Image Restoration Based on Multi-Scale Feature Enhancement and Fusion

One of the most challenging issues in computer vision is image restoration with large-scale irregular missing patches. Considering existing image restoration methods do not adequately integrate structure and texture information, this causes the problem of structure distortion and texture blurring in the final generated image. This paper proposes a multi-scale feature enhancement and fusion image restoration network. Based on a dual-stream restoration network, a multi-scale augmented attention method is produced to capture multi-level semantic information in images and facilitates the fusion of structural and textural elements. In the feature fusion module, fast Fourier convolution is added to the processing of structure features to draw attention to edge features and enhance the contrast and clarity of the edges in the image. Numerous experiment results show that our method can effectively repair large-scale irregularly missing images in complex backgrounds, and metrics like peak signal-to-noise ratio and structural similarity perform better than existing comparison techniques.