基于特征交互引导的弱监督显著目标检测  

Weakly supervised salient object detection via feature interaction guidance

作　　者：李永强 石艳娇 张晴 LI Yong-qiang;SHI Yan-jiao;ZHANG Qing(School of Computer Science and Information Engineering,Shanghai Institute of Technology,Shanghai 201418,China)

机构地区：[1]上海应用技术大学计算机科学与信息工程学院,上海201418

出　　处：《计算机工程与设计》2025年第4期1234-1240,F0003,共8页Computer Engineering and Design

基　　金：国家自然科学基金项目(61806126);上海市自然科学基金项目(21ZR1462600);上海应用技术大学科技发展基金项目(ZQ2023-15)。

摘　　要：为解决现有的基于稀疏标记的弱监督显著目标检测模型需要引入额外标签恢复显著目标结构、模型检测效果不佳等问题,提出一种基于涂鸦注释的特征交互引导网络用于弱监督显著目标检测,无需预处理/后处理操作或额外标签。模型集成特征增强机制分别对浅层和深层的多尺度特征进行扩展与聚合,轮廓重定位模块恢复并优化显著目标结构,所提全局特征对齐损失辅助模型感知显著目标的全局结构。实验结果表明,所提出模型优于现有的弱监督方法。To address the issues of introducing additional labels to restore the salient object structure and poor model detection performance in existing weakly supervised salient object detection(SOD)models based on sparse labeling,a feature interaction guidance network(FIGNet)based on scribble annotation was proposed for weakly supervised salient object detection,without the need for preprocessing/post-processing operations or additional labels.The model integrated feature enhancement mechanism to extend and aggregate multi-scale features of shallow and deep layers,recovering and optimizing the structure of salient object through contour repositioning module(CRM),and global feature alignment(GFA)loss was proposed to assist model perceiving the global structure of salient object.Experimental results show that the proposed model outperforms existing weakly supervised methods.

关 键 词：显著目标检测 弱监督 涂鸦注释 特征交互引导 特征增强 多尺度特征 特征对齐 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]