnya kurang bisa melakukan generalisasi pada data lain yang diujikan. Banyak hal yang mempengaruhi performa dari NMT tersebut, salah satunya adalah ukuran hyperparameter yang digunakan dan arsitektur model yang digunakan. Namun belum ada ukuran pasti yang dapat digunakan untuk menghasilkan model dengan performa yang terbaik. Sehingga penelitian ini bertujuan untuk mengembangkan arsitektur model NMT dan melakukan simulasi pada masing-masing hyperparameter Neural Network dan ukuran pada arsitektur modelnya, antara lain batch size, epoch, optimizer, activation function, dan dropout rate. Hasil yang didapatkan adalah model pengembangan dapat mengatasi masalah overfitting dari model sebelumnya dengan akurasi sebesar 72,24% dan skor BLEU sebesar 45,83% yang dilakukan pada data uji lainnya.

There have been 350,000 tweets generated by the interaction of social networks with different cultures and educational backgrounds in the last ten years. Various sentiments are expressed in the user comments, from support to hatred. The sentiments regarded the United States General Election in 2020. This dataset has 3,000 data gotten from previous research. We augment it becomes 15,000 data to facilitate training and increase the required data. Sentiment detection is carried out using the CNN-BiLSTM architecture. It is chosen because CNN can filter essential words, and BiLSTM can remember memory in two directions. By utilizing both, the training process becomes maximum. However, this method has disadvantages in the activation. The drawback of the existing activation method, i.e., "Zero-hard Rectifier" and "ReLU Dropout" problem to become the cause of training stopped in the ReLU activation, and the exponential function cannot be set become the activation function still rigid towards output value in the SERLU activation. To overcome this problem, we propose a novel activation method to repair activation in CNN-BiLSTM architecture. It is namely the ASERLU activation function. It can adjust positive value output, negative value output, and exponential value by the setter variables. So, it adapts more conveniently to the output value and becomes a flexible activation function because it can be increased and decreased as needed. It is the first research applied in architecture. Compared with ReLU and SERLU, our proposed method gives higher accuracy based on the experiment results.

Neurodegenerative disease defined about death of some brain parts. This disease is the dangerous disease to cure with devastating results. In addition to the elderly, the neurodegenerative disease threatening the pregnant women since it affects the foetus. The existing studies related with neurodegenerative disease is very few. The diagnosis of the neurodegenerative disease occur in foetal during pregnancy is the major challenge in medical field. In this proposed study, the neurodegenerative disease for foetus are classified based on the novel VGG-19 feature extraction and improved deep ResNet classifier. This proposed concept follows feature extraction and transfer learning based classification. For effective feature extraction, the Visual Geometry Group -VGG-19 has been widely used and for better classification of the neurodegenerative disease in foetal MRI brain the proposed improved Deep Residual Network-ResNet classifier is implemented. ResNet allows the alternate shortcut path and reduces the vanishing gradient problem. If the current layer is not required, CNN weight layer is bypassed by ResNet identity mapping. The training set over fitting problem is thus avoided. The VGG-19 network major role is to increase the CNN depth with the help of 3 × 3 filter size. The proposed study are evaluated in terms of various performance metrics and several activation function compared with proposed swish activation function. The results shows that the proposed method resulted in better values compared with existing studies used the foetal brain images.

Recent works based on deep learning and facial priors have performed well in superresolving severely degraded facial images. However, due to the limitation of illumination, pixels of the monitoring probe itself, focusing area, and human motion, the face image is usually blurred or even deformed. To address this problem, we properly propose Face Restoration Generative Adversarial Networks to improve the resolution and restore the details of the blurred face. They include the Head Pose Estimation Network, Postural Transformer Network, and Face Generative Adversarial Networks. In this paper, we employ the following: (i) Swish-B activation function that is used in Face Generative Adversarial Networks to accelerate the convergence speed of the cross-entropy cost function, (ii) a special prejudgment monitor that is added to improve the accuracy of the discriminant, and (iii) the modified Postural Transformer Network that is used with 3D face reconstruction network to correct faces at different expression pose angles. Our method improves the resolution of face image and performs well in image restoration. We demonstrate how our method can produce high-quality faces, and it is superior to the most advanced methods on the reconstruction task of blind faces for in-the-wild images; especially, our 8 × SR SSIM and PSNR are, respectively, 0.078 and 1.16 higher than FSRNet in AFLW.

