Baseline Convolutional Neural Network Research Articles

Hybrid nonlinear convolution filters for image recognition

Typical convolutional filter only extract features linearly. Although nonlinearities are introduced into the feature extraction layer by using activation functions and pooling operations, they can only provide point-wise nonlinearity. In this paper, a Gaussian convolution for extracting nonlinear features is proposed, and a hybrid nonlinear convolution filter consisting of baseline convolution, Gaussian convolution and other nonlinear convolutions is designed. It can efficiently achieve the fusion of linear features and nonlinear features while preserving the advantages of traditional linear convolution filter in feature extraction. Extensive experiments on the benchmark datasets MNIST, CIFAR10, and CIFAR100 show that the hybrid nonlinear convolutional neural network has faster convergence and higher image recognition accuracy than the traditional baseline convolutional neural network.

Polyp-Net: A Multimodel Fusion Network for Polyp Segmentation

Computer-aided diagnosis of disease primarily depends on proper vision-based measurement (VBM). The traditional approach followed for diagnosis of colorectal cancer includes a manual screening of colorectum via a colonoscope and resection of polyps for histopathological analysis to decide the grade of malignancy. This procedure is time-consuming and expensive, and removal of benign polyp for analysis signifies the inefficiency of the diagnosis system. These drawbacks inspired us to develop an automatic vision-based analysis method for preliminary in vivo malignancy analysis of the polyp region. In this work, we have proposed a fusion-based polyp segmentation network, namely, Polyp-Net. Recently, convolutional neural networks (CNNs) have shown immense success in the domain of medical image analysis as it can exploit in-depth significant features with high discrimination capabilities. Therefore, motivated by these insights, we have proposed an enriched version of CNN with a nascent pooling mechanism, namely dual-tree wavelet pooled CNN (DT-WpCNN). The resultant segmented mask contains some surplus high-intensity regions apart from the polyp region. These shortcomings are avoided using a new variation of the region-based level-set method, namely, the local gradient weighting-embedded level-set method (LGWe-LSM), which shows a significant reduction of false-positive rate. The pixel-level fusion of the two enhanced methods shows more potentiality in the segmentation of the polyp regions. Our proposed network is trained on CVC-colon DB and tested on CVC-clinic DB. It achieves a dice score of 0.839, volume-similarity of 0.863, precision of 0.836, recall of 0.811, F1-score of 0.823, F2-score of 0.815, and Hausdorff distance of 21.796 which outperforms the existing baseline CNN’s and recent state-of-the-art methods.

Improving convolutional neural network for text classification by recursive data pruning

In spite of the state-of-the-art performance of deep neural networks, shallow neural networks are still the choice in applications with limited computing and memory resources. Convolutional neural network (CNN), in particular the one-convolutional-layer CNN, is a widely-used shallow neural network in natural language processing tasks such as text classification. However, it was found that CNNs may misfit to task-irrelevant words in dataset, which in turn leads to unsatisfactory performance. To alleviate this problem, attention mechanism can be integrated into CNN, but this takes up the limited resources. In this paper, we propose to address the misfitting problem from a novel angle - pruning task-irrelevant words from the dataset. The proposed method evaluates the performance of each convolutional filter based on its discriminative power of the feature generated at the pooling layer, and prunes words captured by the poorly-performed filters. Experiment results show that our proposed model significantly outperforms the CNN baseline model. Moreover, our proposed model produces performance similar to or better than the benchmark models (attention integrated CNNs) while demanding less parameters and FLOPs, and is therefore a choice model for resource limited scenarios, such as mobile applications.

AdaBoost-CNN: An adaptive boosting algorithm for convolutional neural networks to classify multi-class imbalanced datasets using transfer learning

