基于多任务判别器的图像阴影去除算法  

Image shadow removal algorithm based on multi-task discriminator

作　　者：江灵雅 黄友锐[1] 韩涛[1] 徐善永[1] JIANG Ling-ya;HUANG You-rui;HAN Tao;XU Shan-yong(College of Electrical and Information Engineering,Anhui University of Science and Technology,Huainan 232001,China)

机构地区：[1]安徽理工大学电气与信息工程学院,安徽淮南232001

出　　处：《哈尔滨商业大学学报（自然科学版）》2023年第2期139-146,共8页Journal of Harbin University of Commerce：Natural Sciences Edition

基　　金：国家自然科学基金(61772033);安徽省高校协同创新项目(GXXT-2019-048,GXXT-2020-54)。

摘　　要：针对现实生活中存在的阴影对于计算机视觉任务产生不利影响,现有图像阴影去除算法有效性与准确性低的问题,提出了一种基于生成对抗网络的阴影去除方法.该方法主要包括两个部分,采用了U-Net网络结构作为生成器部分的网络,用来提取图像中的有效信息,同时在生成器网络中引入自注意力机制,提高网络对阴影区域特征的关注度;采用多任务网络结构的判别器,通过对生成器生成图像的整体区域、局部区域以及与约束条件的匹配程度,进行多方面的鉴别,用来提升判别器鉴别能力,进而督促生成器,生成更加符合真实情况的无阴影图像.该方法在公开数据集ISTD上进行实验,实验结果表明所提方法在图像的结构相似性上以及峰值信噪比方面均有所提升.Aiming at the adverse effects of shadows in real life on computer vision tasks and the low effectiveness and accuracy of existing image shadow removal algorithms,a shadow removal method based on generative adversarial networks was proposed.The method mainly includes two parts.The U-Net network structure was used as the generator part of the network to extract the effective information in the image.At the same time,the self-attention mechanism was introduced into the generator network to improve the attention of the network to the shadow area features.The discriminator of multi-task network structure was used to identify the whole area,local area and matching degree with constraints of the image generated by the generator,so as to improve the discriminator s ability to identify,and then urge the generator to generate a shadow-free image more in line with the real situation.The method was tested on the public data set ISTD.The experimental results showed that the proposed method improved the structural similarity of the image and the peak signal-to-noise ratio.

关 键 词：阴影去除 生成对抗网络 自注意力机制 马尔可夫判别器 U-Net网络 多任务判别器 

分 类 号：TP317.4[自动化与计算机技术—计算机软件与理论]