Weakly supervised camouflaged object detection as Progressive Perception Learning

Tianxin Han

doi:10.1016/j.knosys.2025.113993

Return

Weakly supervised camouflaged object detection as Progressive Perception Learning

2025-06-30

PRE

DOI：10.1016/j.knosys.2025.113993

Original

Original request for help

Save

Abstract

En 中文

Scribble-Supervised Camouflaged Object Detection (SCOD) aims to detect camouflaged objects that seamlessly blend into their surroundings using sparse and coarse scribble annotations as weak supervision. However, existing SCOD methods face two main natural deficiencies. First, the weak richness of pixel training samples (WRPS) hinders robust camouflaged object perception learning, as scribble annotations provide insufficient pixel-level supervision. Second, the poor structural integrity of camouflaged objects (PSIO) makes direct structural learning infeasible, resulting in imprecise segmentation. To address these challenges, we propose the Progressive Perception Learning (PPL) Framework, which employs the PPL optimization function to progressively enhance pixel-level perception, semantic alignment, and global understanding combined with three key components: the Clue-Guided Region Restoration Module (CGRRM), Semantic Interaction Refinement Fusion Module (SIRFM), and Global Perception Module (GPM). Specifically, CGRRM improves object localization by guiding the restoration of key regions based on sparse scribble annotations. SIRFM refines the interaction between low-level spatial features and high-level semantic information. GPM captures long-range dependencies across the image, providing a global understanding while preserving local details. Additionally, we design a Feature Alignment Fusion Module (FAFM) to further refine structural alignment. Extensive experiments demonstrate that our method outperforms existing state-of-the-art SCOD approaches and is competitive with several fully-supervised methods. Our code and data are available at: https://github.com/NGI-vision/PPL .