电子学报 ›› 2020, Vol. 48 ›› Issue (7): 1293-1302.DOI: 10.3969/j.issn.0372-2112.2020.07.007

• 学术论文 • 上一篇    下一篇

基于低尺度细节恢复的单幅图像阴影去除方法

吴文1, 万毅2   

  1. 1. 新疆理工学院信息工程系, 新疆阿克苏 843100;
    2. 温州大学电气与电子工程学院, 浙江温州 325035
  • 收稿日期:2019-06-12 修回日期:2020-01-03 出版日期:2020-07-25 发布日期:2020-07-25
  • 通讯作者: 万毅
  • 作者简介:吴文 男,1994年5月出生,湖北武汉人.2019年毕业于湖北大学,获理学硕士学位,现为新疆理工学院信息工程系教员,主要研究方向为图像处理、深度学习等.E-mail:1119764335@qq.com
  • 基金资助:
    浙江省科技计划基础公益研究计划(No.LGG18F040002);浙江省自然科学基金(No.LY19F020035)

Single Image Shadow Removal Using Low-Scale Detail Recovering

WU Wen1, WAN Yi2   

  1. 1. Department of Information Engineering, Xinjiang Institute of Technology, Aksu, Xinjiang 843100, China;
    2. School of Electrical and Electronic Engineering, Wenzhou University, Wenzhou, Zhejiang 325035, China
  • Received:2019-06-12 Revised:2020-01-03 Online:2020-07-25 Published:2020-07-25

摘要: 为了在光照复杂、纹理丰富的图像上获得更好的去阴影效果,基于生成对抗网络提出了一种新颖的阴影去除方法.首先,所提网络中的阴影检测子网为阴影图像生成阴影掩膜,基于该检测结果提出一种光照敏感的多尺度图像分解方法,在几乎不损失光照信息的同时提取图像纹理信息;然后,蒙版生成子网为分解后的低尺度图像生成相应的蒙版用于去除其中阴影;其次,边界复原子网修复阴影边界实现友好的过渡;最后,使用自适应衰减因子引导图像进行细节恢复以得到纹理丰富的结果.实验结果表明所提方法可以有效地提高阴影去除效果.

关键词: 图像处理, 阴影去除, 生成对抗网络, 光照敏感, 多尺度分解

Abstract: To obtain a better shadow removal result on images with complex illumination and texture,we proposed a novel approach based on generative adversarial networks.Firstly,shadow mask is generated by the shadow detection sub-net from input shadow image.Based on this detection result,we proposed an illumination sensitive multi-scale image decomposition method to extract the texture information with less or no illumination information loss.Secondly,shadow matte is generated by the matte generation sub-net for the low scale shadow image to remove shadows in it.Thirdly,the shadow boundary are naturally recovered by the boundary completion sub-net.Finally,the shadow removal result is obtained using a detail recovering method guided by adaptive attenuation factor.Experimental results show that the proposed method can improve the removal performance effectively.

Key words: image processing, shadow removal, generative adversarial networks, illumination sensitive, multi-scale decomposition

中图分类号: