Scribble-Guided Weakly-Supervised Edge-Enhanced Saliency Detection with Centroid Optimization
Xin Fang, Ming Zhao · 2025
Salient Object Detection (SOD) aims to identify and locate objects in images. And it is widely used in tasks such as target recognition and image segmentation. Recently, the weakly supervised methods using limited labelled data. However, existing methods have difficulties in boundary accuracy and detail representation. In order to these problems, we propose a new weakly supervised saliency target detection method that integrates which integrates several innovative modules. First, we introduce an Edge Detection Network (EDN) to refine the boundary of the saliency graph by enhancing its structure and using gating to remove structural perceptual loss. Next, we introduce a feature fusion branch that combines low-level saliency features with edge and semantic cues to enhance fine-grained details in the details in the saliency graph. In addition, we propose a centroid-guided optimization method that uses graffiti annotations to refine regions of interest, thereby improving the accuracy of the accuracy of saliency localization. Experimental results show that our method outperforms existing techniques on multi-criteria datasets, especially in terms of boundary accuracy and detail preservation.