空洞卷积与注意力融合的对抗式图像阴影去除算法  

An antagonistic image shadow removal algorithm based on dilated convolution and attention mechanism

作　　者：刘万军[1] 佟畅 曲海成[1] LIU Wanjun;TONG Chang;QU Haicheng(Software College,Liaoning Technical University,Huludao 125105,China)

机构地区：[1]辽宁工程技术大学软件学院,辽宁葫芦岛125105

出　　处：《智能系统学报》2021年第6期1081-1089,共9页CAAI Transactions on Intelligent Systems

基　　金：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)。

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。To solve the problem of the unobvious shadow removal effect in dark areas or complex textured and penumbra 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, expanding 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.

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 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]