Adaptive Dynamic Filtering Network for Image Denoising

In image denoising networks, feature scaling is widely used to enlarge the receptive field size and reduce computational costs. This practice, however, also leads loss of high-frequency information fails consider within-scale characteristics. Recently, dynamic convolution has exhibited powerful capabilities in processing (e.g., edges, corners, textures), but previous works lack sufficient spatial contextual filter generation. To alleviate these issues, we propose employ improve learning multi-scale features. Specifically, design a spatially enhanced kernel generation (SEKG) module convolution, enabling context with very low complexity. Based on SEKG module, block (DCB) (MDCB). The former enhances via preserves low-frequency skip connections. latter utilizes shared adaptive kernels idea dilated achieve efficient extraction. proposed multi-dimension integration (MFI) mechanism further fuses features, providing precise contextually enriched representations. Finally, build an network DCB MDCB, named ADFNet. It achieves better performance complexity real-world synthetic Gaussian noisy datasets. source code available at https://github.com/it-hao/ADFNet.