Dual degradation image inpainting method via adaptive feature fusion and U-net network

Abstract

Most existing image inpainting methods are designed to address a single specific task, such as super-resolution, denoising, or colorization, with few models capable of handling dual degradation simultaneously. Moreover, current algorithms that tackle multiple image degradation problems often suffer from complex structures, prolonged training times, and high labor costs. In this paper, we propose a Dual Degradation Network via Adaptive Feature Fusion and U-Net (AFFU). The network employs a Self-Guided Module (SGM) to fuse multi-scale image information, effectively eliminating certain defects in the image. A coder-decoder module with null convolution is utilized to consolidate the semantic information of the image, enabling intermediate image colorization. Additionally, an Adaptive Multi-feature Fusion Module (AMF) and Information Transfer Mechanism (ITM) are introduced to link these two major structures, adaptively selecting and retaining image features during network progression to prevent the loss of useful information. Experimental results demonstrate that the proposed dual image degradation restoration network model, based on adaptive multi-feature fusion, achieves optimal visual generation. Evaluations on CelebA dataset and Landscape dataset show that the proposed method outperforms comparable approaches in terms of Structural Similarity (SSIM) and Learned Perceptual Image Patch Similarity (LPIPS).