Object classification, such as handwritten Arabic character recognition, is a computer vision application. Deep learning techniques such as convolutional neural networks (CNNs) are employed in character recognition to overcome the processing complexity with traditional methods. Usually, a CNN is followed by an activation function such as a rectified linear unit (ReLU) or leaky ReLU to filter the extracted features. Most handwritten character recognition endures an imbalanced number of positive and negative vectors. This issue decreases CNN performance when adopting ReLU and leaky ReLU for the next deep layers in the architecture. Hence, this study proposed an optimized leaky ReLU to retain more negative vectors using a CNN architecture with a batch normalization layer to address this weakness. To evaluate the proposed method, four datasets are used: Arabic Handwritten Characters Dataset (AHCD), self-collected, Modified National Institute of Standards and Technology (MNIST), and AlexU Isolated Alphabet (AIA9K). The proposed method shows significant performance in terms of accuracy, precision, and recall measures compared to the state-of-art methods. The results showed outstanding improvement over the known leaky ReLU as follows: 99% for AHCD, 95.4% for self-collected data, 90% for HIJJA dataset and 99% for Digit MNIST. The proposed CNN architecture with the proposed optimized leaky ReLU showed a stable accuracy performance and error rates between the training, validation, and testing phases. This indicates that most samples are trained and classified correctly.

Cancer is one of the world’s leading causes of human mortality, and the most prevalent type is breast cancer. However, when diagnosed early, breast cancer may be treated. In this paper, a 5-layer feed-forward neuronet model, trained by a novel fuzzy WASD (weights-and-structure-determination) algorithm, called FUZWASD, is introduced and employed to predict whether the breast cancer is benign or malignant. In general, WASD-trained neuronets are known to overcome the limitations of traditional back-propagation neuronets, including slow training speed and local minimum; however, multi-input WASD-trained neuronets with no dimension explosion weakness are few. In this work, a novel FUZWASD algorithm for training neuronets is modeled by embedding a fuzzy logic controller (FLC) in a WASD algorithm, and a multi-input FUZWASD neuronet (MI-FUZWASDN) model for classification problems with no dimension explosion weakness is proposed. The FUZWASD algorithm uses a FLC to map the input data into a specific interval that enhances the accuracy of the weights-direct-determination (WDD) method. In this way, the FUZWASD algorithm detects the optimal weights and structure of the MI-FUZWASDN using a power softplus activation function and while handling the model fitting and validation. Applications on two diagnostic breast cancer datasets validate and demonstrate the MI-FUZWASDN model’s exceptional learning and predicting performance. In addition, for the intrigued user, we have created a MATLAB kit, which is freely accessible via GitHub, to promote and support the results of this work.

Activation function is an important component of the convolutional neural network. Recently, nonlinear nonmonotonic activation functions such as Swish and Mish have illustrated good performance in deep learning structures. In this paper, we propose a new nonlinear nonmonotonic activation function called Logish, which can be represented by f ( x ) = x · ln [ 1 + sigmoid ( x ) ] . Firstly, we take the logarithmic operation to reduce the numerical range of sigmoid ( x ) + 1 , then we employ variable x to make the negative output have a strong regularization effect. Furthermore, we evaluate the image classification performance of Logish and its variant f ( x ) = α x · ln [ 1 + sigmoid ( β x ) ] in simple and complex networks with top–1 accuracy. Experimental results demonstrate that Logish’s variant ( α = 1 , β = 10 ) can achieve 94.8% top-1 accuracy with ResNet–50 network on CIFAR 10 dataset, and can reach 99.24% top-1 accuracy with DenseNet on MNIST dataset and 88.52% top-1 accuracy with SE-Inception-v4 network on SVHN dataset respectively. It is higher than the Sigmoid, Tanh, ReLU, Swish and Mish activation functions in the corresponding dataset. It also verifies the performance and effectiveness of Logish.

In this paper, we propose a novel multi-manifold feature fusion (MMFF) based deep learning framework for automatic target recognition (ATR) in complex-valued synthetic aperture radar (SAR) images. The conventional real-valued convolutional neural networks (CNN) usually ignore the intrinsic geometry structure of complex numbers, and the existing complex- and manifold-valued CNNs are not in accordance with the multimodality of SAR images. Founded on parallel manifold convolution branches for magnitude and phase data, the proposed MMFF model takes both the underlying geometry and the multimodal information of complex-valued SAR data into consideration, to learn manifold features of military and civilian targets for recognition tasks. Besides, we investigate manifold-valued layers such as activation function and batch normalization to leverage the MMFF architecture as well as the multimodal data. We validate our methods on two public SAR datasets in single- and dual-polarization, and achieve obviously better classification performance with less model parameters, compared with previously proposed real-, complex- and manifold-valued baselines.

