基于无监督学习的两阶段显著性目标检测

Two-Stage Unsupervised Salient Object Detection

摘要:

针对传统显著性目标检测依赖大量人工标注数据,成本高且泛化能力有限等问题,提出了一种基于无监督学习的两阶段显著性目标检测方法；该方法主要通过两个阶段实现在复杂场景中快速定位并分割出最具吸引力的区域,一阶段构建融合全局与局部特征的伪标签生成器(GLEPG),通过差分运算与自适应权重机制提升伪标签质量；第二阶段设计一个RGB显著图细化网络(SMRNet)对初始显著图进行细化,并利用高质量伪标签对标准检测网络进行训练,从而获得更准确的显著性预测结果；通过在几个常用RGB数据集上的实验结果表明,该方法在保持无监督学习优势的同时,有效提升了显著图的完整性与细节精度,取得了较优的检测效果。

Abstract:

Aiming at the problems of traditional salient object detection, such as reliance on large amounts of manually annotated data, high cost, and limited generalization ability, this paper proposed a novel two-stage unsupervised salient object detection method. The method rapidly located and segmented the most visually attractive regions in complex scenes through two stages. In the first stage, it constructed a Global-Local feature-based Pseudo Generator(GLEPG).This generator enhanced pseudo-label quality by applying differential operations and an adaptive weighting mechanism. In the second stage, it designed a RGB Saliency Map Refinement Network to refine the initial saliency maps. It also used the high-quality pseudo-labels to train a standard detection network for more accurate saliency predictions. Experimental results on multiple RGB datasets demonstrate that the proposed method not only preserves the benefits of unsupervised learning but also significantly enhances the completeness and detail accuracy of the saliency maps, achieving competitive detection performance.