Year: 2026
Type: article
Abstract: The existing manual prior based image dehazing methods are prone to edge smoothing and halo artifacts in non-uniform foggy scenes, while deep learning based dehazing methods improve overall restoration performance, but still have shortcomings in preserving edge details and accurately estimating depth. Therefore, this article proposes an edge preserving dehazing algorithm based on multi-scale fusion and depth estimation. A U-shaped Transformer depth estimation subnetwork is constructed, utilizing a sliding-window multi-head self-attention mechanism to perform weighted encoding of spatial semantic information at different scales, thereby enhancing edge depth and obtaining high-resolution scene depth maps. A feature fusion module combining parallel multi-scale convolutions with dense residual connections is designed. By modeling the multi-branch collaboration of 1 × 1, 3 × 3, and 5 × 5 convolutional kernels and pooling branches, the module enhances the depth map’s ability to represent details in multi-scale edge structures, while introducing a channel-spatial dual attention mechanism to suppress interference from irrelevant features. By combining guided filtering with the à-trous wavelet transform, and using the original foggy image as a guide, we perform edge-preserving residual high-frequency compensation on the preliminarily restored image, thereby removing residual fog effects while preserving sharp edges. Experimental results show that the proposed algorithm achieves a depth estimation AbsRel error of < 0.12 in the far-field range on the KITTI depth completion benchmark, an edge gradient amplitude of 0.78, a perceptual similarity (LPIPS) of 0.11 on paired synthetic fog datasets, and a transmittance edge variance of 0.63 on real hazy images from RTTS. Ablation experiments validated the cumulative contributions of the multiscale fusion and edge-preservation modules to the sharpness of edges in the depth map and the abrupt changes in transmittance boundaries, demonstrating effective image restoration in complex foggy environments. Clinical trial number Not applicable.
Source: Discover Computing
Authors: Yue Qin, Chao Liu, Jingsong Zhang
Institution: Qiqihar University
Language: English
Topic: Image Enhancement Techniques
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $1,520
OpenAlex ID: W7236822789
