A novel CNN architecture for image restoration with implicit frequency selection
Jiaxing Hu, Zhibo Wang · Connection Science · 2025
Image restoration aims to recover clear images from degraded ones, with deep neural networks becoming the dominant approach. While earlier methods focused on spatial-domain information, recent models have explored frequency-domain data to improve performance. However, explicit frequency-domain processing introduces significant computational overhead. To address this, we propose the Implicit Frequency Selective Image Restoration Network (IFSR-Net), which implicitly captures frequency information without explicit transformations, achieving high performance with reduced computational cost. Our analysis indicates that the main spectral variations between the clear and degraded images are centred on the high-frequency components in the feature maps; the convolution operator tends to amplify the amplitude and variance of these components. Building on this observation, we designed an Implicit Frequency Selection Module (IFSM) to enrich high-frequency components and an Implicit Frequency Selection Attention (IFSA) mechanism to emphasize and integrate beneficial frequency features. We integrated and optimized design elements from existing image restoration models to further refine the overall architecture of IFSR-Net. Extensive experiments across seven datasets and three tasks demonstrate the effectiveness of our approach. Ablation studies confirm the validity of our design choices, offering insights for future research in image restoration.