Research On Face Detection Algorithm Based On Super-resolution Reconstruction Technology

In recent years,face detection technology has been developed rapidly with the unremitting efforts of researchers,but there is still a big gap with the actual application demand.Under unconstrained conditions,due to the influence of the photographing posture,imaging distance,and lighting,the face images we obtain are usually blurred,which is very unfavorable for image analysis and feature extraction tasks,and also leads to the detection rate of the face greatly reduced.Therefore,how to solve the accuracy,stability and efficiency of face detection is an active research topic in the field of pattern recognition and computer vision.In order to improve the application of face detection in the actual scene,the research content of this dissertation is as follows:(1)In view of the limitations and disadvantages of existing models such as MTCNN in detecting low-resolution human face,this dissertation proposes to build a multi task depth network model based on auxiliary vision task.First of all,through the construction of stacked hourglass network model,the face features are successfully extracted from the face image,which makes up for the high-level semantic information that the shallow features do not have in the convolution neural network,and enhances the target texture information that the deep features are easy to lose.On this basis,the face detection model(TFPGAN)based on the fusion of priori information and Generative Adversarial Network is studied,and the image super-resolution reconstruction is regarded as the auxiliary visual task of face detection.A new generation countermeasure network model is proposed,which can effectively alleviate the feature rough problem of small targets,at the same time,it can improve the detection accuracy of small-scale face while restoring the image details,and it is proved to be effective by experiments with other existing models.(2)This dissertation studies how to improve ours model through self-attention mechanism.The attention mechanism can use the normalization function to calculate the probability distribution of the information about the face and background semantic segmentation in the deep features.Through different probability distribution,we can directly screen the location and scale of face in fine-grained shallow features.Otherwise,the interactive sharing structure composed of residual blocks and self-attention mechanism is used to make attention probability distribution fullyexpress the target semantic context information makes full use of the global feature information of the face image,which greatly improves the reconstruction effect and detection accuracy of the image,and further improves the robustness of ours model to different real scenes.