Research On Capsule Network Based On Sharing Transformation Matrix And Its Cross-view Gait Recognition Method

With the improvement of people’s awareness of public safety and privacy,more and more biometric-based recognition technologies are applied to the security field.However,during the epidemic,traditional biological features such as iris,face,fingerprints,etc.,have requirements for close contact and no shelter during identification,which will bring the risk of infection,and the gait has the advantages of requiring no contact,non-invasion and easy detection,which can solve the above problems.However,for the cross-view gait recognition,the contour of pedestrian will change with the viewpoint of the image capture device and the pedestrian.For this reason,based on the improved capsule network,this paper studies the cross-view gait recognition.The main work is as follows:1.Studying on capsule networks sharing transformation matrix.the traditional capsule network has many parameters,which is difficult to apply to high-resolution images and multisample data set,and the capsule represents the local entity under the current viewpoint,and is projected to another viewpoint through the transformation matrix,therefore,in order to reduce the network training parameters and optimize the dynamic routing algorithm,the capsule network with sharing transformation matrix is proposed.The experimental result on MNIST and Fashion-MNIST with the traditional capsule network shows that the capsule network with sharing transformation matrix can greatly reduce the network parameters,improve network generalization performance,and keep the recognition accuracy basically unchanged.2.Studying on cross-view gait recognition fused with view features.the contour of pedestrians varies with the change of device’s viewpoint,which will affect the accuracy of gait recognition,for this reason,a cross-view gait recognition model fused with view features was proposed to improve the model’s anti-interference ability against changes in viewpoints.First,a viewpoints recognition model based on the sharing transformation matrix capsule network was established to verify the feasibility of identifying different viewpoints and extracting view features.Then,a gait feature extraction model based on convolutional neural network was established,and the extracted gait features and view features are fused to form the final discriminant feature.The model training is based on the Siamese network,and the network is trained through Contrastive loss and Margin loss.The experimental results on CASIA-B show that the use of capsule network to extract view features has obvious advantages.Compared with the traditional gait recognition model,the accuracy of the cross-view gait recognition model fused with view features has been significantly improved,and the visualization shows that the cross-view gait recognition model fused with view features has better performance in the case of a larger view.3.Studying on cross-view gait recognition based on capsule network.Viewpoints is the most important factor affecting the accuracy of cross-view gait recognition,and the capsule can well represent the posture,spatial position and other information between entities.Therefore,in order to make full use of the characteristics of the capsule network,a cross-view gait recognition model based on the capsule network is proposed.The model uses a triple network architecture,and uses the Triplet loss and Margin loss to train the network.Experimental result on CASIAB dataset show that under the same experimental conditions,the cross-view gait recognition model based on the capsule network has an accuracy rate of 4.13% higher than that of the traditional gait recognition model,shows the effectiveness of using capsule network to extract gait features.