首页|基于双阶段特征解耦网络的单幅图像去雨方法

基于双阶段特征解耦网络的单幅图像去雨方法

扫码查看
原文链接
  • NETL
  • NSTL
  • 万方数据
  • 维普
针对现有的单幅图像去雨方法无法有效地平衡背景图像细节恢复与有雨分量去除问题，提出一种基于双阶段特征解耦网络的单幅图像去雨方法，采用渐进式的学习方式从粗到细分阶段进行单幅图像去雨。首先构建压缩激励残差模块，实现背景图像与有雨分量的初步分离;然后设计全局特征融合模块，其中特别引入特征解耦模块分离有雨分量和背景图像的特征，实现细粒度的图像去雨;最后利用重构损失、结构相似损失、边缘感知损失和纹理一致性损失构成的复合损失函数训练网络，实现高质量的无雨图像重构。实验结果表明，在Testl00合成雨图数据集上，所提方法峰值信噪比为25。57dB，结构相似性为0。89;在100幅真实雨图上，所提方法的自然图像质量评估器为3。53，无参考图像空间质量评估器为20。16;在去雨后的RefineNet目标分割任务中，平均交并比为29。41％，平均像素精度为70。06％;视觉效果上，该方法能保留更多的背景图像特征，有效地辅助下游的目标分割任务的开展。
Two-Stage Feature Disentanglement Network for Single Image Rain Removal
Existing single image de-raining methods generally fail to balance the relationship between the detail restoration of rain-free background image and the removal of rain streaks.In this work,we proposed a single image de-raining method based on two-stage features disentanglement network from coarse to fine progressively.Firstly,we constructed the squeeze and excitation residual module to separate the background image and rainy components roughly.Then,we devised a novel context feature fusion and introduced a dis-entanglement module to decouple the feature relationship between rain streaks and the background image for a fine-grained de-rained image.Besides,we devised a composite loss function to train our network model for the restoration of high-quality rain-free images using Reconstruction loss function,SSIM loss function,Edge loss function and Texture loss function.Comparative experiments on Testl 00 dataset showed that our meth-od has a peak signal to noise ratio value of 25.57dB,a structural similarity value of 0.89.On 100 real rain rain images,a natural image quality evaluator value of 3.53 and a blind/referenceless image spatial quality evaluator value of 20.16 were obtained.On target segmentation task of RefineNet,the mean intersection over union and the mean pixel accuracy of proposed method were 29.41％and 70.06％,respectively.The proposed method shows clear visual results with more features of background image preserved for image rain removal and the assist of the downstream target segmentation task.

features disentanglement networksqueeze and excitation residual modulecontext feature fusion modulecomposite loss functionsingle image de-raining

汤红忠、熊珮全、王蔚、王晒雅、陈磊

展开 >

湘潭大学自动化与电子信息学院 湘潭 411105

湘潭大学智能计算与信息处理教育部重点实验室 湘潭 411105

湖南科技大学电气与信息工程学院 湘潭 411201

特征解耦网络 压缩激励残差模块 全局特征融合模块 复合损失函数 单幅图像去雨

国家自然科学基金区域创新发展联合基金子课题湖南省自然科学基金湖南省自然科学基金湖南省自然科学基金广东省基础与应用基础研究联合基金重点项目湘潭大学智能计算与信息处理教育部重点实验室开放课题

U19A20832020JJ45882020JJ409020JJ41512020B15151200502020ICIP06

2024

10.3724/SP.J.1089.2024.19828

计算机辅助设计与图形学学报
中国计算机学会

计算机辅助设计与图形学学报

CSTPCD北大核心
影响因子：0.892
ISSN：1003-9775
年,卷(期)：2024.36(2)
  • 29