The paper deals with continuous and compact mappings generated by the Fourier transform between distinguished Besov spaces B_(p)^(s)(R^(n))=B_(p,p)^(s)(R^(n)),1≤p≤∞,and between Sobolev spaces Hs p(R^(n)),1
Due to the scattering and absorption of light,underwater imaging yields suboptimal results characterized by low brightness,diminished image contrast,and the loss of details.Current learning-based underwater image enhancement(UIE)methods mainly focus on addressing these issues in the spatial domain,with limited attention to the Fourier frequency domain.We propose the Fourier transform guided dual-channel underwater image enhancement diffusion network(FDNet),aiming to fully leverage frequency domain information and the characteristics of a diffusion-based generation model.The amplitude-phase dualchannel not only effectively provides feedback on the energy distribution of the image but also incorporates crucial image structure and texture information.Capitalizing on these advantages,we first introduce a Fourier transform-based amplitude-phase dualchannel enhancement front-end network for UIE.This network significantly improves the quality of the network's image input,including brightness and contrast,by incorporating phase congruency edge enhancement and prior-driven amplitude mapping strategies.The preceding network mitigates the overall training challenges caused by poor input image quality.Simultaneously,the application of denoising networks based on lightweight transformers effectively improves the computational time required for each iteration during model training.The results indicate that FDNet outperforms many current learning-based UIE algorithms in restoring images across multiple real underwater image datasets,demonstrating more robust generalization capabilities in degraded underwater scenarios.The competitive performance advantage achieved in enhancing visual quality underscores the effectiveness of our method.
Zhen ZHUXiaobo LIQianwen MAJingsheng ZHAIHaofeng HU