Despite many advances, Handwritten Devanagari Character Recognition (HDCR) remains unsolved due to the presence of complex characters. For HDCR, the traditional feature extraction and classification techniques are limited to the datasets developed in the respective laboratory that are not available publicly. A standard benchmarking dataset is not available for HDCR that helps to develop deep learning models. To progress the performance of HDCR, in this study, we produced a dataset of 38,750 images of Devanagari numerals, and vowels are generated and made publicly available for fellow researchers in this domain. This data is collected from more than 3000 subjects of different age groups. Each character is extracted by a segmentation technique proposed here, which is limited to this application. Experiments are conducted on the dataset; three different Convolution Neural Networks (CNN) architecture is developed. 1. CNN, 2. Modified Lenet CNN (MLCNN) and 3. Alexnet CNN (ACNN). A Modified LCNN is proposed by changing the architecture of Lenet 5 CNN. Regular Lenet 5 has $$\mathrm{tanh}(x)$$ as its activation function. Since the Devangari characters are nonlinear, non-linearity is introduced in the Networks by using Rectified Linear Unit. This solves the problem of vanishing gradient problem by $$\mathrm{tanh}(x)$$ . We achieved a recognition rate of 96% on training data and 94% on unseen data using CNN. MLCNN obtained an accuracy rate of 99% and 94% with less computational cost. Whereas, ACNN attained a recognition rate of 99% and 98% on unseen data. A series of experiments were conducted on the data with different combination splits of data and found a minimum loss of 0.001%. Such developments fill a significant percentage of the huge gap between real-world requirements and the actual performance of Devanagari recognizers.

This article introduces a multiple classifier method to improve the performance of concatenate-designed neural networks, such as ResNet and DenseNet, with the purpose of alleviating the pressure on the final classifier. We give the design of the classifiers, which collects the features produced between the network sets, and present the constituent layers and the activation function for the classifiers, to calculate the classification score of each classifier. We use the $$L2 \left( \sqrt{e^x}\right)$$ normalization method to obtain the classifier score instead of the $$Softmax$$ normalization. We also determine the conditions that can enhance convergence. As a result, the proposed classifiers are able to improve the accuracy in the experimental cases significantly and show that the method not only has better performance than the original models, but also produces faster convergence. Moreover, our classifiers are general and can be applied to all classification related concatenate-designed network models.

The lack of interpretability of the neural network algorithm has become the bottleneck of its wide application. We propose a general mathematical framework, which couples the complex structure of the system with the nonlinear activation function to explore the decoupled dimension reduction method of high-dimensional system and reveal the calculation mechanism of the neural network. We apply our framework to some network models and a real system of the whole neuron map of Caenorhabditis elegans. Result shows that a simple linear mapping relationship exists between network structure and network behavior in the neural network with high-dimensional and nonlinear characteristics. Our simulation and theoretical results fully demonstrate this interesting phenomenon. Our new interpretation mechanism provides not only the potential mathematical calculation principle of neural network but also an effective way to accurately match and predict human brain or animal activities, which can further expand and enrich the interpretable mechanism of artificial neural network in the future.

In this article, hardware-based spiking neural networks (SNNs) using capacitor-less positive feedback (PF) neuron devices are designed. It was reported that the PF device can simultaneously process the excitatory and inhibitory signals. The PF device shows very steep subthreshold slope (SS < 1 mV/dec) due to the PF opertaion, leading to low-power and reliable neuron device. The PF devices also show the behavior of leaky integrate and fire (LIF) neuron, which is the most popular neuron model in SNNs. For hardware configuration, the neuron characteristics of PF device are investigated with the transient behavior of the anode current. Based on the PF neuron devices, the SNN shows the accuracy of 98.19% for the Modified National Institute of Standards and Technology (MNIST) database classification in four-hidden layer, fully-connected neural network, which is near the accuracy (98.46%) of the artificial neural networks using rectified linear unit (ReLU) activation function.

The leading cause of cancer-related death globally has been identified as lung cancer. Early lung nodule detection is critical for lung cancer therapy and patient survival. The Gard Cam++ Class Activation Function is used with a squeeze-and-excite network to provide a revolutionary method for differentiating malignant from benign lung nodules on CT scans. The new SENET (Squeeze-and-Excitation Networks) Grad Cam++ module, which combines the features calibration and discrimination benefits of SENET, has been shown to have a substantial potential for improving feature discriminability in lung cancer classification. According to the publicly available LUng Nodule Analysis 2016 (LUNA16) database, when assessed on 1230 nodules, the technique achieved an AUC of 0.9664 and an accuracy of 97.08% (600 malignant and 630 benign). The favorable results demonstrate the robustness of our technique to nodule classification, which we anticipate will be valuable in the future. The technology's objective is to aid radiologists in evaluating diagnostic data and differentiating benign from malignant lung nodules on CT images. To our knowledge, no systematic evaluation of SENET usefulness in classifying lung nodules has been done.

