Generative Adversarial Network-Based Deep Image Enhancement Algorithms

With the development of 3D(three dimension)imaging technology,depth image plays an increasingly important role.At the same time,the birth of low-cost 3D scanning devices such as Microsoft Kinect and time of flight(TOF)cameras has opened the door for new applications in different research fields,such as computer vision,graphics,human-computer interaction and virtual reality.However,the depth image captured by depth camera will produce various types of distortion,which makes it difficult to accurately estimate the depth information from the depth image and affects people’s experience.In order to solve the problem of depth image quality degradation,appropriate depth image enhancement methods can be used to improve the quality of 3D images and the comfort of people watching stereo images.The purpose of depth image enhancement is to repair damaged isolated pixels and small areas,and improve image details,especially the edge of depth image.Aiming at various types of distortion in low-quality depth image,this paper proposes a depth image enhancement algorithm based on generation countermeasure network.The main contents of this paper are as follows:1.A structure of generative adversarial networks based on deep convolution is proposed.Aiming at the problems of low distortion image and traditional generative adversarial networks,such as gradient explosion,gradient disappearance and model degradation,the above network structure is improved,and a model based on residual network is proposed.Residual network and jump connection are introduced.The experimental results show that the peak signal-to-noise ratio(PSNR)of the generated depth image is increased by 6%,and the structural similarity(SSIM)is increased by 4% after introducing the residual network,which can restore the texture structure information of the image better.2.Aiming at the problems of traditional generative countermeasure network,such as difficult training and convergence,a discriminant network model based on Wasserstein distance is proposed.Wasserstein distance is used as an improved loss function at the end of the discriminant network.The experimental results show that after introducing Wasserstein distance,the peak signal-to-noise ratio and structure similarity of the enhanced depth image are improved by 11.7% and 5.1% respectively.3.Aiming at the problems of chessboard effect and image detail loss in generating depth image,an improved loss function for generating network model is proposed.The loss function includes mean square error loss,self smoothing loss and gradient loss.The mean square error loss aims to improve the peak signal-to-noise ratio of the image.The self smoothing loss can remove the chessboard effect in the image,and the gradient loss can enhance the details of the image.Through the improvement of network loss,the depth image quality is improved,and the edge information of the image is retained effectively.Experimental results show that the proposed model can improve the PSNR and structure similarity of the enhanced image by17% and 8.2% respectively.