Ensemble models achieve high accuracy by combining a number of base estimators and can increase the reliability of machine learning compared to a single estimator. Additionally, an ensemble model enables a machine learning method to deal with imbalanced data, which is considered to be one of the most challenging problems in machine learning. In this paper, the capability of Adaptive Boosting (AdaBoost) is integrated with a Convolutional Neural Network (CNN) to design a new machine learning method, AdaBoost-CNN, which can deal with large imbalanced datasets with high accuracy. AdaBoost is an ensemble method where a sequence of classifiers is trained. In AdaBoost, each training sample is assigned a weight, and a higher weight is set for a training sample that has not been trained by the previous classifier. The proposed AdaBoost-CNN is designed to reduce the computational cost of the classical AdaBoost when dealing with large sets of training data, through reducing the required number of learning epochs for its ingredient estimator. AdaBoost-CNN applies transfer learning to sequentially transfer the trained knowledge of a CNN estimator to the next CNN estimator, while updating the weights of the samples in the training set to improve accuracy and to reduce training time. Experimental results revealed that the proposed AdaBoost-CNN achieved 16.98% higher accuracy compared to the classical AdaBoost method on a synthetic imbalanced dataset. Additionally, AdaBoost-CNN reached an accuracy of 94.08% on 10,000 testing samples of the synthetic imbalanced dataset, which is higher than the accuracy of the baseline CNN method, i.e. 92.05%. AdaBoost-CNN is computationally efficient, as evidenced by the fact that the training simulation time of the proposed method is 47.33 s, which is lower than the training simulation time required for a similar AdaBoost method without transfer learning, i.e. 225.83 s on the imbalanced dataset. Moreover, when compared to the baseline CNN, AdaBoost-CNN achieved higher accuracy when applied to four other benchmark datasets including CIFAR-10 and Fashion-MNIST. AdaBoost-CNN was also applied to the EMNIST datasets, to determine its impact on large imbalanced classes, and the results demonstrate the superiority of the proposed method compared to CNN.

Facial Expression Recognition Based on Weighted-Cluster Loss and Deep Transfer Learning Using a Highly Imbalanced Dataset.

Facial expression recognition (FER) is a challenging problem in the fields of pattern recognition and computer vision. The recent success of convolutional neural networks (CNNs) in object detection and object segmentation tasks has shown promise in building an automatic deep CNN-based FER model. However, in real-world scenarios, performance degrades dramatically owing to the great diversity of factors unrelated to facial expressions, and due to a lack of training data and an intrinsic imbalance in the existing facial emotion datasets. To tackle these problems, this paper not only applies deep transfer learning techniques, but also proposes a novel loss function called weighted-cluster loss, which is used during the fine-tuning phase. Specifically, the weighted-cluster loss function simultaneously improves the intra-class compactness and the inter-class separability by learning a class center for each emotion class. It also takes the imbalance in a facial expression dataset into account by giving each emotion class a weight based on its proportion of the total number of images. In addition, a recent, successful deep CNN architecture, pre-trained in the task of face identification with the VGGFace2 database from the Visual Geometry Group at Oxford University, is employed and fine-tuned using the proposed loss function to recognize eight basic facial emotions from the AffectNet database of facial expression, valence, and arousal computing in the wild. Experiments on an AffectNet real-world facial dataset demonstrate that our method outperforms the baseline CNN models that use either weighted-softmax loss or center loss.

LEARNING TO DETECT BRAIN LESIONS FROM NOISY ANNOTATIONS.

Supervised training of deep neural networks in medical imaging applications relies heavily on expert-provided annotations. These annotations, however, are often imperfect, as voxel-by-voxel labeling of structures on 3D images is difficult and laborious. In this paper, we focus on one common type of label imperfection, namely, false negatives. Focusing on brain lesion detection, we propose a method to train a convolutional neural network (CNN) to segment lesions while simultaneously improving the quality of the training labels by identifying false negatives and adding them to the training labels. To identify lesions missed by annotators in the training data, our method makes use of the 1) CNN predictions, 2) prediction uncertainty estimated during training, and 3) prior knowledge about lesion size and features. On a dataset of 165 scans of children with tuberous sclerosis complex from five centers, our method achieved better lesion detection and segmentation accuracy than the baseline CNN trained on the noisy labels, and than several alternative techniques.