In the real world, the detection of heart diseases is a challenging process. For detecting testicular anomalies with or without congenital heart disease, 2D ultrasound images through a computer-aided diagnostic support system are favorable. Existing work achieves classification through an adaptive neuro-fuzzy inference system (ANFIS) with an F1 score of 0.9673. Although the classification accuracy is good, there are some small changes in detection, leading to misclassification in some miniature valve detection scenarios. To increase the efficiency, a multilayer deep detection perceptron that over performs an ANFIS, a radial basis function neural network, and a multilayer back propagation neural network is proposed. It has the capability of optimizing the classification with higher performance accuracy. During the valuable phases of processing, classification of prenatal ventricular septal defects was achieved with help of the novel multilayer perceptron. Further, gradient descent is used to increase efficiency and improve the filter rate by using a linear algebra activation function. Both backpropagation and the least squares ensured that the smallest error occurred in classification and detection. The accuracy of this proposed method reaches 98%.

In this paper, the monthly long-term precipitation of the city of Ardabil from 1976 to 2020 is simulated by a modern hybrid learning machine. To this end, the Wavelet and Outlier Robust Extreme Learning Machine (ORELM) models are integrated to produce a hybrid model called “Wavelet-Outlier Robust Extreme Learning Machine (WORELM)”. First, the observed data are normalized, and the best normalization coefficients for the study are acquired. About 70% of the observed data are employed to train the artificial intelligence models and the remaining (i.e., 30%) to test them. After that, the optimal mother wavelet and the best decomposition level of the wavelet model are computed. Then, the optimal number of the hidden layer neurons and the best activation function of the ORELM model is chosen by performing a trial and error procedure. In this study, the regularization parameter of the ORELM model is also optimized. Moreover, using the autocorrelation function (ACF), the most influencing lags of the time-series data are detected, and 14 WORELM models are developed. By analyzing the WORELM models, the best model and the effective lags of the time-series data are introduced. The outcomes of the hybrid WORELM model are finally compared with other machine learning (ML) algorithms to prove the superiority of the WORELM model.

In this paper, we propose a multi-stage genetic algorithm that allows to automatically initialize deep multilayer perceptron neural network models to train it for prediction of robot execution failures. The proposed genetic algorithm system is divided on three stages, the first stage consists of initializing number of hidden layers. The second stage aims to fix number of neurons in each hidden layer. The final stage generates the activation function and the optimizer used to train neural network models. The next step is the application of the generated neural network models to predict robot execution failures. The aim of this approach is giving a robot many models so it can better take a more precise decision, since there is no scientific method to choose neural network model, genetic algorithm allows to generate many models automatically. Results obtained in this study show the efficiency of deep neural networks on robotic failures detection, as well as the efficiency of genetic algorithms to generate different models automatically which prevent the manual setup.

Beef is a food ingredient that has a high selling value. Such high prices make some people manipulate sales in markets or other shopping venues, such as mixing beef and pork. The difference between pork and beef is actually from the color and texture of the meat. However, many people do not understand these differences yet. In addition to socialization related to understanding the differences between the two types of meat, another solution is to create a technology that can recognize and differentiate pork and beef. That is what underlies this research to build a system that can classify the two types of meat. Convolutional Neural Network (CNN) is one of the Deep Learning methods and the development of Artificial Intelligence science that can be applied to classify images. Several regularization techniques include Dropout, L2, and Max-Norm were applied to the model and compared to obtain the best classification results and may predict new data accurately. It has known that the highest accuracy of 97.56% obtained from the CNN model by applying the Dropout technique using 0.7 supported by hyperparameters such as Adam's optimizer, 128 neurons in the fully connected layer, ReLu activation function, and 3 fully connected layers. The reason that also underlies the selection of the model is the low error rate of the model, which is only 0.111.Keywords: Beef and Pork, Model, Classification, CNN

Finger-vein verification is a considerable problem to be addressed in a biometric system. A lot of research works have been intended for finger-vein authentication with aid of diverse data mining algorithms. But, the verification accuracy using conventional algorithms was minimal. Also, the time complexity involved during the finger-vein verification was maximal. To overcome the above drawbacks, Steepest Deep Bipolar Cascade Correlative Machine Learning (SDBCCML) technique is proposed. The proposed SDBCCML technique is designed to efficiently perform the finger-vein verification process when considering the large size of the dataset as input. The proposed SDBCCML technique contains three main components namely input, hidden, and output units for effective finger-vein authentication. The input unit in the proposed SDBCCML technique takes a number of finger vein images as input and then sent it to the hidden units. The proposed SDBCCML technique employs more numbers of hidden units with aiming to deeply learn the input finger vein images and thereby find the significant vein features by using the Gabor filter. Subsequently, the discovered vein features at hidden units are forwarded to the output unit. In the proposed SDBCCML technique, the output unit applies a bipolar activation function that compares the extracted vein features with pre-stored templates in the dataset. After that, the output unit gives the verification result. If the output unit result is +1, then the input finger vein image is classified as an authorized person. If the output unit result is −1, then the input finger vein image is classified as an unauthorized person. Thus, the main contribution of the proposed SDBCCML technique increases the authentication performance of finger-vein with higher accuracy and minimal time. The simulation of the proposed SDBCCML technique is conducted using metrics such as accuracy, time complexity, error rate, F-Score, and space complexity for a diverse number of finger-vein images.

During the COVID-19 pandemic, many offline activities are turned into online activities via video meetings to prevent the spread of the COVID 19 virus. In the online video meeting, some micro-interactions are missing when compared to direct social interactions. The use of machines to assist facial expression recognition in online video meetings is expected to increase understanding of the interactions among users. Many studies have shown that CNN-based neural networks are quite effective and accurate in image classification. In this study, some open facial expression datasets were used to train CNN-based neural networks with a total number of training data of 342,497 images. This study gets the best results using ResNet-50 architecture with Mish activation function and Accuracy Booster Plus block. This architecture is trained using the Ranger and Gradient Centralization optimization method for 60000 steps with a batch size of 256. The best results from the training result in accuracy of AffectNet validation data of 0.5972, FERPlus validation data of 0.8636, FERPlus test data of 0.8488, and RAF-DB test data of 0.8879. From this study, the proposed method outperformed plain ResNet in all test scenarios without transfer learning, and there is a potential for better performance with the pre-training model. The code is available at https://github.com/yusufrahadika-facial-expressions-essay.

Purpose of the research.The purpose of this research is to study the problems of the features of teaching technologies of modern artificial neural networks for carrying out the procedure of unambiguous authentication of students according to a pre-formed reference base of digital biometric characteristics of the authorized users in the field of distance educational technologies.In the modern world, artificial neural networks are successfully used in both applied and scientific fields. The problem of authenticating a human personality, implemented using artificial neural networks, finds practical application in solving problems such as the protection of stateandcorporateinformationresources,robotics,accesscontrolsystems,information retrieval, control systems, etc., and is already beginningto find application in the field of distance educational technologies.InMarch2021,theGovernmentoftheRussianFederationdevelopeda decree on the basis of which higher educational institutions areallowed to use distance learning technologies. Conducting remotelyactivitiesofintermediateandfinalcertification,aswellasmonitoringthe current progress of both distance learning students and full-timeandpart-timestudentswitha temporary transition to distance learningin a pandemic, the problem of identifying the student’s personalityarises in order to achieve unambiguous recognition of the authorized users for the purpose of reliable assessment of learning outcomes, which can be solved using modern technologies of artificial neural networks.Materials and methods.Methods of reviewing scientific literature onthe research topic, methods of collecting, structuring and analyzing the information obtained were used as materials and methods.Research results.The results of the study allow us to draw the following conclusions: to solve the problem of authenticating studentsin distance education systems it is first necessary to form the actualbase of biometric characteristics of the authorized users, which willbe compared with the biometric data of the identified users, and forthe recognition procedure, the neural network must be trained in advance on special trainers datasets. The identification procedure must be repeated several times during a session to ensure that the identity of the authorized user is verified.Conclusion.Realizing the set goal to study the problematics oflearningtechnologiesofmodernartificialneuralnetworksforcarryingouttheprocedureofunambiguousauthenticationofstudentsaccordingto a pre-formed reference base of digital biometric characteristicsof authorized users in the field of distance learning technologies,and relying on the results obtained in the course of generalizationandanalysisofexistingexperienceandourownstudies,theauthors identified two independent stages in the algorithm for the implementation of the task of identifying the student’s personality: the formation of a reference base of digital biometric characteristics ofauthorized users and user authentication according to the previouslyformed reference base, and also revealed that when training a neural network, it is necessary to take into account a sufficiently largenumberofdifferent attributesaffectingit. Withaninsufficient numberof training sets (datasets), neural networks begin to perceive errorsas reliable information, which, as a result, will lead to the need toretrain neural networks. With a sufficiently large number of trainingsets (dataset), more versions of dependencies and variability appear, which makes it possible to create rather complex machine learningmodels of neural networks, in which retraining takes the main place.

