Gans-based Research On Motion Deblurring And Color Enhancement

Image is a kind of information carrier with great use value.Not only that,people with certain aesthetic ability can find beauty from the composition and color of images.From the perspective of social development,image plays an important role.There are two classic problems in the field of image processing,motion deblurring and image color enhancement.With the development of deep learning,the research on these two problems has made remarkable progress for the help of neural network model.In the past,two different models were used to deal with them.This is bound to increase the total processing time.The research of this thesis aims to enhance the two effects at the same time without reducing the two effects.The main research work of this thesis is as follows:1.Based on the convolutional network structure of encoder-decoder,an image generator which can restore the blurred image to clear and beautiful color is constructed.In the generator,the residual connection is used to improve the sensitivity of the model to capture image features.In the middle of the generator,the nested residual connection is applied,which is conducive to model optimization and can give model the ability to extract complex features.The skip concatenate operation is applied between encoder and decoder to avoid the loss of image feature information after multi-layer convolution operation.2.Based on Wasserstein distance,a generative adversarial network is constructed for this research task.Due to the binary loss of the discriminator of the original generative adversarial network,the model can not achieve the gradual approach between the generated image and the target image.Therefore,in the case of using L1 loss as content loss,and based on WGAN,a network model which can enhance both effects is constructed.Compared with the single task model,this model greatly reduces the image processing time.3.Inspired by gradient penalty and image style transfer,the perceptual style loss which can measure the subtle style difference of images is added to the generator loss,and the WGAN based model is improved with gradient penalty.According to the analysis of image feature information extracted from different convolutions of VGG16 model,after a series of experiments about adding different proportion and different features of perception style loss,we successfully get the generator loss setting mode to improve the effect of the model.