Image restoration based on SimAM attention mechanism and constraint adversarial network

Image restoration has important application value in the fields of old photo recovery, object removal, video editing, etc. Traditional image restoration methods The existing image restoration methods use the single resolution image as the network input for image feature extraction, and cannot fully use of the different features of images with different resolution rates for image restoration, resulting in poor restoration of image texture details and even artifacts. In order to improve the restoration effect, an image restoration algorithm based on the SimAM attention mechanism and constrained adversarial network is proposed. Firstly, a low-resolution image generation module is designed, which uses the encoding and decoding structure to generate the global structure information of multiple low-resolution input images. Then, a SimAM attention feature fusion module is constructed, and the low-resolution image generation results are fused with the fine-grained features contained in the high-resolution input missing images to help complete the high-resolution images. A parameter-free SimAM attention mechanism is introduced to infer the three-dimensional attention weight of feature maps by considering the correlation between spatial and channel dimensions, to characterize locally significant features and suppress useless features, and to improve the effectiveness of target region information. Finally, an improved multi-scale adversarial loss function is proposed. The function guarantees the quality and authenticity of the discriminator image at different scales by constraining the local fineness of the generated image and the consistency of the repaired image on the high-level structure. The experimental results show that on the Paris StreetView, CelebA and Places2 data sets, aiming at three different image damage rates, the experiment is conducted based on three quantitative evaluation indexes including SSIM, PSNR and L1. The results show that the proposed method can generate more reliable local details and improve the effect of image restoration.