Single Image Dehazing Based on Haze Prior Residual Perception Learning

Image dehazing is a pixel-level reconstruction issue that encounters significant challenges. The model-driven dehazing algorithms rely on strong nonlinear terms to describe haze features through explicit mathematical expressions. However, parameter identification is complicated due to the variable nature of haze across diverse scenarios. Data-driven algorithms, while utilizing deep neural networks to learn optimal solutions from large datasets, suffer from limited interpretability and discard prior knowledge encapsulated in mathematical expressions. In this paper, we propose a prior residual perception framework that integrates prior knowledge into the deep learning dehazing network. Instead of treating haze as an inherent part of the image, we regard it as perturbation information and express it through prior explicit model. It can be learned more efficiently by isolating from the overall image features. Specifically, in the dehazing pathway, our method reconstructs clear images by eliminating haze features derived from the prior model. Besides of above, structural and textural details are more indispensable in image recovery, we combine difference convolution with vanilla convolution in the gated mechanism to emphasize critical details and overcome vanilla convolution's limitations in extracting gradient information. Based on the proposed prior residual perception framework, we design a classification discrepancy loss that maximizes the separation between haze and clear regions, ensuring accurate haze feature learning for image reconstruction. Extensive experiments across four benchmark datasets and comparing seven state-of-the-art dehazing methods, demonstrate that our methodology achieves superior performance in both objective metrics and subjective assessment.