Research On Image Restoration Based On Generative Adversarial Network

In today's Internet world,images can convey more visual and intuitive information than words and sounds.Image processing technology has also achieved practical applications in many fields,such as surveillance video,unmanned driving,medical diagnosis,etc.However,during the process of acquisition,transmission and editing of images as information carriers,some random noise may be generated due to the problem of collection equipment.When it is in some bad weather such as rainy days,the shooting equipment will inevitably be attached by raindrops and cause shooting The image was contaminated by raindrops.Such images not only affect the human subjective visual understanding,but also bring a certain negative impact on subsequent computer vision algorithm processing.Therefore,the image restoration technology has great practical application value.Based on generative adversarial networks,this paper conducts research on the noise in the image and the raindrop recovery technology in the image.The main work is as follows:(1)In order to solve the problem of insufficient expression ability caused by the loss of features in the training process of the generated adversarial network,a residual model feature extraction network combined with dilated convolution is proposed.The introduction of dilated convolution in the generated network can effectively increase the receptive field of the network and reduce the amount of calculation.At the same time,the residual module is added to the generated network to directly transfer the shallow features to the deep network,effectively avoiding the problem of gradient disappearance during the network training process and improving the quality of the restored image.And through experiments to verify the effectiveness and feasibility of the algorithm.(2)The presence of image noise will interfere with people's understanding of the image.The effect of common image denoising methods is often not ideal,it is easy to cause image texture to be blurred and the restoration area is too smooth.An improved generative adversarial network algorithm,which increases the width of the generative network to obtain more image features,and adds a global residual to extract and learn the features of the input noise image to avoid the loss of features.The network uses the weighted sum of the loss against reconstruction and the reconstruction loss,which can effectively retain the detailed information of the image while removing noise.Experimental results show that the algorithm can effectively remove image noise and improve the visual perception of the image.(3)Because raindrops do not have a fixed shape and have different transparency,the method of directly removing raindrops often causes some damage to the image background information while removing raindrops.Aiming at this problem,this paper proposes an algorithm for raindrop removal in image generative based on differential learning.The generative network does not directly output the image without raindrops,but learns the difference between the image with raindrops and the image without raindrops,and then subtracts the learned difference image with the image with raindrops to generate an image without raindrops.In order to better learn this difference,the reconstruction loss is introduced into the generative network,and the pre-trained VGG-16 network is used to extract the difference between the generated image and the real image and calculate the mean square error.The experimental results show that the method in this paper can not only remove the raindrops in the image,but also reconstruct the image information of the part blocked by the raindrops.