SAM2-DFBCNet: A Camouflaged Object Detection Network Based on the Heira Architecture of SAM2

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Segmentation

Background:

• Camouflaged Object Detection (COD) faces challenges like low contrast, complex textures, and blurred boundaries.
• Existing deep learning methods struggle with robust segmentation in these difficult scenarios.

Purpose of the Study:

• To propose SAM2-DFBCNet, a novel network designed to overcome limitations in current Camouflaged Object Detection techniques.
• To enhance the accuracy and robustness of segmenting camouflaged objects in challenging visual environments.

Main Methods:

• Developed SAM2-DFBCNet based on the SAM2 Hiera architecture.
• Incorporated three key modules: Camouflage-Aware Context Enhancement Module (CACEM), Cross-Scale Feature Interaction Bridge (CSFIB), and Dynamic Boundary Refinement Module (DBRM).
• Utilized attention mechanisms and bidirectional convolutional GRU for feature fusion and boundary refinement.

Main Results:

• SAM2-DFBCNet outperformed twenty state-of-the-art methods on CAMO, COD10K, and NC4K datasets.
• Achieved maximum improvements of 7.4% in Sα, 5.78% in Fβ, and 4.78% in Eϕ.
• Reduced Mean Absolute Error (M) by 37.8%.

Conclusions:

• The proposed SAM2-DFBCNet demonstrates superior performance and robustness for Camouflaged Object Detection.
• The network effectively addresses challenges posed by low contrast, complex textures, and blurred boundaries in camouflage scenarios.