第16卷第6期 智能系统学报 Vol.16 No.6 2021年11月 CAAI Transactions on Intelligent Systems Nov.2021 D0:10.11992/tis.202011022 网络出版地址：https:/ns.cnki.net/kcms/detail/23.1538.TP.20210901.1633.008.html 空洞卷积与注意力融合的对抗式图像阴影去除算法 刘万军，佟畅，曲海成 (辽宁工程技术大学软件学院，辽宁葫芦岛125105) 摘要：为了解决暗区域、纹理复杂或半影区域的阴影去除效果不明显的问题，提出了空洞卷积与注意力机制 融合的对抗式图像阴影去除算法。该算法基于生成对抗网络的总体思想，将空洞卷积引入残差网络中，用自定 义的空洞残差块进行特征提取，扩大了特征提取的感受野。在注意力编码阶段，加入4层相同结构的空洞卷 积，确保最小计算量的情况下为解码阶段提供更抽象、更本质的全局的语义特征。运用多重注意力机制，引导 判别网络对无阴影图像的鉴别，提高判别网络能力。该算法分别在ISTD(image shadow triplets dataset))与 SRD(shadow removal dataset)公开数据集上进行检验，SSIM(structural similarity)值达到97.77%。该算法图像特征 信息保存完整，画面请晰，暗区域及地物复杂的区域阴影去除效果较好，对半影区域，也有具有良好的表现。 关键词：生成对抗网络；空洞卷积；多重注意力：残差网络；多尺度；自编码：长短记忆法：阴影去除 中图分类号：TP391.41文献标志码：A文章编号：1673-4785(2021)06-1081-09 中文引用格式：刘万军，佟畅，曲海成.空洞卷积与注意力融合的对抗式图像阴影去除算法J.智能系统学报，2021,16(6)： 1081-1089. 英文引用格式：LIUWanjun,TONG Chang,QU Haicheng.An antagonistic image shadow removal algorithm based on dilated con- volution and attention mechanismJ.CAAI transactions on intelligent systems,2021,16(6):1081-1089. An antagonistic image shadow removal algorithm based on dilated convolution and attention mechanism LIU Wanjun,TONG Chang,QU Haicheng (Software College,Liaoning Technical University,Huludao 125105,China) Abstract:To solve the problem of the unobvious shadow removal effect in dark areas or complex textured and penum- bra areas,an antagonistic image shadow removal algorithm is proposed based on a dilated convolution and attention mechanism.The algorithm is based on the general idea of generative adversarial networks.First,the dilated convolution is introduced into the residual network,and the user-defined hole residual block is used for feature extraction,expand- ing the receptive field of feature extraction.Second,in the attention coding stage,four layers of dilated convolution with the same structure are added to provide more abstract and essential global semantic features for the decoding phase with a minimum calculation amount.Finally,the multiple attention mechanism is used to guide the discrimination network to identify the unshadowed image;thus,improving the discrimination network's ability.The proposed algorithm is tested on image shadow triplets dataset and shadow removal public datasets and achieves the structural similarity of 97.77%. The image feature information of the algorithm is well preserved,the picture is clear,the shadow removal effect is good in the dark area and complex area,and the algorithm has good performance for the penumbra area. Keywords:generative adversarial networks;hole convolution;multiple attention;residual network;multi-scale;auto- coder:long short-term memory:shadow removal 光线穿过一切不透明物体都会产生阴影，图 收稿日期：2020-11-20.网络出版日期：2021-09-01. 基金项目：国家自然科学基金项目(41701479)：辽宁省教育厅 像中阴影的存在十分常见。图像阴影会对目标检 科学研究经费项目(L2019几010)：辽宁省自然科学 测与识别、图像分割等问题产生一定影响。因此， 基金面上项目(20180550529). 通信作者：佟畅.E-mail:1879031567@qq.com 去除图像中的阴影是一个十分重要的研究内容。