中文 | English

J. Hefei Univ. Tech. (Nat. Sci.)

合肥工业大学学报(自然科学版)

JOURNAL OF HEFEI UNIVERSITY OF TECHNOLOGY (NATURAL SCIENCE)

Competent Authorities: Ministry of Education of the People's Republic of China

Sponsor by: Hefei University of Technology

ISSN 1000-1646, CN 34-1001/N

Multi-stream selective kernel dehazing network

Download Full Text

多流选择性核去雾网络

Journal Information

J. Hefei Univ. Tech. (Nat. Sci.), Mar.2025, Vol.48 No.3: 320-326

DOI: 10.3969/j.issn.1003-5060.2025.03.005

Authors

HOU Ce, YANG Yizhong, LIU Xueqing

(School of Microelectronics, Hefei University of Technology, Hefei 230601, China)

Abstract and Keywords

Abstract: Although current image dehazing algorithms have achieved high indicators, the restored images still have problems such as color distortion, haze residue, and image artifacts. Therefore, this paper proposes an end-to-end multi-stream selective kernel (MSSK) dehazing network, which can optimize the dehazing effect by using the low-frequency and high-frequency information of the image. The color fidelity of the image is improved under the guidance of low-frequency information, and the edges details of the image are optimized under the guidance of high-frequency information. The network is a multi-branch structure, and the information interaction between branches is realized by branch guidance blocks. It uses the selective residual block to adaptively adjust the size of the receptive field according to the haze concentration. A comprehensive evaluation was carried out on multiple benchmark datasets. The experimental results show that the proposed network has high structural similarity (SSIM) and peak signal-to-noise ratio (PSNR), which can solve the problems of color distortion, haze residue and image artifacts.

Keywords: deep learning; image dehazing; multi-stream structure; adaptive fusion; end-to-end training

作者信息

侯策，杨依忠，刘雪晴

（合肥工业大学微电子学院，安徽合肥230601）

摘要和关键词

摘要: 目前图像去雾算法虽已取得较高的指标,但其恢复出的图像仍存在着颜色失真、雾霾残留、图像伪影等问题。文章提出一种可端到端训练的多流选择性核(multi-stream selective kernel, MSSK)去雾网络,借助图像的低频信息和高频信息优化去雾效果,实现在低频信息的指导下提高图像的颜色保真度,在高频信息的指导下优化图像的边缘和细节。该网络为多分支结构,通过分支指导块实现分支间的信息交互;采用选择残差块依据雾霾质量浓度自适应地调整感受野大小。在多个基准数据集上对网络进行评估的实验结果表明,文章所提出的网络具有较高的结构相似度(structural similarity, SSIM)和峰值信噪比(peak signal-to-noise ratio, PSNR)指标,能较好地解决颜色失真、雾霾残留、图像伪影等问题。

关键词: 深度学习；图像去雾；多流结构；自适应融合；端到端训练

Funding

基金项目:安徽省自然科学基金资助项目(2208085MF177);中央高校基本科研业务费专项资金资助项目(JZ2021HGQA0262)

Previous Next

Back to Contents