Deep neural networks have emerged as a widely used and effective means for tackling complex, real-world problems. However, a major obstacle in applying them to safety-critical systems is the great difficulty in providing formal guarantees about their behavior. We present a novel, scalable, and efficient technique for verifying properties of deep neural networks (or providing counter-examples). The technique is based on the simplex method, extended to handle the non-convex Rectified Linear Unit (ReLU) activation function, which is a crucial ingredient in many modern neural networks. The verification procedure tackles neural networks as a whole, without making any simplifying assumptions. We evaluated our technique on a prototype deep neural network implementation of the next-generation airborne collision avoidance system for unmanned aircraft (ACAS Xu). Results show that our technique can successfully prove properties of networks that are an order of magnitude larger than the largest networks that could be verified previously.

Deep neural network (DNN) has been recently used in the field of fault diagnosis, but still their applicability is restricted to high computational complexity. In addition, useless information transformation between adjacent layers of the network could have a negative influence on the diagnosis accuracy. In this paper, a new DNN structure with sparse gate is designed to highlight the role of neurons contributed more by making it directly transfer through layers rather than transfer via an activation function. So it can reduce the computational complexity of network training since only those contributed less are required to be transferred via a nonlinear transformation. The proposed sparse denoising DNN (SD-DNN)-based fault diagnosis method can achieve more accurate diagnosis result with less computational complexity. It shows significant superiority to other-related methods in the case when only small size of training samples polluted by strong noise is available, which is very common for the engineering field of fault diagnosis. The experimental testing of fault diagnosis for rolling bearings verifies the effectiveness of the proposed method.

COVID-19 is a disease or virus that has recently spread worldwide. The disease has also taken many casualties because the virus is notoriously deadly. An examination can be carried out using a chest X-Ray because it costs cheaper compared to swab and PCR tests. The data used in this study was chest X-Ray image data. Chest X-Ray images can be identified using Computer-Aided Diagnosis by utilizing machine learning classification. The first step was the preprocessing stage and feature extraction using the Gray Level Co-Occurrence Matrix (GLCM). The result of the feature extraction was then used at the classification stage. The classification process used was Extreme Learning Machine (ELM). Extreme Learning Machine (ELM) is one of the artificial neural networks with advanced feedforward which has one hidden layer called Single Hidden Layer Feedforward Neural Networks (SLFNs). The results obtained by GLCM feature extraction and classification using ELM achieved the best accuracy of 91.21%, the sensitivity of 100%, and the specificity of 91% at 135° rotation using linear activation function with 15 hidden nodes.

The novel coronavirus-2019 (Covid-19), a contagious disease became a pandemic and has caused overwhelming effects on the human lives and world economy. The detection of the contagious disease is vital to avert further spread and to promptly treat the infected people. The need of automated scientific assisting diagnostic methods to identify Covid-19 in the infected people has increased since less accurate automated diagnostic methods are available. Recent studies based on the radiology imaging suggested that the imaging patterns on X-ray images and Computed Tomography (CT) scans contain leading information about Covid-19 and is considered as a potential automated diagnosis method. Machine learning and deep learning techniques combined with radiology imaging can be helpful for accurate detection of the disease. A deep learning approach based on the multilayer-Spatial Convolutional Neural Network for automatic detection of Covid-19 using chest X-ray images and CT scans is proposed in this paper. The proposed model, named as the Multilayer Spatial Covid Convolutional Neural Network (MSCovCNN), provides an automated accurate diagnostics for Covid-19 detection. The proposed model showed 93.63% detection accuracy and 97.88% AUC (Area Under Curve) for chest x-ray images and 91.44% detection accuracy and 95.92% AUC for chest CT scans, respectively. We have used 5-tiered 2D-CNN frameworks followed by the Artificial Neural Network (ANN) and softmax classifier. In the CNN each convolution layer is followed by an activation function and a Maxpooling layer. The proposed model can be used to assist the radiologists in detecting the Covid-19 and confirming their initial screening.

