Self-Anchored Progressive Framework With Noise Mitigation for Unsupervised Camouflaged Object Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Unsupervised Camouflaged Object Detection (UCOD) is challenging due to object-background similarity and lack of annotations.
• Existing methods using pixel-level pseudo-labeling for Unsupervised Salient Object Detection (USOD) are unreliable for UCOD, leading to noisy labels and poor performance.

Purpose of the Study:

• To propose SAPNet, a self-anchored progressive framework to address the limitations of current UCOD methods.
• To improve the accuracy and boundary precision of camouflaged object detection in unsupervised settings.

Main Methods:

• SAPNet leverages semantically reliable foreground and background regions as high-confidence anchors, transforming UCOD into a weakly supervised problem.
• It employs a semantic-driven region detector (SDRD) with cascaded convolutions and residual attention to filter noise and enhance context.
• A region-based context inference module (RCIM) iteratively refines object boundaries using multi-level semantic features and region-level anchors.

Main Results:

• SAPNet significantly outperforms state-of-the-art unsupervised methods on four benchmark COD datasets.
• The proposed method demonstrates improved detection accuracy and boundary delineation compared to existing approaches.
• The framework effectively mitigates issues caused by noisy pseudo-labels in UCOD.

Conclusions:

• SAPNet offers a robust and effective solution for Unsupervised Camouflaged Object Detection.
• The self-anchored progressive approach provides reliable supervision signals, reducing learning difficulty and overfitting.
• The framework demonstrates the potential of leveraging high-confidence semantic regions for challenging detection tasks.