Research On Multi-task Face Recognition System Based On Deep Learning

Face detection and recognition research has always been a popular research direction in the field of computer vision.Face detection and recognition refers to the process of detecting and locating static faces or dynamic faces using information technology and relying on face information to achieve identity recognition.However,in practical applications,traditional face detection methods and face recognition methods often fail to achieve fast and accurate results.For dynamic faces and multiple faces,it is difficult to capture faces quickly and stably using traditional methods.Moreover,due to the limitation of the acquisition device,the low-resolution image will also have a greater impact on the face recognition effect.Based on the above,this thesis implements a multi-task face recognition system based on deep learning algorithms for static face detection and recognition,dynamic face detection and recognition,and low-resolution face detection and recognition.The overall system can achieve face detection,face tracking,face recognition and image super-resolution reconstruction.In order to take into account the performance of the algorithm as well as its practicality,this thesis adopts the MTCNN(Multi-task Convolutional Neural Network)face detection algorithm and improves on it.This thesis achieves faster face detection under the premise of guaranteeing the detection accuracy.Firstly,using fixed-size convolution to replace the pooling layer to extract more informative features,and then using the depth separable method for the second improvement to significantly reduce the model computation amount.After testing,the improved model achieved a recall rate of 94.55% on the FDDB dataset,only a decrease of about 0.5%,but the average inference speed increased by about 74%.Secondly,in order to enhance the detection effect of multiple faces,the Soft-NMS algorithm is introduced to replace the NMS(Non-Maximum Supression)algorithm,and the model’s ability to detect faces and side faces in low light environments is improved by enhancing the data.In the process of dynamic face detection,if there are too many faces,faces are not correct,faces move or face interlacing occurs,dynamic faces cannot be captured stably by relying on the face detection algorithm only.Therefore,in order to achieve effective and stable capture of dynamic faces,this thesis builds a face tracking module using Kalman filter and Hungarian function.This module assists the face detection module to achieve multiple faces tracking at the same time,which improves the stability of dynamic face capture.The face recognition algorithm in this thesis adopts Mobile Face Net.In order to improve the accuracy of face recognition,this thesis introduces the channel attention module and HDC(Hybrid Dilated Convolution)structure to improve Mobile Face Net.It enhances the channel information attention by adding an attention module for adaptive feature optimisation and uses the HDC structure to improve the front-end part of the network.Compared with Mobile Face Net,the improved model increases the test accuracy by 0.24% on the LFW dataset.Aiming at the problem of poor face recognition due to low image resolution,this thesis research builds a super-resolution reconstruction module for image clarity processing.The super-resolution algorithm is improved based on ESPCN(Efficient Sub-pixel Convolutional Neural Network),which firstly uses tandem small convolution instead of large convolution,and then improves the traditional convolution into MLPConv structure to improve the nonlinear expression ability of the model,and uses Smooth L1 loss in training.Compared with ESPCN,the improved model improves the mean PSNR(Peak Signal to Noise Ratio)values by 0.35 d B,0.29 d B,and 0.28 d B on Set5,Set14,and BSDS200 datasets,respectively,and the mean SSIM(Structural Similarity Index Measure)values by 0.0023,0.0022,and 0.0019,respectively.In this thesis,the operation of the system is tested through a variety of scenarios such as static face recognition test,dynamic face recognition test,low-resolution face recognition test,and multiple face recognition test in a real environment.The tests are all based on Pytorch framework to achieve model inference.When the camera acquisition resolution is 640×480,the CPU-based running system detects dynamic single face and four faces in about 43 ms and 70 ms,respectively,and the time used for face recognition is about 36ms;and when testing the image superscoring module with50×70 images,the time used for processing images is about 14 ms.