Construction activity recognition with convolutional recurrent networks

Although heavy equipment is an indispensable resource in many construction projects, it is often underutilized. Inefficient usage patterns and frequent idling contribute to increased emissions and project costs. Efforts to improve usage patterns often begin with activity tracking. Recent research into automated activity tracking has leveraged sensing devices and Internet-of-Things (IoT) frameworks to power machine learning models that can predict the behaviors of monitored equipment. However, shallow machine learning models require complex manual feature engineering that could be further automated with more recent deep learning approaches. Deep learning approaches not only increase automation but also promise improved accuracies by avoiding biases introduced by manual feature design. This paper proposes a construction equipment activity recognition framework that uses deep learning architectures to predict the activities of heavy construction equipment monitored via accelerometers and applies this framework to a roller compactor and an excavator performing real work. The performance of a simple baseline convolutional neural network (CNN) is compared to a hybrid network that contains both convolutional and recurrent long short-term memory (LSTM) layers. The hybrid model outperforms the baseline model in all instances studied. In the task of classifying the activities of the roller compactor, the hybrid model achieves a validation accuracy of 77.1% when presented with six activities and a validation accuracy of 96.2% when distinguishing only direction. In the task of classifying seven activities of the excavator, the hybrid model achieves a validation accuracy of 77.6%, with some confusion between isolated activities and a Various category that includes elements of the isolated activities. With the Various category removed, the hybrid model achieves a validation accuracy of 90.7%. This study demonstrates that deep learning frameworks can model the activities of construction equipment with high accuracy. In particular, this work shows that convolutional and LSTM layers can each form effective parts of deep learning models that characterize equipment activities based on accelerometer data, and furthermore that these components can produce more effective models when combined. The findings of this study can be leveraged by researchers and industry professionals to develop reliable automated activity recognition systems for tracking and monitoring equipment performance and for measuring the productivity and the efficiency of the work performed.

Enhancement of surgical hand gesture recognition using a capsule network for a contactless interface in the operating room

Background and objective:Hand gesture recognition systems in operating rooms (ORs) are crucial for browsing and controlling computer-aided devices, which have been developed to decrease the risk of contamination during surgical procedures. Methods:We proposed the use of hand gesture recognition to enhance accuracies and recognition areas with the capsule network (CapsNet) of deep neural network and Leap Motionâ Our method includes the i) extraction and preprocessing of infrared (IR) images (60 frames per second) from Leap Motion™, ii) training of various types of networks, and iii) gesture recognition evaluation in the OR. We trained the images of training dataset (N=903) and tested images (N=100) using five types of surgical hand gestures including hovering, grab, click, one peak, and two peaks by 10 subjects with various types of augmentation methods including rotate (0∘, 90∘, 180∘), scale, translation, illumination, and resize. Results: CapsNet achieved a classification accuracy of 86.46% (around 10% improvement) compared with 73.67% for the baseline convolutional neural network (CNN) and 76.4% for VGG16. Conclusions: In conclusion, the accuracy of hand gesture recognition with CapsNet was better than that of conventional CNNs, which could be used to navigate and manipulate various types of computer-aided devices and applications through contactless gesture interaction.

Improving cardiac MRI convolutional neural network segmentation on small training datasets and dataset shift: A continuous kernel cut approach.

Cardiac magnetic resonance imaging (MRI) provides a wealth of imaging biomarkers for cardiovascular disease care and segmentation of cardiac structures is required as a first step in enumerating these biomarkers. Deep convolutional neural networks (CNNs) have demonstrated remarkable success in image segmentation but typically require large training datasets and provide suboptimal results that require further improvements. Here, we developed a way to enhance cardiac MRI multi-class segmentation by combining the strengths of CNN and interpretable machine learning algorithms. We developed a continuous kernel cut segmentation algorithm by integrating normalized cuts and continuous regularization in a unified framework. The high-order formulation was solved through upper bound relaxation and a continuous max-flow algorithm in an iterative manner using CNN predictions as inputs. We applied our approach to two representative cardiac MRI datasets across a wide range of cardiovascular pathologies. We comprehensively evaluated the performance of our approach for two CNNs trained with various small numbers of training cases, tested on the same and different datasets. Experimental results showed that our approach improved baseline CNN segmentation by a large margin, reduced CNN segmentation variability substantially, and achieved excellent segmentation accuracy with minimal extra computational cost. These results suggest that our approach provides a way to enhance the applicability of CNN by enabling the use of smaller training datasets and improving the segmentation accuracy and reproducibility for cardiac MRI segmentation in research and clinical patient care.

