.png){kind=icon}
.png){kind=icon}
Return
Fighting background misjudgment for unsupervised salient object detection
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2026.132504.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Unsupervised salient object detection (SOD) methods based on contrastive learning pre-trained models have achieved remarkable progress. However, due to the inherent foreground bias in such models, existing approaches often suffer from background misjudgment, particularly in regions adjacent to foreground objects. Moreover, current evaluation metrics for background errors (e.g., false positives) fail to reflect the severity of misjudgments across different spatial regions. For SOD task, misjudging background pixels near the foreground can severely disrupt object structure and compromise the performance of downstream tasks, while misjudgments in distant background regions are generally more tolerable. This spatial discrepancy is crucial for model evaluation but has been overlooked. To address this issue, we propose a novel evaluation metric, Background Misjudgment Rate (BMR omega), which dynamically weights misjudged background pixels based on their spatial distance to the foreground, offering region-sensitive guidance for downstream applications. Furthermore, we introduce a refinement module to mitigate background misjudgment in existing unsupervised SOD methods. Built upon the initial saliency maps generated by a base model (e.g., A2S-v3), our approach consists of a Foreground-Background Partition Component (FBPC) and a Pseudo-Foreground Suppression Component (PFSC). FBPC adaptively determines the ratio of foreground and background blocks by analyzing the global saliency distribution of image patches, breaking the limitations of fixed thresholds or local contrast-based segmentation. PFSC further leverages the statistical separability of foreground and background features by constructing a Gram matrix-based feature difference enhancement mechanism. It introduces a background misjudgment penalty loss to enlarge the foreground-background distance in feature space, effectively suppressing pseudo-foreground activation. The proposed module is model-agnostic and can be readily integrated into various unsupervised SOD frameworks. Extensive experiments on five standard benchmark datasets demonstrate that our method consistently outperforms eight state-of-the-art unsupervised approaches, with significant advantages in reducing background misjudgment.
Keywords:
Salient object detection
Unsupervised learning
Background misjudgment
Journal
IF:
7.5
Papers:
2.9W
Citations:
10.2W