Deep Models For Low Resolution Face Detection And Recognition In Video Surveillance

The great breakthrough of face recognition technology is driven by fast iterative computer vision.Although current face recognition system has achieved high recognition accuracy on controlled high-resolution faces,while in real scenes such as video surveillance environment,due to influence of distance,posture,illumination,occlusion,expression,noise and other factors,Face images captured by surveillance cameras are often blurred,which brings great challenges to face detection and face recognition.Therefore,it is of great significance to improve the accuracy of low-resolution face detection and recognition in real scenes,whether in academic research or practical application fields.For low-resolution face images in video surveillance scenarios,this paper proposes an effective solution through literature research and experimental exploration to improve the overall performance of face detection and recognition in video surveillance.The main work of the article is as follows:1.In view of the low-resolution face recognition in video surveillance scenes,based on the summary of the research status of face detection and recognition algorithms at home and abroad,combined with the particularity of video surveillance scenes,the research difficulties of low-resolution face detection and recognition algorithms in video surveillance scenes are pointed out,and then the ideas of low-resolution face detection and recognition algorithms in this paper under the framework of deep learning are introduced.2.It is known that current state-of-the-art face detection algorithms are not suitable for terminal equipment such as video surveillance due to large model memory and slow reasoning,this paper designs a light and low-resolution face detection model DLFace,which takes into account the model memory size while taking into account multi-scale faces Detection.By improving the depthwise separable convolution,the feature loss can be effectively reduced.By introducing the improved deformable convolution(DCNv2)and Lambda Layer,the feature extraction is refined and the context information is effectively enhanced.Compared with current face detection algorithms with superior performance,DLFace achieves a balance of performance and speed.It has verified the superiority of DLFace in different scenarios,indicating that DLFace can be better suited for low-resolution faces in video surveillance scenarios.3.It is found that insufficient of current face recognition algorithms for real low-resolution faces,this paper designs a low-resolution face recognition model KDDA based on knowledge distillation and domain adaptation.The Res2 Net module is introduced into the backbone network to express facial features in a more fine-grained way.Taking into account the bottleneck problem of the teacher network for student network guidance in the process of knowledge distillation,the mean square error loss is applied to each stage of the backbone network to strengthen the ability of student network feature extraction.Considering that artificial down sampling can not fit the degradation mechanism of low-resolution face in real scene,knowledge distillation and domain adaptation are effectively fused to learn the feature representation with domain invariance,so as to improve the recognition accuracy of the model on low-resolution face in real video surveillance.The results show that KDDA is suitable for low-resolution face recognition in video surveillance scene.