PolSAR Image Classification Based on Low-Frequency and Contour Subbands-Driven Polarimetric SENet

In order to more efficiently mine the features of PolSAR images and build a more suitable classification model that combines the features of the polarimetric domain and the spatial domain, this article proposes a PolSAR image classification method, called low-frequency and contour subbands-driven polarimetric squeeze-and-excitation network (LC-PSENet). First, the proposed LC-PSENet introduces the nonsubsampled Laplacian pyramid to decompose polarimetric feature maps, so as to construct a multichannel PolSAR image based on the low-frequency subband and contour subband of these maps. It guides the network to perform feature mining and selection in the subbands of each polarimetric map in a supervised way, automatically balancing the contributions of polarimetric features and their subbands and the influence of interference information such as noise, making the network learning more efficient. Second, the method introduces squeeze-and-excitation operation in the convolutional neural network (CNN) to perform channel modeling on the polarimetric feature subbands. It strengthens the learning of the contributions of local maps of the polarimetric features and subbands, thereby, effectively combining the features of the polarimetric domain and the spatial domain. Experiments on the datasets of Flevoland, The Netherlands, and Oberpfaffenhofen show that the proposed LC-PSENet achieves overall accuracies of 99.66%, 99.72%, and 95.89%, which are 0.87%, 0.27%, and 1.42% higher than the baseline CNN, respectively. The isolated points in the classification results are obviously reduced, and the distinction between boundary and nonboundary is more clear and delicate. Also, the method performs better than many current state-of-the-art methods in terms of classification accuracy.

Identity Recognition by Walking Outdoors Using Multimodal Sensor Insoles

Recently, gait attracts attention as a practical biometric for devices that naturally possess walking pattern sensing. In the present study, we explored the feasibility of using a multimodal smart insole for identity recognition. We used sensor insoles designed and implemented by us to collect kinetic and kinematic data from 59 participants that walked outdoors. Then, we evaluated the performance of four neural network architectures, which are a baseline convolutional neural network (CNN), a CNN with a multi-stage feature extractor, a CNN with an extreme learning machine classifier using sensor-level fusion and CNN with extreme learning machine classifier using feature-level fusion. The networks were trained with segmented insole data using 0%, 50%, and 70% segmentation overlap, respectively. For 70% segmentation overlap and both-side data, we obtained mean accuracies of 72.8% ±0.038, 80.9% ±0.036, 80.1% ±0.021 and 93.3% ±0.009, for the four networks, respectively. The results suggest that multimodal sensor-enabled footwear could serve biometric purposes in the next generation of body sensor networks.

Capsule Network for Cyberthreat Detection

In cybersecurity, analyzing social network data has become an essential research area due to its property of providing real-time updates about real-world events. Studies have shown that Twitter can contain information about security threats before some specialized sites. Thus, the classification of tweets into security-related and not security-related can help with early warnings for such attacks. In this study, the use of a capsule network (CapsNet), the new deep learning algo-rithm, is investigated for the first time in the field of security attack detection using Twitter. The aim was to increase the accuracy of tweet classification by using CapsNet rather than a convolutional neural network (CNN). To achieve the research objective, the original implementation of CapsNet with dynamic routing is adapted to be suitable for text analysis using tweet data set. A random search technique was used to tune the model’s hyperparameters. The experimental results showed that CapsNet exceeded the baseline CNN on the same data set, with accuracy of 92.21% and a 92.2% F1 score; also, word2vec embedding performed better than a random initialization.

Hierarchical Text Classification Using CNNs with Local Approaches

In this paper, we discuss the application of convolutional neural networks (CNNs) for hierarchical text classification using local top-down approaches. We present experimental results implementing a local classification per node approach, a local classification per parent node approach, and a local classification per level approach. A 20Newsgroup hierarchical training dataset with more than 20 categories and three hierarchical levels was used to train the models. The experiments involved several variations of hyperparameters settings such as batch size, embedding size, and number of available examples from the training dataset, including two variation of CNN model text embedding such as static (stat) and random (rand). The results demonstrated that our proposed use of CNNs outperformed flat CNN baseline model and both the flat and hierarchical support vector machine (SVM) and logistic regression (LR) baseline models. In particular, hierarchical text classification with CNN-stat models using local per parent node and local per level approaches achieved compelling results and outperformed the former and latter state-of-the-art models. However, using CNN with local per node approach for hierarchical text classification underperformed and achieved worse results. Furthermore, we performed a detailed comparison between the proposed hierarchical local approaches with CNNs. The results indicated that the hierarchical local classification per level approach using the CNN model with static text embedding achieved the best results, surpassing the flat SVM and LR baseline models by 7 % and 13 %, surpassing the flat CNN baseline by 5 %, and surpassing the h-SVM and h-LR models by 5 % and 10 %, respectively.

