.png){kind=icon}
.png){kind=icon}
Return
A Weakly Supervised Salient Object Detection Framework Based on Structured Scribble
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neucom.2025.131057.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Weakly Supervised Salient Object Detection (WSOD) methods based on scribble annotations provide a solution to the annotation burden from pixel-level annotations. However, due to the sparse nature of scribble, models built upon them suffer from structure deficiency. In this paper, we propose to learn object structural information from the structured scribbles to address this issue. Structured scribbles incorporate object structural information into the scribble annotations by utilizing the object skeleton feature. On this basis, We propose a novel simple yet effective WSOD learning framework, which consists of a Multiscale Connection Module (MCM), a Detail Mining Module (DMM), and a Structure Guidance Module (SGM). MCM decodes and filters features at different scales to enhance the decoder’s ability to restore structures. DMM suppresses noise and extracts detail features via dual-stream attention. SGM integrates structure and detail features through a fusion, enhancement, and selection strategy to obtain the final prediction map. Moreover, we propose a plug-and-play Proportional Data Augmentation method to alleviate the impact of feature deformation and enhance the model’s generalization and practicality. Extensive experiments demonstrate that our method outperforms state-of-the-art WSOD methods, even reaching the performance level of fully supervised saliency object detection methods. The dataset and code will be available at https://github.com/Rouiy/SeNet .
Keywords:
scribble annotations
weakly supervised salient object detection
structured scribbles
multiscale connection module
structure guidance module