Research On Attention-guided High-resolution Reconstruction Method Of Defect Image

Complete high-resolution images have important applications in cultural relics protection,information dissemination,criminal investigation analysis,etc.However,due to the limitations of imaging cost,imaging environment,transmission bandwidth,and natural or human factors,it is easy to cause spatial information damage and detailed information loss.To solve the problem of spatial information damage,image inpainting technology can be used to fill the defect content.At the same time,aiming at the problem of detail information loss,the image details can be reconstructed by super-resolution technology.Generative adversarial network(GAN)is the classic strategy in the field of image inpainting and super-resolution.It shows good reconstruction performance through the zero-sum game and has attracted the attention of many researchers.This thesis introduced the attention mechanism into the generative adversarial framework and proposed effective image inpainting and super-resolution methods,aiming at the problem that the image inpainting network architecture is not fully utilized and the super-resolution network architecture does not pay enough attention to the key areas of the image processing.The main research contents and contributions are as follows:(i)Through a detailed comparison and analysis of the experimental results for the mainstream image inpainting and super-resolution methods,the following deficiencies have been found.The results of the mainstream image inpainting network architecture showed fuzzy and severe edge responses.The high-resolution image reconstructed by the typical image super-resolution method of generating adversarial has serious detail distortion and artifact.(ii)The mainstream codec network architecture losses the information in the encoding stage,it is difficult to reconstruct the realistic texture structure by using the small scale features with a serious lack of detailed information in the decoding stage,which leads to the problems of fuzzy results and serious edge response.Considering the rich hole edge information contained in the features of different scales in the encoding stage,this thesis proposed an image inpainting method based on a multi-stage decoding network.The proposed image inpainting method used the features decoded by multiple sub-decoders from the encoder at different scales and works with the main decoder to establish the final output.Experimental results showed that the proposed multi-stage decoding network image inpainting method can restore the damaged area more clearly and the edge transition is smoother,compared with the representative methods.(iii)Because the super-resolution network architecture based on generation confrontation does not pay enough attention to the key areas of the image,does not have enough constraints on the image generation process,and generates random error information,it leads to the problem of image detail distortion and serious visible artifacts.This thesis introduced attention mechanism into the high-resolution image reconstruction network and proposed a super-resolution method combining attention mechanisms,based on the fact that the attention mechanism allocates weights to different representations and strengthens the attention of important information.This method differentiates the textured region from the smoothed region by applying attention constraints so that the network can focus the reconstruction target on the region with rich high-frequency details and perform high-frequency compensation.Experimental results showed that the detail is more realistic and the overall image observation effect is more realistic of reconstructed images by the super-resolution method of the combining attention mechanism proposed in this thesis,compared with the representative method.