王琪1,,,
毕晓君2
1.哈尔滨理工大学电气与电子工程学院 哈尔滨 150000
2.哈尔滨工程大学信息与通信工程学院 哈尔滨 150000
基金项目:国家自然科学基金(51779050)
详细信息
作者简介:柳长源:男,1970年生,副教授,工学博士,硕士生导师,研究方向为模式识别与图像处理,机器学习与优化方法
王琪:女,1996年生,硕士生,研究方向为模式识别与图像处理
毕晓君:女,1964年生,教授,博士生导师,研究方向为数字图像处理、信息智能处理技术及通信信息处理技术
通讯作者:王琪 1208401521@qq.com
中图分类号:TN911.73; TP391.4计量
文章访问数:1520
HTML全文浏览量:377
PDF下载量:115
被引次数:0
出版历程
收稿日期:2019-09-29
修回日期:2020-05-28
网络出版日期:2020-07-13
刊出日期:2020-09-27
Research on Rain Removal Method for Single Image Based on Multi-channel and Multi-scale CNN
Changyuan LIU1,Qi WANG1,,,
Xiaojun BI2
1. School of Electrical and Electronic Engineering, Harbin University of Science and Technology, Harbin 150000, China
2. School of Information and Communication Engineering, Harbin Engineering University, Harbin 150000, China
Funds:The National Natural Science Foundation of China(51779050)
摘要
摘要:雨天等恶劣天气会严重影响到图像成像质量,从而影响到视觉处理算法的性能。为了改善雨天图像的成像质量,该文提出一种基于多通道多尺度卷积神经网络的去雨算法,建立了多通道多尺度卷积神经网络结构来提取雨线特征。首先利用小波阈值引导的双边滤波将有雨图像进行分解,得到高频雨线图像和轮廓保持度高的低频背景图像。然后为了使图像高频部分的雨线信息更为明显,减少雨线特征学习时高频图像中的背景误判,将得到的高频雨线图像再一次通过滤波器得到减弱背景信息同时增强雨线信息的到更高频雨线图像。其次针对低频背景图像上也残留了大量雨痕,该文提出将低频背景图像和更高频雨线图像一起送入卷积神经网络进行特征学习,其中对图像提取的是多尺度特征信息,最后得到雨线去除更彻底的复原图像。同时在构造网络模型时利用空洞卷积代替标准卷积来提取图像的特征信息,得到更丰富的图像特征,提高了算法的去雨性能。从实验结果可以看出去雨之后的图像清晰,细节保持度较高。
关键词:深度学习/
空洞卷积/
图像分解/
多尺度提取特征
Abstract:Rainy days and other severe weather will seriously affect the image quality, thus affecting the performance of vision processing algorithms. In order to improve the imaging quality of rain images, a rain removal algorithm based on multi-channel multi-scale convolution neural network to extract rain line features is proposed. Firstly, the rain images are decomposed by wavelet threshold-guided bilateral filtering to obtain high-frequency rain line images and low-frequency background images with high contour preservation. Then, in order to make the rain line information in the high-frequency part of the image more obvious and reduce the background misjudgment in the high-frequency image during the rain line feature learning, the obtained high-frequency rain line image is passed through a filter again to obtain a higher-frequency rain line image with reduced background information and enhanced rain line information. Secondly, in view of the large amount of raindrop imprint left on the low-frequency background image, it is proposed to send the low-frequency background image and the higher-frequency rain line image together into the convolution neural network for feature learning, in which multi-scale feature information is extracted from the image, finally, a more complete restoration image with rain line removal is obtained. At the same time, when constructing the network model, hole convolution is used instead of standard convolution to extract the feature information of the image, thus obtaining richer image features and improving the rain removal performance of the algorithm. From the experimental results, after removing rain, the image is clear and the detail retention is high.
Key words:Deep learning/
Dilated convolution/
Image decomposition/
Multi-scale feature extraction
PDF全文下载地址:
https://jeit.ac.cn/article/exportPdf?id=1232074f-9076-42be-bfa3-e727aca03355