Convolutional Neural Networks (CNNs) have proven to be an effective approach for solving image classification problems. The output, the accuracy and the computational efficiency of a CNN are determined mainly by the architecture, the convolutional filters, and the activation functions. Based on the importance of an activation function, in this paper, nine new activation functions based on combinations of classical functions such as ReLU and sigmoid are presented. Also, a study about the effects caused by the activation functions in the performance of a CNN is presented. First, every new function is described, also, their graphs, analytic forms and derivatives are presented. Then, a traditional CNN model with each new activation function is used to classify three 10-class databases: MNIST, Fashion MNIST and a handwritten digit database created by us. Experimental results illustrate that some of the proposed activation functions lead to better performances on classifying than classical activation functions. Moreover, our study demonstrated that the accuracy of a CNN could be increased by 1.18% with the new proposed activation functions.

Achieving high detection accuracy of pavement cracks with complex textures under different lighting conditions is still challenging. In this context, an encoder-decoder network-based architecture named CrackResAttentionNet was proposed in this study, and the position attention module and channel attention module were connected after each encoder to summarize remote contextual information. The experiment results demonstrated that, compared with other popular models (ENet, ExFuse, FCN, LinkNet, SegNet, and UNet), for the public dataset, CrackResAttentionNet with BCE loss function and PRelu activation function achieved the best performance in terms of precision (89.40), mean IoU (71.51), recall (81.09), and F1 (85.04). Meanwhile, for a self-developed dataset (Yantai dataset), CrackResAttentionNet with BCE loss function and PRelu activation function also had better performance in terms of precision (96.17), mean IoU (83.69), recall (93.44), and F1 (94.79). In particular, for the public dataset, the precision of BCE loss and PRelu activation function was improved by 3.21. For the Yantai dataset, the results indicated that the precision was improved by 0.99, the mean IoU was increased by 0.74, the recall was increased by 1.1, and the F1 for BCE loss and PRelu activation function was increased by 1.24.

For the lower limb rehabilitation robot, how to better realize intention recognition is the key issue in the practical application. Recognition of the patient’s movement intention is a challenging research work, which needs to be studied from the shallow to the deep. Specifically, it is necessary to ensure that the movement intention of the normal person can be accurately recognized, and then improve the model to realize the recognition of the movement intention of the patients. Therefore, before studying the patient’s movement intention, it is essential to consider the normal person first, which is also for safety considerations. In recent years, a new Hill-based muscle model has been demonstrated to be capable of directly estimating the joint angle intention in an open-loop form. On this basis, by introducing a recurrent neural network (RNN), the whole prediction process can achieve more accuracy in a closed-loop form. However, for the traditional RNN algorithms, the activation function must be convex, which brings some limitations to the solution of practical problems. Especially, when the convergence speed of the traditional RNN model is limited in the practical applications, as the error continues to decrease, the convergence performance of the traditional RNN model will be greatly affected. To this end, a projected recurrent neural network (PRNN) model is proposed, which relaxes the condition of the convex function and can be used in the saturation constraint case. In addition, the corresponding theoretical proof is given, and the PRNN method with saturation constraint has been successfully applied in the experiment of intention recognition of lower limb movement compared with the traditional RNN model.

This article introduces a method for realizing the Gaussian activation function of radial-basis (RBF) neural networks with their hardware implementation on field-programmable gaits area (FPGAs). The results of modeling of the Gaussian function on FPGA chips of different families have been presented. RBF neural networks of various topologies have been synthesized and investigated. The hardware component implemented by this algorithm is an RBF neural network with four neurons of the latent layer and one neuron with a sigmoid activation function on an FPGA using 16-bit numbers with a fixed point, which took 1193 logic matrix gate (LUTs—LookUpTable). Each hidden layer neuron of the RBF network is designed on an FPGA as a separate computing unit. The speed as a total delay of the combination scheme of the block RBF network was 101.579 ns. The implementation of the Gaussian activation functions of the hidden layer of the RBF network occupies 106 LUTs, and the speed of the Gaussian activation functions is 29.33 ns. The absolute error is ± 0.005. The Spartan 3 family of chips for modeling has been used to get these results. Modeling on chips of other series has been also introduced in the article. RBF neural networks of various topologies have been synthesized and investigated. Hardware implementation of RBF neural networks with such speed allows them to be used in real-time control systems for high-speed objects.

