边缘深度挖掘的弱监督显著性目标检测  

Weakly Supervised Salient Object Detection via Edge Depth Mining

作　　者：李军侠 王星驰 殷梓 石德硕 LI Junxia;WANG Xingchi;YIN Zi;SHI Deshuo(School of Computer Science,Nanjing University of Information Science and Technology,Nanjing 210044,China;School of Automation,Nanjing University of Information Science and Technology,Nanjing 210044,China)

机构地区：[1]南京信息工程大学计算机学院,南京210044 [2]南京信息工程大学自动化学院,南京210044

出　　处：《计算机工程》2023年第7期169-178,共10页Computer Engineering

基　　金：科技创新2030—“新一代人工智能”重大项目(2018AAA0100400)。

摘　　要：基于深度学习的显著性目标检测算法大多依赖于大规模标注数据下的监督学习模式,但是,样本的像素级标签存在获取困难、标注成本高的问题。为此,设计一种边缘深度挖掘的弱监督显著性目标检测算法,仅使用图像级类别标签,从输入图像的显著性目标边缘角度得到能够较准确描述目标轮廓信息的边缘特征图,以生成伪标签对显著性模型进行监督训练。通过粗糙边缘生成模块对显著性目标轮廓特征进行简单标定,用于获取粗糙边缘特征图。在此基础上,利用精细边缘生成模块优化特征以得到精细边缘特征图,能够更准确地反映目标的边界信息,在完整刻画轮廓信息的同时可以更好地抑制背景噪声。伪标签生成模块基于精细边缘特征图生成像素级伪标签。实验结果表明,相比MSW、MFNet、NSAL等算法,该算法能够准确识别显著性区域,获得的预测图具有较完整的细节信息,其中,在ECSSD数据集上S-measure值和E-measure值相较于第2名NSAL算法分别提高1.1和0.6个百分点。Salient Object Detection(SOD)algorithms based on deep learning primarily use supervised learning for large-scale labeled data.However,obtaining pixel-level labels for samples is difficult and labeling costs are high.To this end,a weakly supervised SOD algorithm for edge depth mining is designed.It uses image-level category labels to obtain edge feature maps that accurately describe the target's contour information from the perspective of the salient object edge of the input image,producing pseudo labels for supervised training of the saliency model.Through the Coarse Edge Generation(CEG)module,simple calibration of salient target contour features is performed to obtain rough edge feature maps.The Fine Edge Generation(FEG)module is subsequently used to optimize features to obtain a fine edge feature map,which accurately reflects the boundary information of the target and suppresses background noise for complete contour detection.The Pseudo Label Generation(PLG)module generates pixel-level pseudo labels based on fine edge feature maps.Experimental results show that in contrast to algorithms such as MSW,MFNet,and NSAL,the proposed algorithm accurately identifies significant regions and obtains prediction maps with complete detail information.Compared to the second place NSAL algorithm,the S-measure and E-measure values of the proposed algorithm increase by 1.1 and 0.6 percentage points on ECSSD dataset,respectively.

关 键 词：弱监督 显著性目标检测 深度学习 图像级类别标签 伪标签 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]