.png){kind=icon}
.png){kind=icon}
Return
Multi-task information propagation network for camouflaged object detection
.png){kind=icon}
PRE
AI
DOI：10.1016/j.dsp.2025.105172.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Camouflaged object detection (COD) aims to discover camouflaged objects embedded in the background. However, most existing COD methods focus solely on extracting single-task features for inference, and rely on relatively simple supervision data, which hampers their generalization ability in real-world deployment. To meet these challenges, we design a novel camouflaged object detection framework, the Multi-Task Information Propagation Network (MIP-Net). In the feature extraction stage, the Multi-Task Feature Extraction Module (MFEM) learns multi-task feature representations that preserve boundary, texture, and object information to enhance COD performance. During the feature fusion encoding stage, the Cross-Level Multi-Task Fusion Propagation (CMFP) module effectively integrates multi-task features from different levels, adaptively weighting the encoder's features so that the decoder receives more effective information for improved COD outcomes. In the feature decoding inference stage, the Cross-Layer Progressive Refinement Module (CSRM) captures the complementarity of cross-level features, further refining the final inference results. We conducted comprehensive experiments on four benchmark datasets: CHAMELEON, CAMO, COD10K, and NC4K. The comprehensive results prove that the proposed MIP-Net framework outperforms existing methods, showcasing its effectiveness and robustness.
Keywords:
Camouflaged object detection
Swin transformer
Multi-task learning
Feature fusion
Feature propagation