Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (16): 234-241.DOI: 10.3778/j.issn.1002-8331.2101-0092

• Graphics and Image Processing • Previous Articles     Next Articles

Image Shadow Removal Algorithm Based on Attention and Multi-Scale Fusion

QU Haicheng, TONG Chang, LIU Wanjun   

  1. School of Software, Liaoning Technical University, Huludao, Liaoning 125105, China
  • Online:2022-08-15 Published:2022-08-15

注意力与多尺度融合的图像阴影去除算法

曲海成,佟畅,刘万军   

  1. 辽宁工程技术大学 软件学院,辽宁 葫芦岛 125105

Abstract: In order to solve the problem of incomplete shadow removal for complex objects or dark areas with similar texture to shadow areas, an image shadow removal algorithm based on attention and multi-scale fusion is proposed. The algorithm is based on the framework of generative adversarial network. Firstly, the user-defined hole residual block is used for feature extraction to obtain accurate shadow feature information and input into the attention guided coding network. Secondly, multi-scale feature fusion is carried out in the encoding stage of attention guidance to obtain different levels of receptive field, so that the coder can capture context information at different scales. Finally, the multiple attention mechanism is used to guide the discriminant network to identify the unshadowed image, thus reducing the loss of key information of the discrimination network and improving the identification effect of the discrimination network. The algorithm is tested on ISTD(dataset with image shadow triplets) and SRD(dataset for shadow removal) public datasets respectively. The results show that the algorithm has good visual effect. Comparing the single shadow removed image with the real unshadowed image in the dataset, SSIM(structural similarity) can reach 0.978, PSNR(peak signal to noise ratio) can reach 32.2 dB, RMSE(root mean squared error) can reach 6.2. Compared with the similar algorithms, it has significant advantages, and has good shadow removal effect for complex objects or dark areas.

Key words: attention mechanism, multi-scale feature fusion, generative adversarial networks, shadow removal

摘要: 针对图像阴影去除算法中复杂地物或与阴影区域纹理相似的暗区域阴影去除不完全的问题,提出了一种注意力与多尺度融合的图像阴影去除算法。该算法基于生成对抗网络框架构建。利用自定义的空洞残差块进行特征提取,获得精确的阴影特征信息并输入到注意力引导的编码网络;在注意力引导的编码阶段进行多尺度的特征融合,获取不同层次的感受野,使编码器能够在不同尺度上捕捉上下文信息;利用多重注意力机制引导判别网络对生成的无阴影图像进行鉴别,进而减少判别网络关键信息损失,提高判别网络的鉴别效果。分别在ISTD(dataset with image shadow triplets)与SRD(dataset for shadow removal)公开数据集上进行验证,实验结果表明:该算法视觉效果表现良好,单幅阴影去除后的图片与数据集中真实无阴影图片进行对比,SSIM(structural similarity)可达到0.978,PSNR(peak signal to noise ratio)可达到32.2?dB,RMSE(root mean squared error)可达到6.2,相比同类算法,具有显著优势,且对复杂地物或暗区域阴影去除效果良好。

关键词: 注意力机制, 多尺度特征融合, 生成对抗网络, 阴影去除