Image denoising network based on involution kernel
Ji Xia, Zishuai Hu · 2024
This article presents CIUNet, an innovative deep neural network designed for image denoising, with the objective of efficiently eliminating noise and enhancing denoising task performance. By leveraging cutting-edge convolutiveinvolution modules, CIUNet adeptly captures non-local image information without the need for additional parameters, and is finely tuned based on the UNet architecture, facilitating a synergistic improvement of both local and non-local features. The network is engineered for end-to-end learning, ensuring minimal temporal and spatial complexity. Extensive testing on both synthetic and real datasets has underscored CIUNet's exceptional efficiency and effectiveness in image denoising, particularly notable in its handling of low-noise images. Additionally, the paper includes analyses of model complexity and ablation studies, establishing CIUNet's ability to deliver high denoising performance with reduced complexity, affirming its viability as a lightweight network solution. In sum, CIUNet's introduction marks a significant advancement in the image denoising domain, offering a novel approach and insights, alongside substantial practical application potential due to its innovative network design and denoising proficiency