Research On Methods Of Repairing Moving Targets By Integrating Saliency And Depth Information

In recent years,with the rapid development of computer vision technology,people have higher requirements for the quality of videos and images,these images with occlusion and damage can not meet the needs of people.Image inpainting is to restore the demaged image area based on the existing image information,the key point is to use the existing information of the image to fill the pixels in the damaged area,and restore the integrity of image structure and content authenticity.Convolutional Neural Networks(CNN)has powerful image feature extraction and representation capabilities,and initially shows good recovery effects in the field of image restoration.Therefore,how to design excellent network structure and loss function for different image restoration tasks has been a research hotspot for scholars.To solve the problem of moving target restoration,a moving target restoration algorithm which combines saliency detection and depth information is proposed.The main research work of this thesis includes:1.In order to solve the problem of missing information and too many parameters in traditional classification network structure.The generative adversarial networks(GAN)based on deep learning is used,and improves the generator structure and adds loss function based on structural similarity in this thesis,which makes the network model train faster,and the dege details of restored images are more abundant,the overall image are more authentic,and has better performance.2.In order to solve the problem of the position of the location of the occlusion cannot be determined for the video sequence.Through combining the video saliency detection algorithm and the monocular visual depth estimation algorithm to analyze the occlusion of the moving target in the video sequence in this thesis,which can estimate the mask area of the occlusion,and combine the image restoration method to remove the occlusion,so as to achieve the effect of moving target restoration.3.In order to solve the problem of the image is blurred and unreal after restoration.Based on the structure of the discriminator network,the global discriminator loss function and local discriminator loss function is proposed in this thesis,which improves the convolutioal mode to increase the receptive field of the network,and the integrity and content authenticity of the repaired image structure,and is beneficial to retain the local details of the image.Based on the above research,a complete solution for the problem of occlusion of moving targets is proposed in this thesis.Through a large number of experimental comparisons,the effectiveness and feasibility of the proposed algorithm are verified.