Research On Face Detection And Recognition Based On Lightweight Neural Network And Low Quality Images

Due to the progress of information technology and the arrival of the 5G era,face detection and recognition algorithms based on deep learning emerge in endlessly,and have been successfully applied in fields such as autonomous driving,smart cities,and mobile payment.Although the development of deep learning has promoted the development of the field of face detection and recognition,in practical applications,there are situations where the quality of face images is not high,such as low illumination,occlusion,different scales,and incomplete face display.Therefore,low quality face recognition algorithms have important research value and application space.In recent years,with the increasing scale and complexity of neural networks,the training of deep learning models requires increasingly high hardware resources.Typically,network models can only be successfully run on highly computational servers.However,for devices with limited hardware resources such as mobile devices,it is very difficult to run complex deep learning network models,which has become a bottleneck in current deep learning applications.This article focuses on the analysis and research of algorithms for face detection and recognition based on lightweight networks and low quality images.The content is as follows:First,in order to deploy the model on mobile devices with limited hardware resources and computational power,the idea of deep separable convolution and the multi-scale feature of human faces are utilized to add branches in the backbone network,thereby improving information loss between convolutional layers and the loss of multi-scale information in features.In addition,the classical regression loss function is compared and the GIo U Loss is selected to replace the Smooth L1 Loss in the original loss function according to the experimental results,so that the face detection network can obtain a better detection effect.Secondly,in order to further improve the performance of Mobile Face Net,based on the idea of attention mechanism,an efficient channel attention mechanism ECA is introduced into the backbone network,which only adds a small number of parameters to achieve cross channel interaction of features.By exploring the impact of different training loss function on the model,Ada Cos is finally used as the training loss function to avoid manual parameter adjustment to find the parameter value of the best effect.Finally,in order to extract a lower quality facial feature map with more representational ability,Swin Transformer is used as the network backbone,and a multi Receptive field fusion module is designed.In this module,the residual connection method is used,and the expansion convolution is fused at the same time.The input feature map is fused with different Receptive field.In addition,Ada Face loss function is used to train the network.The loss function considers the impact of sample image quality on the network,adjusts the weight of sample images with different quality,and further improves the robustness of the network.