Research On Face Super-resolution Algorithm Based On Video Stream

The performance improvement of computer algorithms can bring disruptive changes to various applications in daily life.In this paper,face super-resolution processing for video streams,the face image in the original low-resolution video is super-resolution processed,which can effectively improve the accuracy of face recognition while improving the clarity of the target face image.This technology can provide better application scenarios for the application of face recognition and provide good technical support for the field of public safety.This paper combines two key technologies of video super-resolution algorithm and face super-resolution algorithm,and implements the deployment of the algorithm on edge computing platform--FPGAs.This paper starts from the classic super-resolution algorithm based on sparse coding and dictionary learning,summarizes it into the algorithm framework of neural network,and finally uses neural network method to achieve video super-resolution task.In the process of video processing,this paper designs and proposes an efficient general video processing framework based on the general characteristics of simultaneous video processing algorithms.At the same time,face feature extraction is used in the implementation of video super-resolution algorithm,and the training strategy of neural network construction using face loss can improve the accuracy of face recognition in the process of improving the subjective perception of the image after face super-resolution.The video processing algorithm needs the information of the adjacent frames before and after.In this paper,combined with the general video processing framework,a real-time processing framework with a one-frame delay algorithm is planned and implemented.The experimental results show that the processing of low-resolution face images can achieve a structural similarity of 0.933 and a peak signal-to-noise ratio of 32.998 d B under the super-resolution algorithm in this paper,which is comparable to the 0.273 index of the face distance corresponding to the high-resolution image.Compared with the more state-of-the-art super-resolution algorithms such as FSRCNN,EDSR,and ESRGAN,our algorithm has obvious advantages,not only in low model parameters but also in outstanding evaluation indexs.The demonstration results of video super-resolution also show that when processing continuous video images,the algorithm in this paper is natural and fluent in processing the target face,and the recovered face detail information is correct and rich,which has a good practical scene.In addition,during the execution of video stream super-resolution algorithm,this paper uses an algorithm supplemented by face posterior information in order to solve the problem of the decline in accuracy of face recognition in situations such as the movement of the recognition object,the turning of the face,etc.,which can effectively improve the accuracy of recognition.In the samples tested in the experiment,the face video super-resolution algorithm supplemented with posterior information can stabilize the average distance in face recognition below 0.3.This article also implements the super-resolution algorithm on FPGA.The experimental results under the Vitis-AI framework show that the video super-resolution algorithm in this paper can reach 2.45 GB/s operating bandwidth under the general video processing framework,and a throughput rate of 3.87 TMACs Performance,the processing speed of163.3 frames can be reached for the input image of 56×56 size,which satisfies the real-time nature of video processing.