Learning Implicit Neural Degradation Representation for Unpaired Image Dehazing

TL;DR

The paper addresses unpaired image dehazing in scenes with highly nonuniform haze by introducing NeDR-Dehaze, which combines a Kolmogorov-Arnold-based KAN-CID module with implicit neural degradation modeling. It advances the field with an Implicit Dense Residual Module (IDRM) and a Dense Residual Enhanced Module (DREM) to learn haze degradation as a continuous function and remove redundant high-frequency information, enabling end-to-end training without explicit physical priors. Experiments on RESIDE and real-world datasets using PSNR, SSIM, and LPIPS demonstrate competitive or superior performance, particularly in complex hazy environments, while maintaining efficiency. The approach offers a practical, unpaired dehazing solution with strong robustness and potential for real-world deployment, with code made publicly available.

Abstract

Image dehazing is an important task in the field of computer vision, aiming at restoring clear and detail-rich visual content from haze-affected images. However, when dealing with complex scenes, existing methods often struggle to strike a balance between fine-grained feature representation of inhomogeneous haze distribution and global consistency modeling. Furthermore, to better learn the common degenerate representation of haze in spatial variations, we propose an unsupervised dehaze method for implicit neural degradation representation. Firstly, inspired by the Kolmogorov-Arnold representation theorem, we propose a mechanism combining the channel-independent and channel-dependent mechanisms, which efficiently enhances the ability to learn from nonlinear dependencies. which in turn achieves good visual perception in complex scenes. Moreover, we design an implicit neural representation to model haze degradation as a continuous function to eliminate redundant information and the dependence on explicit feature extraction and physical models. To further learn the implicit representation of the haze features, we also designed a dense residual enhancement module from it to eliminate redundant information. This achieves high-quality image restoration. Experimental results show that our method achieves competitive dehaze performance on various public and real-world datasets. This project code will be available at https://github.com/Fan-pixel/NeDR-Dehaze.

Figures (12)

• Figure 1: A single image haze removal example. (a) The hazy image, (b) The dehazing result using the D4 method, (c) The dehazing result of our method, and (d) The ground-truth image.
• Figure 2: The architecture of the method consists of the dehazing network. (a) represents the channel-independent and channel-dependent mechanisms based on the KAN network that we designed, and (b) denotes a dense residual enhancer based on Implicit Neural Representation. The feature extraction section denotes the extracted non-sized feature map; where $A$, $\beta$, $t$, and $d$ are intermediate parameters of the atmospheric scattering model; the ChnMapper module converts feature maps with different numbers of channels into feature maps with a specific number of extended channels.
• Figure 3: Visual comparison between the input image and the IDRM result. It can be clearly observed that IDRM has a significant effect on reducing the high-intensity pixel values caused by haze. This method effectively mitigates the interference of haze in the image and reconstructs the underlying haze-free image, thereby significantly enhancing the clarity and detail representation of the image.
• Figure 4: The detail branch of the dehazing network. (a) Feature Fusion Module, (b) Dense Residual Enhanced Module, (c) Residual Convolutional Module, and (d) Residual Dense Block.
• Figure 5: Visual comparison of haze removal on samples from the SOTS-Outdoor datasets.