Blind Repair Of Face Occlusion Based On Deep Learning

Occlusion is a common phenomenon in real life.The occluded objects will lose some information,which seriously affects the computer's understanding and analysis of images.Therefore,occlusion has been hindering the development of some important research on computer vision.The occluded image is not convenient for our analysis and processing of the image,but also affects our research on image target detection and recognition.Image inpainting refers to restoring the information of defect position in the image.It is mainly through the characteristics of the existing information in the image to restore the missing part of the image.As an important branch of image inpainting,face inpainting has important application value in real life.However,the existing image inpainting methods have some limitations.They need to input the specific information of the defect location while inputting the defect image,which is difficult to apply to the actual scene.In real life,such as security video monitoring system,we usually capture the video image face will have partial occlusion,to a certain extent,the face key point information is lost,which has a great impact on recognition.Therefore,it is necessary to design a method to repair the occluded face.In recent years,with the development of deep neural network,especially the generation of confrontation network,many excellent image inpainting methods have emerged.The images restored by these methods are not only reasonable in semantics,but also true in content,reaching the point where it is difficult to distinguish the true from the false,and the effect of face repair is also improved.However,for the occluded face,due to the random pixels in the occluded area,which seriously affects the feature flow in the repair process,it is difficult to accurately repair the occluded area,and the final repair result is difficult to achieve the ideal value in detail.At present,there is no molding method for face occlusion compensation.Therefore,this paper studies this issue and constructs a face occlusion repair model based on deep learning.Due to the randomness of occlusion,it is difficult to repair the occluded parts directly.Therefore,a model with occlusion segmentation and occlusion repair function is designed to solve this problem.In this paper,a multi-scale feature-based codec network is proposed,which can segment the occluded part of the face image accurately and repair the occluded area.The whole network structure is divided into two stages.In the first stage,the occlusion region is segmented and the occlusion is rough repaired.In the second stage,the repair results of the first stage are refined.In the first stage,according to the characteristics of gated convolution,we fuse the missing features of occlusion repair encoder into the occlusion segmentation encoder,so as to better predict the occlusion part.In addition,we construct a feature pyramid network to fuse the upper and lower coding features in the repair encoder,and fuse the coding pyramid features of all levels in the decoding process,so as to guide the decoder to recover the pure face image better.The second stage is to fine tune the results of the first stage.In order to get a more realistic and clear image,we add two layers of hole convolution after decoding to expand the receptive field.Experiments on synthetic face data sets show that the proposed method has a good effect in blind face occlusion repair.