Research On Surveillance Face Image Super-resolution Method Based On Deep Learning

In video surveillance,the distance between the target person and the camera is far and the device itself has certain limitations,so that the face image obtained from the video has low resolution and is difficult to recognize.Therefore,it is an urgent problem to improve the resolution of the obtained face image.The superresolution algorithm based on deep learning has a certain effect on the reconstruction of the image,but the effect of the existing method is not obvious in the case of the face image under surveillance.This paper carries out high-quality reconstruction for the method dedicated to low-quality face design.We use a priori information combined with an improved loss function and an improved activation function for face superresolution reconstruction.The main contents are as follows:(1)Data enhancement through a simulated monitoring environment.Surveillance video is limited by many objective factors,such as the camera lens is not focused,the tape is used and copied repeatedly.These factors make the video blurry.In response to this phenomenon,we have enhanced the data in a simulated monitoring environment.(2)Weaken the impact of background information.The difference between different faces is small,the structure of the face is similar,and even the structures of the eyes,nose,and mouth are similar in shape.These features are advantageous for face positioning.This paper uses this feature to focus on the facial features of the facial features,and to deepen the influence of background information on the model through deep neural networks.(3)Weaken the impact of non-primary brightness information.Actual lighting conditions are usually not uniform.The side light causes too dark,the polarized light is too bright,and the high light causes too dark shadow.These will reduce the accuracy of face super-resolution reconstruction.In this paper,by designing the activation function,filtering part of the illumination,reducing the difference in brightness of each face image,and weakening the influence of non-primary brightness information on the model through deep neural network.