Efficient denoising method for real-world noise image using Scalable Convolution and Channel Interaction Attention

We propose a three-stage image denoising method, called Scalable Convolution and Channel Interaction Attention (SC–CIA), to address the high computational cost, complexity, and suboptimal performance of traditional convolution image denoising networks when dealing with real-world noise. In the first stage, we use a variant of dynamic convolution called Scalable Convolution for shallow feature extraction. This method utilizes a set of adaptive small convolution kernels with spatial variations and applies them to full-resolution feature mapping through slicing operations. By combining bilinear interpolation and Scalable Convolution operations, it minimizes computational resources while enhancing the network’s ability to capture position and shape information from images. In addition, our method incorporates an attention mechanism for channel interaction. This mechanism groups every two channels of the input feature map, generating attention maps for each subgroup. The output features are then aggregated and rearranged to achieve channel information interaction and feature enhancement, thereby improving the model’s ability to remove real noise. Compared to various mainstream denoising networks, our method achieves excellent PSNR/SSIM performance on SIDD and DND datasets, while significantly reducing computational complexity (measured in MACs). In particular, when processing 512 \(\times\) $\times$ 512 images, our method only uses 10% of the MACs used by MIRNet, and has an inference speed that is 3.73 times faster than MIRNet. These results highlight the potential of our method for fast denoising applications.