Deformable Attention Multiscale Feature Fusion Network‐Dehaze: An Adaptive Image Dehazing Method Based on Deformable Attention Multiscale Fusion Network

Image dehazing, a fundamental computer vision task, is crucial for the stability of multiple downstream systems. Traditional methods have limited generalization, while deep learning methods face bottlenecks like error accumulation and poor adaptability. To address these, this paper proposes deformable attention multiscale feature fusion network‐dehaze, a deformable attention‐based multiscale feature fusion network with three core modules: the revised residual shrinkage unit uses deformable convolution to dynamically adjust receptive fields via real‐time fog density learning for precise adaptation to irregular fog regions; the hierarchical multiscale attention integrates channel and spatial attention to screen key features and restore fog‐obscured details; and the cross‐scale feature fusion adopts bidirectional transmission to fuse high‐level semantic and low‐level spatial features, compensating for downsampling‐induced spatial loss. Experiments on RESIDE and D‐HAZE show the method outperforms mainstream approaches, with peak signal‐to‐noise ratio improved by over 20% and SSIM by over 10% in some real fog scenes, verifying its superior accuracy and generalization.