计算机工程与应用 ›› 2021, Vol. 57 ›› Issue (15): 215-222.DOI: 10.3778/j.issn.1002-8331.2005-0178

• 图形图像处理 • 上一篇    下一篇

基于注意力生成对抗网络的单幅图像去雨方法

朱德利,熊昌,胡雪奎,李炜,王青   

  1. 1.重庆师范大学 计算机与信息科学学院,重庆 401331
    2.重庆市数字农业服务工程技术研究中心,重庆 401331
  • 出版日期:2021-08-01 发布日期:2021-07-26

Single Image De-raining Method Based on Attention Generation Adversarial Network

ZHU Deli, XIONG Chang, HU Xuekui, LI Wei, WANG Qing   

  1. 1.College of Computer and Information Science, Chongqing Normal University, Chongqing 401331, China
    2.Chongqing Digital Agricultural Service Engineering Technology Research Center, Chongqing 401331, China
  • Online:2021-08-01 Published:2021-07-26

摘要:

下雨是一种常见的天气现象,而滞留在图像上的雨条纹降低了图像的清晰度以及影响了基于该图像的后续图像处理。从图像中去除雨的关键是如何准确、鲁棒地识别图像中的雨区域。使用导向滤波器和Haar小波变换组成的雨线提取模块来增强雨条纹特征提取,然后通过空间关注模块生成雨线注意力图,以准确定位雨条纹的位置。两者结合后,获得降雨条纹的前景信息再通过生成对抗网络训练机制中相互博弈的特征,可以增强雨条位置识别能力,并有效地去除雨条纹。在综合测试数据集和真实图像上进行实验,对比几种深度网络去雨方法,峰值信噪比(PSNR)和结构相似比(SSIM)都得到提升。实验表明,该网络具有出色的性能,对于不同的雨条纹密度具有较高的泛化能力同时可以更好地保持图像的原有信息,避免图像模糊现象。

关键词: 生成对抗网络, 注意力机制, 图像去雨

Abstract:

Raining is a common weather phenomenon, and the rain streaks left on the image reduce the clarity of the image and affect the subsequent image processing based on the image. The key to removing rain from an image is how to accurately and robustly identify rain areas in the image. In this paper, the rain line extraction module composed of guided filter and Haar wavelet transform is used to enhance the rain stripe feature extraction, and then the rain line attention map is generated by the spatial attention module to accurately locate the position of the rain stripe. After combining the two, the foreground information of the rain stripes can be obtained and then generating the characteristics of the mutual game in the training mechanism of the adversarial network can enhance the position recognition ability of the rain bars and effectively remove the rain stripes. Experiments are conducted on the comprehensive test data set and real images. Compared with several deep network de-raining methods, the Peak Signal-to-Noise Ratio(PSNR) and Structural Similarity ratio(SSIM) are improved. Experiments show that the network in this paper has excellent performance, and has a high generalization ability for different rain streak densities. At the same time, it can better maintain the original information of the image and avoid image blur.

Key words: generative adversarial networks, attention mechanism, image derain