Research On Image Superpartition Reconstruction And Coloring Based On Generative Adversarial Network

In recent years,deep neural networks have improved the ability of feature extraction and data fitting by increasing the depth of the model.Compared with the shallow model,it shows great advantages in image processing,such as image classification,image recognition,image segmentation,etc.,and provides a new way for image super-resolution reconstruction and black and white image coloring.Generative adversarial networks use the generative model and the discriminative model to achieve the camouflage of the maximum similarity of the generated pictures,so that the generation and conversion of images to images is best to be almost indistinguishable.For image super-resolution reconstruction and black-and-white image coloring,how to improve the utilization of image features,reduce the computational complexity,and solve the problems of fewer images and taking full advantage of existing images has become a research hotspot.The problems of low reconstruction quality of existing models,high model sub-complexity,and poor rendering accuracy of existing black and white images have become urgent problems.Based on the above problems,this research is based on the optimization of the generative network structure,the parallelization of the network layer,and the ingenious combination with existing deep learning frameworks.The advantages of the proposed model are compared experimentally.The specific work is as follows:(1)A single image super-resolution reconstruction model Res?WGAN based on WGAN and Res Ne Xt is proposed.This model uses a Res Ne Xt network to construct a generator,which reduces the computational complexity of the model generator,which is only 1/8 of that of SRGAN.To solve the problem of SRGAN model instability,the discriminator uses a WGAN network.The network structure was pruned to remove the BN normalization operation.The experimental results show that the model proposed in this thesis is obtained in subjective and objective evaluations compared with the existing single-image super-resolution reconstruction model on four public data sets More superior performance.(2)In order to solve the problems of large amount of training data and time consuming of existing models,a single image super-resolution reconstruction model SR?Sin GAN based on Sin GAN was proposed.As far as the amount of training data is concerned,SR?Sin GAN can obtain super-resolution pictures by training only one picture.From the perspective of the reconstruction effect,the improved SR?Sin GAN model is more effective than the original Sin GAN network in ensuring the training data amount With more cases,reconstruction of image details can be better achieved.(3)In order to realize the problems of black and white image coloring and low feature utilization,a black and white image coloring model SA?Cycle GAN based on Cycle GAN is proposed.From the comparison of the presence or absence of attention,SA?Cycle GAN with added attention mechanism uses the self-attention mechanism to weight and process the features of the convolution.Compared with the model without added attention mechanism,the PSNR and SSIM values are improved.Secondly,it can be seen from the comparison that the choice of different normalization methods also affects the performance of the network.Finally,experimental comparison with some existing models proves that the SA?Cycle GAN model is as close to the original picture as possible when restoring the picture,achieving The grayscale image achieves the purpose of coloring.