CFNIF-DehazeNet: cross-scale feedback neural implicit function network for robust single image dehazing

Single image dehazing remains a challenging ill-posed problem due to complex atmospheric scattering, non-uniform haze distribution, and severe loss of structural information under dense haze conditions. Although recent deep learning methods have achieved significant progress, most existing approaches still suffer from cross-scale feature inconsistency, loss of fine textures, and reconstruction artifacts caused by discrete pixel-grid decoding. To address these limitations, this paper proposes CFNIF-DehazeNet, a progressive four-stage dehazing framework that integrates a Cross-Scale Feedback Unit (CFU) and a Neural Implicit Function (NIF) module. The proposed CFU establishes bidirectional interaction between adjacent hierarchical stages through discrepancy-aware feature refinement, enabling effective coarse-to-fine semantic propagation and fine-to-coarse structural feedback correction for improved cross-scale consistency. Furthermore, the NIF module replaces the conventional decoder with a continuous coordinate-based representation, allowing resolution-agnostic reconstruction while reducing over-smoothing and grid-induced artifacts. The network is trained using a combination of pixel-level reconstruction, structural similarity, and perceptual losses to preserve both global visibility and local texture fidelity. Extensive experiments conducted on RESIDE-6K, I-HAZE, O-HAZE, and Dense-Haze benchmarks demonstrate that the proposed CFNIF-DehazeNet consistently achieves superior quantitative and qualitative performance compared with existing state-of-the-art dehazing methods. The experimental results demonstrate the effectiveness of the proposed bidirectional cross-scale refinement and continuous implicit reconstruction strategy for robust single image dehazing.