| As an image processing technique,blind image separation aims to recover the original image from the observed image that has been linearly mixed and noise-added.Traditional independent component analysis,singular value decomposition,etc.achieve blind separation by assuming mutual independence between the source images.In practice,however,the possible correlation between the source images makes traditional methods difficult.Generative adversarial network is a new type of deep learning method,which autonomously learns the features of mutual mapping between different domains,suitable for solving the problem of source image correlation,to achieve more accurate source image estimation.This thesis improves the basic generative adversarial network and proposes a one-to-many generative adversarial network based on channel attention guidance,and a generative adversarial network combining Transformer and UNet,which achieves better performance in the single-channel blind source image separation task,and the specific innovations are as follows.1.A channel attention-guided one-to-many generative adversarial network model is proposed to address the fact that the source and mix images are similar in structural distribution,and the blind separation process is susceptible to pixel-level noise and artifacts caused by the interference source information in the mixed images.A multi-channel generative adversarial structure with one generator paired with two discriminators is used,corresponding to the generation of source images in each domain,and a channel attention mechanism is added to each generation channel for guiding the feature extraction of different source information.Using both adversarial loss,source image L1 loss,and reconstruction loss of the mixed image in the loss function of the generator,the generator is constrained to map different source information to different source domains.2.To address the problem of losing the detailed information and texture features of the source image in the separated image when the generative adversarial network is used for image reconstruction employing deconvolution in the blind source image separation task,a generative adversarial network structure combining Transformer and UNet is proposed.The UNet network is first used for feature extraction and coding and decoding,based on which the Transformer module is used to achieve positional coding and information interaction of the input features,and finally,the separated source information is reconstructed by the generative adversarial network.After quantitative and qualitative analysis of the experimental results,it is demonstrated that the method can retain more detailed information and texture features while maintaining high separation accuracy. |
