多阶段多尺度去雨网络
Multi-Stage and Multi-Scale Deraining Network
张玉波 1王晓彤1
作者信息
- 1. 东北石油大学电气信息工程学院,黑龙江 大庆 163319
- 折叠
摘要
针对编解码器结构容易导致图像的纹理细节缺失,而单尺度结构处理的图像全局语义信息不够准确的问题,设计多阶段多尺度去雨网络(MMDNet).具体来说,首先提出了一个有效的特征注意模块,利用上述模块组成编解码器、单尺度模块,以在不同尺度上学习雨纹的特征;此外,引入语义特征融合模块来融合编解码器的特征,将丰富的语义信息传入下一阶段;更进一步,在网络的阶段间设计了信息过渡模块,上述模块一方面将浅层信息过渡到下一阶段,另一方面起到监督作用.实验结果表明,提出的算法在Rain100H、Rain100L和Test100 数据集上的性能与其它最新的去雨算法相比具有竞争力.
Abstract
Aiming at the problem that the encoder-decoder structure easily leads to the lack of texture details of the image,and the global semantic information of the image processed by the single-scale structure is not accurate e-nough,a Multi-stage and Multi-scale Deraining Network(MMDNet)is designed.Specifically,an effective feature at-tention module is firstly proposed,which is used to form an encoder-decoder and a single-scale module to learn the features of rain streaks at different scales;In addition,a semantic feature fusion module is introduced to fuse the fea-tures of the encoder and decoder,and the rich semantic information is passed to the next stage;Furthermore,an infor-mation transition module is designed between the stages of the network,which on the one hand transitions the shallow information to the next stage,and on the other hand plays a supervisory role.Comprehensive experiments demonstrate that the performance of the proposed algorithm is competitive with other state-of-the-art rain removal algorithms on the Rain100H,Rain100L and Test100 datasets.
关键词
图像去雨/卷积神经网络/注意力机制Key words
Image deraining/Convolutional neural network/Attention mechanism引用本文复制引用
基金项目
东北石油大学青年基金项目(15071202202)
出版年
2024
NETL
NSTL
万方数据