SPEGNet: Synergistic Perception-Guided Network for Camouflaged Object Detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Camouflaged object detection is challenging due to objects blending with backgrounds.
• Existing methods often use complex, independently added components, leading to computational inefficiency.
• Reduced resolution processing in current methods sacrifices fine details crucial for camouflage detection.

Purpose of the Study:

• To introduce SPEGNet, a novel architecture for synergistic camouflaged object detection.
• To address the fragmentation and computational burden of current detection methods.
• To improve the accuracy and efficiency of segmenting camouflaged objects across various scales.

Main Methods:

• SPEGNet integrates multi-scale features using channel calibration and spatial enhancement.
• The architecture ensures semantic-spatial alignment for emergent boundaries from context-rich representations.
• Progressive refinement with scale-adaptive edge modulation is employed for balanced boundary precision and regional consistency.

Main Results:

• SPEGNet achieved high performance metrics: 0.887 Sα on CAMO, 0.890 on COD10K, and 0.895 on NC4K.
• The model demonstrates real-time inference speed.
• It excels in detecting objects of varying scales, including intricate and pattern-similar ones, and handles occlusion.

Conclusions:

• SPEGNet offers an effective and efficient solution for camouflaged object detection.
• The synergistic design overcomes limitations of fragmented, component-heavy approaches.
• The method provides robust performance across diverse and challenging camouflage scenarios.