Pinwheel-Assisted Dual-Stream Image Restoration Enhancing Texture and Structure Harmony

This study proposes a novel dual-stream image inpainting network, which innovatively integrates pinwheel-shaped con- volutions. The architecture consists of two branches with encoder-decoder structures, focusing on texture and structural inpainting respectively. These two branches interact with each other, where the features extracted by the texture branch guide the inpainting process of the structural branch, and vice versa. The research further introduces Pinwheel-Shaped Convolution to enhance feature extraction and expand the receptive field. Experiments on the JIANDU dataset and the Paris Street View dataset demonstrate that the model has achieved significant results in terms of image inpainting quality, structural fidelity, and perceptual similarity. With a mask rate of 30%-40%, the model's PSNR reaches 26.4433 and the SSIM value is 0.8920, proving the superior performance of the model in image inpainting tasks.