Cloud computing provides different kind of services for users and provides with the help of internet. The Infrastructure as a service is a service model that provides virtual computing resources such as, networking, hardware, and storage services as needed for users. However, cloud-hosting initialization takes several minutes delay in the hardware resource allocation process. To resolve this issue, we need to predict the future amount of computing. In this paper, we propose a convolutional neural network and long short-term memory model for predicting multivariate workload which are the central processing unit, memory, and network usage. Firstly, the input data are analyzed by the vector auto regression method which filters the linear interdependencies among the multivariate data. Then, the residual data are computed and entered into the convolutional neural network layer which extracts complex features of each of the virtual machine usage components after the long short-term memory neural network, which is suitable for modeling temporal information of irregular trends in time series components in the proposed hybrid model latest activation function scaled polynomial constant unit used. The proposed model is compared with other predictive models. Based on the result, the proposed model shows the accuracy rate enhanced by approximately 3.8% to 10.9% and the error percentage rate also reduces by approximately 7% to 8.5% as compared to the other different models. It means the proposed technique improves central processing unit, memory, disk, and network usage in the network taking less amount of time due to the good predication approach compare to other models. In future research, we implement the proposed technique for VM energy section as well data predication system in cloud data center.

Estrus detection in cattle is an important factor in livestock farming. With timely estrus detection, cattle are artificially fertilized and isolated for safety, which directly affects the productivity of livestock farms. Estrus can be successfully detected by identifying the mating posture of cattle. Therefore, in this paper, we propose the identification of cattle mating posture based on video inputs for prompt estrus detection. A deep learning-based object detection network that focuses on real-time processing with high processing speeds is applied. The use of deep learning-based object detection shows high accuracy, even with noise robustness. The performance of the network is improved through the inclusion of an additional layer and a new activation function. The composition of the additional layer enables training by extracting more features required for object detection. The application of the new activation function, Mish, which has a smoother curve, allows for better generalization and improves the accuracy of the results. The data needed for training were gathered by installing cameras at a livestock farm, and various datasets were used depending on camera placement. The results of this study were verified by the evaluation of four networks using test datasets containing image and video data from different environments. The identification of the mating posture of cattle attained 98.5% precision, 97.2% recall, and 97.8% accuracy.

Squeeze-and-Excitation (SE) Networks won the last ImageNet Large-Scale Visual Recognition Challenge (ILSVRC) classification competition and is very popular in today’s vision community. The SE block is the core of Squeeze-and-Excitation Network (SENet), which adaptively recalibrates channel-wise features and suppresses less useful ones. Since SE blocks can be directly used in existing models and effectively improve performance, SE blocks are widely used in a variety of tasks. In this paper, we propose a novel Parametric Sigmoid (PSigmoid) to enhance the SE block. We named the new module PSigmoid SE (PSE) block. The PSE block can not only suppress features in a channel-wise manner, but also enhance features. We evaluate the performance of our method on four common datasets including CIFAR-10, CIFAR-100, SVHN and Tiny ImageNet. Experimental results show the effectiveness of our method. In addition, we compare the differences between the PSE block and the SE block through a detailed analysis of the configuration. Finally, we use a combination of PSE block and SE block to obtain better performance.

In traditional 3D reconstruction methods, using a single view to predict the 3D structure of an object is a very difficult task. This research mainly discusses human pose recognition and estimation based on 3D multiview basketball sports dataset. The convolutional neural network framework used in this research is VGG11, and the basketball dataset Image Net is used for pretraining. This research uses some modules of the VGG11 network. For different feature fusion methods, different modules of the VGG11 network are used as the feature extraction network. In order to be efficient in computing and processing, the multilayer perceptron in the network model is implemented by a one-dimensional convolutional network. The input is a randomly sampled point set, and after a layer of perceptron, it outputs a feature set of n × 16. Then, the feature set is sent to two network branches, one is to continue to use the perceptron method to generate the feature set of n × 1024, and the other network is used to extract the local features of points. After the RGB basketball sports picture passes through the semantic segmentation network, a picture containing the target object is obtained, and the picture is input to the constructed feature fusion network model. After feature extraction is performed on the RGB image and the depth image, respectively, the RGB feature, the local feature of the point cloud, and the global feature are spliced and fused to form a feature vector of N × 1152. There are three branches for this vector network, which, respectively, predict the object position, rotation, and confidence. Among them, the feature dimensionality reduction is realized by one-dimensional convolution, and the activation function is the ReLU function. After removing the feature mapping module, the accuracy of VC-CNN_v1 dropped by 0.33% and the accuracy of VC-CNN_v2 dropped by 0.55%. It can be seen from the research results that the addition of the feature mapping module improves the recognition effect of the network to a certain extent