| 引用本文: | 郭继昌,郭昊,郭春乐.多尺度卷积神经网络的单幅图像去雨方法[J].哈尔滨工业大学学报,2018,50(3):185.DOI:10.11918/j.issn.0367-6234.201704075 |
| GUO Jichang,GUO Hao,GUO Chunle.Single image rain removal based on multi-scale convolutional neural network[J].Journal of Harbin Institute of Technology,2018,50(3):185.DOI:10.11918/j.issn.0367-6234.201704075 |
|
| 摘要: |
| 雨天气会影响户外拍摄图像质量，模糊和覆盖图像信息.为提高受雨天气影响的户外拍摄图像的清晰度，恢复图像特征信息，提高很多计算机视觉算法在雨天气条件下的准确性，提出一种基于多尺度卷积神经网络的单幅图像去雨方法.首先建立多尺度卷积神经网络网络结构，通过多尺度卷积提取图像信息，用于去除雨线和重建图像，然后结合雨线在图像中的低饱和度、高亮度的特征，对网络进行训练，获取网络最优参数值，最终得到可以有效去除雨线的卷积神经网络.实验结果表明：提出的方法相较于现有算法有更好的雨线去除效果，并且可以更好地保持图像的原有信息，避免图像模糊现象.同时，利用多尺度卷积提取图像特征信息可以使特征信息更加丰富，有利于提升卷积神经网络的去雨能力.
|
| 关键词: 图像特征信息 计算机视觉算法 图像去雨 多尺度卷积神经网络 低饱和度、高亮度 |
| DOI：10.11918/j.issn.0367-6234.201704075 |
| 分类号:TP391.4 |
| 文献标识码:A |
| 基金项目:国家重点基础研究发展计划(2014CB340400) |
|
| Single image rain removal based on multi-scale convolutional neural network |
|
GUO Jichang,GUO Hao,GUO Chunle
|
|
(School of Electrical and Information Engineering, Tianjin University, Tianjin 300072, China)
|
| Abstract: |
| Rain weather will affect the quality of the outdoor images and blur and cover the image information. To improve the clarity of outdoor images affected by rain streaks, restore image feature information and improve the accuracy of many computer vision algorithms under rain weather condition, a single image rain streaks removal method based on the multi-scale convolutional neural network is proposed. Firstly, the structure of the multi-scale convolutional neural network is established and the method extracts information of images by multi-scale convolution. The information is utilized to remove the rain streaks and reconstruct images. Then, the method combines the features of rain which include low saturation and high intensity to train the network. The optimal parameters can be obtained in this way. Finally the convolutional neural network can remove the rain streaks in single images effectively. Experimental results show that the proposed method compared with the existing methods has better performance about rain streaks removal. It can better maintain the original information of the image and avoid blurring of the image. Meanwhile, using multiscale convolution to extract feature information can enrich the feature information and it is beneficial to improve the rain removal ability of the convolution neural network.
|
| Key words: image feature information computer vision algorithm rain streaks removal multi-scale convolutional neural network low saturation and high intensity |
