Research On Identity Recognition Method Based On Signature Hand

With the development of information technology and the change of people's consumption concept,bank card payment has become one of the most commonly used means of transaction.At present,in the process of credit card payment,the user's identity is mainly confirmed by password,but the password is not easy to remember and is easy to be obtained by lawless elements.Just using password alone can not effectively prevent the occurrence of credit card fraud,Whether or not the password is used,the user needs to sign to confirm the transaction process,and the process of user signature contains many personal characteristics of the user.This paper takes the user signature video recorded by the camera as the research object,and uses the hand features and motion features in the signature video to identify the user's identity.The main contents of this paper are as follows:Firstly,hand region segmentation in signature video frame.In this paper,hand region extraction is regarded as an instance segmentation task.Since there may be multiple hands in some signature video frames,and the hand which write the name is helpful for identification,other hands will cause confusion in the recognition,so this paper uses instance segmentation task instead of semantic segmentation to complete the task.In order to ensure the speed of segmentation,Polar Mask is used as the basic network in this paper.The network mainly uses FPN to fuse strong high-level lowresolution semantic information feature maps and low-level high-resolution feature maps with weaker but rich special information.But when the backbone network is too deep,the path from the bottom layer to the top layer will be very long,which is not conducive to the use of the bottom layer features.In order to make better use of the location information existing in the bottom layer,this paper adds a bottom-up path based on FPN,which reduces convolution times that the bottom layer features propagate to the top layer features.Experiments show that this network achieves good results in hand region segmentation datasets and egohands dataset.Secondly,the signature video identity recognition algorithm based on object detection.To begin with,the hand region in the signature frame is detected by target detection,and the hand region is cut out according to the detection results.The cut hand image is stored according to the time sequence in the video.In addition,the image classification network and the video classification network are used to train the cut hand image.Experiments show that the image classification network has achieved better results than the video classification network on the test set.Lastly,the task uses the trained target detection network and classification network to construct an endto-end identity recognition framework is constructed to recognize the signer's identity in the signature video by inputting a video.At last,the signature video classification network based on temporal pyramid.In the second part,it is found that the motion information in the video may be lost in the clipped hand image,which may lead to the final classification result not optimal.Therefore,this part takes the complete video frame as the research object,and makes a comparative experiment on the clipped hand image data set and the complete signature video data set.The experiment proves that this conjecture is correct.In order to identify the complete signature video and ensure the speed of identification,ECO network is selected as the basic network.Since ECO does not consider the speed of the signature object in the video when extracting the motion features,this paper adds a time pyramid module on the basis of ECO network to extract the fast and slow features for identification,the recognition accuracy has been improved comparing to ECO network.