QCGFNet: A Lightweight Color Image Denoising Network

Most current denoising methods rely on large-scale models, such as transformers and diffusion models. However, these methods suffer from drawbacks such as excessive network layers and parameters, prolonged training time, and slow denoising processes, which can result in significant costs when applied to real-world scenarios. This work proposes a lightweight color image denoising network, called the quaternion convolution guided filter denoising network (QCGFNet). First, guided filtering is used to separate the base and detail layers of the original image. Then, two quaternion convolutional subnetworks are separately trained for the base and detail layers. Finally, their results are combined to produce the final denoised image. The quaternion convolution module can better fuse information between channels, and the four channels are processed in parallel. Using the detail layer for training simplifies image information and accelerates network learning and convergence. Compared with other excellent methods discussed in this paper, our method reduces the parameters by up to 145 times and speeds up inference time by up to 212 times. It achieves excellent performance on several datasets, especially the Kodak-24 dataset, improving average PSNR by 1.4% to 13.4% and average SSIM by 0.7% to 22.4%. It can be said that QCGFNet achieves outstanding performance at a very low cost.