DOI: 10.11992/tis.202011022 网络出版地址: https://kns.cnki.net/kcms/detail/23.1538.TP.20210901.1633.008.html 空洞卷积与注意力融合的对抗式图像阴影去除算法 刘万军，佟畅，曲海成 （辽宁工程技术大学 软件学院，辽宁 葫芦岛 125105） 摘 要：为了解决暗区域、纹理复杂或半影区域的阴影去除效果不明显的问题，提出了空洞卷积与注意力机制 融合的对抗式图像阴影去除算法。该算法基于生成对抗网络的总体思想，将空洞卷积引入残差网络中，用自定 义的空洞残差块进行特征提取，扩大了特征提取的感受野。在注意力编码阶段，加入 4 层相同结构的空洞卷 积，确保最小计算量的情况下为解码阶段提供更抽象、更本质的全局的语义特征。运用多重注意力机制，引导 判别网络对无阴影图像的鉴别，提高判别网络能力。该算法分别在 ISTD(image shadow triplets dataset) 与 SRD(shadow removal dataset) 公开数据集上进行检验，SSIM(structural similarity) 值达到 97.77%。该算法图像特征 信息保存完整，画面清晰，暗区域及地物复杂的区域阴影去除效果较好，对半影区域，也有具有良好的表现。 关键词：生成对抗网络；空洞卷积；多重注意力；残差网络；多尺度；自编码；长短记忆法；阴影去除 中图分类号：TP391.41 文献标志码：A 文章编号：1673−4785(2021)06−1081−09 中文引用格式：刘万军, 佟畅, 曲海成. 空洞卷积与注意力融合的对抗式图像阴影去除算法 [J]. 智能系统学报, 2021, 16(6): 1081–1089. 英文引用格式：LIU Wanjun, TONG Chang, QU Haicheng. An antagonistic image shadow removal algorithm based on dilated con￾volution and attention mechanism[J]. CAAI transactions on intelligent systems, 2021, 16(6): 1081–1089. An antagonistic image shadow removal algorithm based on dilated convolution and attention mechanism LIU Wanjun，TONG Chang，QU Haicheng (Software College, Liaoning Technical University, Huludao 125105, China) Abstract: To solve the problem of the unobvious shadow removal effect in dark areas or complex textured and penum￾bra areas, an antagonistic image shadow removal algorithm is proposed based on a dilated convolution and attention mechanism. The algorithm is based on the general idea of generative adversarial networks. First, the dilated convolution is introduced into the residual network, and the user-defined hole residual block is used for feature extraction, expand￾ing the receptive field of feature extraction. Second, in the attention coding stage, four layers of dilated convolution with the same structure are added to provide more abstract and essential global semantic features for the decoding phase with a minimum calculation amount. Finally, the multiple attention mechanism is used to guide the discrimination network to identify the unshadowed image; thus, improving the discrimination network’s ability. The proposed algorithm is tested on image shadow triplets dataset and shadow removal public datasets and achieves the structural similarity of 97.77%. The image feature information of the algorithm is well preserved, the picture is clear, the shadow removal effect is good in the dark area and complex area, and the algorithm has good performance for the penumbra area. Keywords: generative adversarial networks; hole convolution; multiple attention; residual network; multi-scale; auto￾coder; long short-term memory; shadow removal 光线穿过一切不透明物体都会产生阴影，图 像中阴影的存在十分常见。图像阴影会对目标检 测与识别、图像分割等问题产生一定影响。因此， 去除图像中的阴影是一个十分重要的研究内容。 收稿日期：2020−11−20. 网络出版日期：2021−09−01. 基金项目：国家自然科学基金项目 (41701479)；辽宁省教育厅 科学研究经费项目 (LJ2019JL010)；辽宁省自然科学 基金面上项目 (20180550529). 通信作者：佟畅. E-mail：1879031567@qq.com. 第 16 卷第 6 期 智 能 系 统 学 报 Vol.16 No.6 2021 年 11 月 CAAI Transactions on Intelligent Systems Nov. 2021