.png){kind=icon}
.png){kind=icon}
Return
Multi-aggregation network based on non-separable lifting wavelet for single image deraining
.png){kind=icon}
PRE
AI
DOI：10.1007/s00530-023-01156-0.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Recently, many methods have utilized Haar wavelet to extract frequency-domain information for image deraining. However, Haar wavelet and other tensor product wavelets only capture high frequencies in horizontal, vertical, and diagonal directions, leading to the loss of high-frequency details in deraining methods. To address this issue, we propose a multi aggregation network (MAGNet) based on non-separable lifting wavelet transform (NLWT), where NLWT is employed to capture high-frequency rain streaks in various directions. MAGNet aggregates neighboring features and incorporates outer skip connections based on the U-Net architecture, effectively utilizing the complementary information of rain patterns in different features. Additionally, a gated fusion module is used to fuse the aggregated features, capturing important rain pattern information and reducing feature redundancy. Moreover, MAGNet employs a scale-guide progressive fusion module to exploit the similarity between rain patterns at adjacent scales for deraining. Experiments on rainy datasets and a joint rain removal and object detection task demonstrate that our MAGNet outperforms advanced methods. The code is available at https://github.com/fashyon/MAGNet.
Keywords:
Image deraining
Non-separable lifting wavelet
Multi-aggregation
Scale-guide
Gated fusion