电子科技2025，Vol.38Issue(9)：49-57,9.DOI:10.16180/j.cnki.issn1007-7820.2025.09.007
基于类双重激活映射和边缘检测的弱监督实例分割
Class Double-Activation Maps for Weakly-Supervised Instance Segmentation with Boundary Detection
摘要
Abstract
Weakly supervised instance segmentation based on image-level category labels has attracted much attention due to its cheap and efficient cost of labeling,in which extracting CAM(Class Activation Map)from classi-fication network is a key step.Most methods use BCE(Binary Cross-Entropy)loss to train classification models.However,since BCE loss is not mutually exclusive,traditional CAM has activation error and activation bias.In order to solve this problem,SCE(Softmax Cross-Entropy)loss is added to the classifier,and CAM is activated twice to ex-tract more efficient and accurate Double-CAM(Class Double-Activation Map).A refined method of edge detection is proposed to mine target boundary clues from Double-CAM,and constrain label propagation more explicitly without adding additional supervisory information,so as to further improve the quality of pseudo-labels.The experimental re-sults show that the pseudo-mask generated by the proposed method has good segmentation performance under mAP50(48.2％)and mAP75(24.7％)indexes,exceeding the mainstream model with the same weak supervision lev-el,and reaching 70％of the performance of the full supervision method.关键词
图像级类别标签/弱监督学习/实例分割/交叉熵损失/类互斥性/类双重激活映射/边缘检测/伪掩码Key words
image-level class labels/weakly supervised learning/instance segmentation/cross-entropy loss/class mutual exclusivity/class double-activation map/boundary detection/pseudo masks分类
信息技术与安全科学引用本文复制引用
彭琎,王永雄,潘志群..基于类双重激活映射和边缘检测的弱监督实例分割[J].电子科技,2025,38(9):49-57,9.基金项目
上海市自然科学基金(22ZR1443700)Natural Science Foundation of Shanghai(22ZR1443700) （22ZR1443700）