Automatic Segmentation of Stroke Lesions in Non-Contrast Computed Tomography Datasets With Convolutional Neural Networks

Non-contrast computed tomography (NCCT) is commonly used for volumetric follow-up assessment of ischemic strokes. However, manual lesion segmentation is time-consuming and subject to high inter-observer variability. The aim of this study was to develop and establish a baseline convolutional neural network (CNN) model for automatic NCCT lesion segmentation. A total of 252 multi-center clinical NCCT datasets, acquired from 22 centers, and corresponding manual segmentations were used to train (204 datasets) and validate (48 datasets) a 3D multi-scale CNN model for lesion segmentation. Post-processing methods were implemented to improve the CNN-based lesion segmentations. The final CNN model and post-processing method was evaluated using 39 out-of-distribution holdout test datasets, acquired at seven centers that did not contribute to the training or validation datasets. Each test image was segmented by two or three neuroradiologists. The Dice similarity coefficient (DSC) and predicted lesion volumes were used to evaluate the segmentations. The CNN model achieved a mean DSC score of 0.47 on the validation NCCT datasets. Post-processing significantly improved the DSC to 0.50 (P <; 0.01). On the holdout test set, the CNN model achieved a mean DSC score of 0.42, which was also significantly improved to 0.45 (P <; 0.05) by post-processing. Importantly, the automatically segmented lesion volumes were not significantly different from the lesion volumes determined by the expert observers (P >0.05) and showed excellent agreement with manual lesion segmentation volumes (intraclass correlation coefficient, ICC = 0.88). The proposed CNN model can automatically and reliably segment ischemic stroke lesions in clinical NCCT datasets. Post-processing techniques can further improve accuracy. As the model was trained and evaluated on datasets from multiple centers, it is broadly applicable and is publicly available.

Unsupervised image translation using adversarial networks for improved plant disease recognition

Acquisition of data in task-specific applications of machine learning like plant disease recognition is a costly endeavor owing to the requirements of professional human diligence and time constraints. In this paper, we present a pipeline that uses GANs in an unsupervised image translation environment to improve learning with respect to the data distribution in a plant disease dataset, reducing the partiality introduced by acute class imbalance and hence shifting the classification decision boundary towards better performance. The empirical analysis of our method is demonstrated on a limited dataset of 2789 tomato plant disease images, highly corrupted with an imbalance in the 9 disease categories. First, we extend the state of the art for the GAN-based image-to-image translation method by enhancing the perceptual quality of the generated images and preserving the semantics. We introduce AR-GAN, where in addition to the adversarial loss, our synthetic image generator optimizes on Activation Reconstruction loss (ARL) function that optimizes feature activations against the natural image. We present visually more compelling synthetic images in comparison to most prominent existing models and evaluate the performance of our GAN framework in terms of various datasets and metrics. Second, we evaluate the performance of a baseline convolutional neural network classifier for improved recognition using the resulting synthetic samples to augment our training set and compare it with the classical data augmentation scheme. We observe a significant improvement in classification accuracy (+5.2%) using generated synthetic samples as compared to (+0.8%) increase using classic augmentation in an equal class distribution environment.

Memory-Reduced Network Stacking for Edge-Level CNN Architecture With Structured Weight Pruning

This paper presents a novel stacking and multi-level indexing scheme for convolutional neural networks (CNNs) used in energy-limited edge-level systems. Basically, the proposed scheme offers multiple accuracy modes by adopting a structured weight pruning method that enables a CNN to be trained once with multiple pruning ratios and thereby allows for adaptive energy-accuracy trade-offs. The memory overhead required to store several different networks is kept to a minimum by adopting a novel method for including smaller lower-accuracy networks as subnetworks of larger higher-accuracy networks and by using a unique multi-level indexing scheme that can effectively store compressed weight data for the proposed stacked-CNN architecture. Experimental results show that the proposed method successfully reduces the memory footprint by up to 33% when compared to a baseline CNN architecture. An FPGA-based multi-mode CNN accelerator that implements the proposed scheme has been designed. Energy usage analysis with a case study shows that the inference energy required for on-device CNN processing can be reduced by up to 1.94 times over the baseline design.

