Research On Video Face Recognition Technology Based On Feature Fusion Network

Video processing has always been an active research direction in the field of computer vision.Among them,video-based face recognition has also attracted extensive attention in this field.The video has one more time dimension than the image,and has the following disadvantages for face recognition: blur,occlusion,light changes,and even the influence of the distance of the lens,which leads to the low quality of individual video frames.Moreover,due to the long face video,the facial pose and face size of the first few frames and the next few frames change greatly,which often generates noise that is not conducive to feature extraction and fusion.How to extract feature representations from videos that are useful for prediction is particularly important.When the length of the video sequence is short,the design of the multi-scale feature extraction network can efficiently extract the local detail features of each frame.The frame structure-aware aggregation network proposed in this paper performs feature fusion on this feature,and finally obtains the face video.The overall features of the sequence are represented and identified.For too long video sequences,the method in this paper segments the video sequences,and combines the proposed multi-head convolutional attention module to extract and fuse the features of long videos and remove redundant noise.The design of segmentation loss can speed up the training speed of the network and improve the anti-interference ability of the model.This paper studies the video face recognition technology combining frame structure processing,multi-scale feature extraction and attention mechanism.The main research work is as follows:(1)For face video sequences with short sequences and poor frame quality,a multi-scale feature extraction network and a frame structure-aware aggregation module are proposed to construct an overall video frame feature representation and identify it.First,the feature representation of the video frame is extracted by the multi-scale feature extraction network,and then the feature fusion network is trained and the corresponding weight of the feature representation of each video frame is assigned to achieve the purpose of evaluating the importance of the frame.Contextual information for efficient modeling.Compared with the traditional method of selecting key frames for recognition,the method in this paper can more efficiently utilize the features of each video frame and its spatial structure information;experiments on two public video face recognition datasets IJB-A and YTF The results show that this scheme has 0.25% and 0.6% improvement in the effect of video face recognition compared with the optimal comparison model.(2)Aiming at the face video sequences with too long and many low-quality frames,a video face recognition method based on segmentation strategy and multi-head convolutional attention is proposed.First,the segmentation strategy is used to map the video,and feature and position embedding are used to avoid the influence of facial changes caused by long sequence spans on feature extraction.The proposed segmentation video face encoder can effectively extract the context information of video frames and perform feature fusion.Finally,a segmentation loss function is constructed to reduce the recognition accuracy of low-quality frames to the model.,improving the model’s ability to handle lengthy video sequences.The experimental results on two public video face recognition datasets IJB-A and YTF show that the recognition accuracy of this scheme is0.47%-2.87% and 1.26%-1.42% higher than the current mainstream video face recognition models,respectively.