基于多级特征编码与双分支引导重构的图像去雾网络

打开文本图片集

关键词：计算机视觉；图像去雾；多级特征融合；双分支结构；注意力机制 中图分类号：TP391.4文献标识码：Adoi：10.37188/CJLCD.2025-0163 CSTR：32172.14.CJLCD.2025-0163

Hierarchical encoding and dual-branch guided reconstruction network for image dehazing

ZHANG Xinqi， TIAN Ying*，LI Yirong

（School of Computer Science and Softrware Engineering， University ofScience and Technology Liaoning， Anshan，China）

Abstract：To address the limitations of existing dehazing methods in detail restoration and global modeling，this paper proposes HEDGR-Net，a deep dehazing network based on hierarchical encoding and dual-branch guided reconstruction. HEDGR-Net adopts structure-enhanced convolution and non-local attention in the encoder to capture multi-scale structural and contextual features.In the decoder，a dualbranch design with a global context-guided module integrates local detail recovery and global semantic consistency. At the output stage，an attention-enhanced residual fusion module combines reconstructed and input features to refine image quality. A composite loss combining L2，SSIM，and TV is further employed to balance pixel fidelity，structural similarity，and smoothness.On the synthetic SOTS-outdoor dataset，HEDGR-Net achieves 27.54 dB PSNR and 0.957 O SSIM. On the real RTTS dataset，it obtains 26.99 BRISQUE and 8.77 Entropy. Compared with AOD-Net，HEDGR-Net improves PSNR by 21.7% ， SSIM by 5.0% ，reduces BRISQUE by 19.2% ，and increases Entropy by 3.9% ，showing clear advantages across multiple metrics. The proposed method enhances detail restoration and global brightness consistency，efectively overcoming the limitations of traditional dehazing methods such as incomplete haze removal and color distortion.

Key words：computer vision；image dehazing；multi-level feature fusion；dual-branch structure；attntion mechanisms

1引言

在计算机视觉技术快速发展的背景下，雾和霾等大气条件在图像采集与传输过程中常导致视觉信息退化，给视觉系统带来挑战[1]，尤其是安全监控、交通管理以及自动驾驶等领域。（剩余15443字）