Compressing the CNN architecture for in-air handwritten Chinese character recognition

Since the convolutional neural network (CNN) has brought great breakthroughs in the field of computer vision, it recently has been introduced to the in-air handwritten Chinese character recognition (IAHCCR) to achieve better recognition performance. However, the CNN is typically over-parameterized and contains lots of redundant filters or parameters. This leads the CNN to suffer from huge computation cost and considerable storage usage, limiting its deployments to resource-constrained devices like mobile phones and intelligent TVs. In this paper, we propose a unified algorithm to effectively compress the CNN for IAHCCR with little accuracy loss. Specifically, we first utilize the channel pruning strategy to simplify the network structure, and then adopt the network quantization technique to represent parameters with lower precision. We conduct experiments on the in-air handwriting dataset IAHCC-UCAS2016, where the baseline CNN achieves the state-of-the-art accuracy of 95.33% with 15.5 MB of storage. After the compression, we achieve 12.4 × storage saving and 1.7 × theoretical acceleration with only 0.17% accuracy loss. Moreover, evaluations on other benchmark datasets including ICDAR-2013 and MNIST further demonstrate the effectiveness of the proposed method.

Cross-Domain Deep Visual Feature Generation for Mandarin Audio–Visual Speech Recognition

There has been a long term interest in using visual information to improve automatic speech recognition (ASR) system performance. Both audio and visual information are required in conventional audio visual speech recognition (AVSR) systems. This limits their wider applications when visual modality is not present. To this end, one possible solution is to use acoustic-to-visual (A2V) inversion techniques to generate visual features. Previous research in this direction used synthetic acoustic-articulatory parallel data in inversion model training. The acoustic mismatch between the audio-visual (AV) parallel data and target data was not considered. In addition, the target language to apply these technologies has been focused on English. In this article, a real 3D Audio-Visual Mandarin Continuous Speech (3DAV-MCS) corpus was used to train deep neural network based A2V inversion models. Cross-domain adaptation of the inversion models allows suitable visual features to be generated from acoustic data of mismatched domains. The proposed cross-domain deep visual feature generation techniques were evaluated on two state-of-the-art Mandarin speech recognition tasks: DAPRA GALE broadcast transcription and BOLT conversational telephone speech recognition. The AVSR systems constructed using the cross-domain generated visual features consistently outperformed the baseline convolutional neural network (CNN) ASR systems by up to 3.3% absolute (9.1% relative) character error rate (CER) reductions after both speaker adaptive training and sequence discriminative training were performed.

Filling Voids in Elevation Models Using a Shadow-Constrained Convolutional Neural Network

We explore the use of convolutional neural networks (CNNs) for filling voids in digital elevation models (DEM). We propose a baseline approach using a fully convolutional network to predict complete from incomplete DEMs, which is trained in a supervised fashion. We then extend this to a shadow-constrained CNN (SCCNN) by introducing additional loss functions that encourage the restored DEM to adhere to geometric constraints implied by cast shadows. At the training time, we use automatically extracted cast shadow maps and known sun directions to compute the shadow-based supervisory signal in addition to the direct DEM supervision. At the test time, our network directly predicts restored DEMs from an incomplete DEM. One key advantage of our SCCNN model is that it is characterized by both CNN data inference and geometric shadow cues. It thus avoids data restoration that may violate shadowing conditions. Both our baseline CNN and SCCNN outperform the inverse distance weighting (IDW)-based interpolation method, with the shadow supervision enabling SCCNN to obtain the best performance.

Learn the Temporal-Spatial Feature of sEMG via Dual-Flow Network

Surface electromyography (sEMG) signals have been widely used in human–machine interaction, providing more nature control expedience for external devices. However, due to the instability of sEMG, it is hard to extract consistent sEMG patterns for motion recognition. This paper proposes a dual-flow network to extract the temporal-spatial feature of sEMG for gesture recognition. The proposed network model uses convolutional neural network (CNN) and long short-term memory methods (LSTM) to, respectively, extract the spatial feature and temporal feature of sEMG, simultaneously. These features extracted by CNN and LSTM are merged into temporal-spatial feature to form an end-to-end network. A dataset was constructed for testing the performance of the network. In this database, the average recognition accuracy by using our dual-flow model reached 78.31%, which was improved by 6.69% compared to the baseline CNN (71.67%). In addition, NinaPro DB1 is also used to evaluate the proposed methods, receiving 1.86% higher recognition accuracy than the baseline CNN classifier. It is believed that the proposed dual-flow network owns the merit in extracting stable sEMG feature for gesture recognition, and can be further applied into practical applications.