Face anti-spoofing recognition method based on action enhancement

Abstract

The invention discloses a face anti-spoofing recognition method based on action enhancement. It belongs to the technical field of digital image processing. The present invention enhances the action information input into the CNN+LSTM network video through the action enhancement technology. In addition, in order to overcome the defect of missing position information in the existing CNN+LSTM framework, the present invention adds the LSTM structure behind the last pooling layer, and Remove the fully connected layer to achieve the purpose of retaining position information, so that the extracted sequence features are more distinguishable; at the same time, the present invention also adds an attention mechanism to the improved framework, and by setting the position confidence matrix, the position changes significantly The increase of the regional confidence value makes LSTM more focused on the area where the action information is concentrated.

Description

technical field [0001] The invention belongs to the technical field of digital image processing, and in particular relates to a human face anti-spoofing recognition method based on motion enhancement. Background technique [0002] As a convenient and efficient biometric authentication, face recognition has attracted more and more attention. Although the face recognition system is developing rapidly, its security faces severe challenges. Face recognition systems are vulnerable to attacks, and using printed photos or rebroadcasted videos to deceive and attack the face system has become a common method for criminals. In general, a practical face recognition system not only needs a high degree of recognition performance, but also has the ability to distinguish whether the authentication information is from an attacker (fake face) or from a real person (real face). [0003] In the face anti-spoofing technology researched now, most of them are based on artificially designed feat...

AI Technical Summary

Problems solved by technology

Therefore, it can learn timing-related information well in the entire video sequence, memorize action enhancement cues, learn action cues by continuously comparing and learning input video frame sequence information, and judge the overall person of this video sequence through the learned action cues. The face is true or false, but there are still many places to be improved in this scheme

It does not do too much preprocessing and information mining on the input data. In addition, its LSTM structure is directly added to the output of the fully connected layer of the CNN network, thus losing the position information on the previous convolutional layer. Position information is very important to the sequence characteristics of video frames

Images