Loading…

Multi-scale dehazing network via high-frequency feature fusion

Numerous learning-based methods have achieved significant improvements in haze removal. However, the dehazed results of these methods still suffer from the loss of edge details. To solve this problem, we propose a novel multi-scale dehazing network via high-frequency feature fusion (HFMDN). HFMDN is...

Full description

Saved in:
Bibliographic Details
Published in:Computers & graphics 2022-10, Vol.107, p.50-59
Main Authors: Xu, YuJie, Zhang, YongJun, Li, Zhi, Cui, ZhongWei, Yang, YiTong
Format: Article
Language:English
Subjects:
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Numerous learning-based methods have achieved significant improvements in haze removal. However, the dehazed results of these methods still suffer from the loss of edge details. To solve this problem, we propose a novel multi-scale dehazing network via high-frequency feature fusion (HFMDN). HFMDN is an end-to-end trainable network, which is mainly composed of four components: a base network (Backbone), a frequency branch network (FBN), a frequency attention module (FAM), and a refine block (RB). The Backbone is a multi-scale feature fusion architecture that can share useful information across different scales. For the training phase, we employ the Laplace Operator to obtain the image’s high-frequency (HF) information, which can specifically represent the details of the image (e.g., edges, textures). The FBN takes the HF derived from the original image as an additional prior and utilizes L1 norm loss to constrain the output of FBN to predict the HF of the haze-free image. We further design a frequency attention module (FAM), which automatically learns the weights map of the frequency features to enhance image recovery ability. Furthermore, a refine block (RB) is proposed to extract the features map by fusing the outputs of FBN and Backbone to produce the final haze-free image. The quantitative comparison of the ablation study shows that high-frequency information significantly improves dehazing performance. Extensive experiments also demonstrate that our proposed methods can generate more natural and realistic haze-free images, especially in the contours and details of hazy images. HFMDN performs favorably against the CNN-based state-of-the-art dehazing methods in terms of PSNR, SSIM, and visual effect. [Display omitted] •An novel multi-scale dehazing network based on high-frequency feature fusion.•The frequency branch is introduced to reconstruct the high-frequency of the image.•The new attention module automatically generates the weights map of high-frequency.•The refine block is exploited to fuse the high-frequency and texture features
ISSN:0097-8493
1873-7684
DOI:10.1016/j.cag.2022.07.001