Edge-Guided Camouflaged Object Detection via Multi-Level Feature Integration.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Camouflaged object detection (COD) is challenging due to low contrast between objects and surroundings.
• Existing methods require improvement in boundary refinement and multi-level feature extraction/fusion.

Purpose of the Study:

• To propose a novel Multi-Level Feature Integration Network (MFNet) for enhanced camouflaged object detection.
• To address limitations in boundary refinement and feature fusion in current COD models.

Main Methods:

• Designed an Edge Guidance Module (EGM) to integrate high-level semantics and low-level spatial details for edge modeling.
• Developed a Multi-Level Feature Integration Module (MFIM) to fuse local and global features across adjacent levels.
• Introduced a Context Aggregation Refinement Module (CARM) for efficient cross-level feature aggregation and refinement.

Main Results:

• The proposed MFNet model demonstrated superior performance in camouflaged object detection.
• MFNet outperformed state-of-the-art models across four key evaluation metrics (Sα, Eϕ, Fβw, MAE).
• Extensive experiments on three benchmark datasets validated the effectiveness of MFNet.

Conclusions:

• MFNet is an effective model for camouflaged object detection.
• The proposed modules (EGM, MFIM, CARM) significantly contribute to improved COD performance.
• MFNet offers a promising solution for accurate segmentation of camouflaged objects.