Structural Highlight Network for Camouflaged Object Detection
Zheng Wang, Junkun Zhao, BiFan Lai, XingHuai Zheng · 2024
Camouflaged targets typically achieve visual integration with their surroundings by adopting appearances similar to the surrounding environment, making them challenging to be easily perceived by the naked eye. This makes the accurate detection and localization of camouflaged targets in images or scenes highly challenging. Inspired by the human tendency to capture subtle structural differences between targets and backgrounds based on structural features, we propose a novel framework called the Structural Highlight Network (SHNet), which consists of three stages: Structural Perception Stage, Structural Feature Enhancement Stage, and Structural Feature Fusion Stage. Specifically, in the Structural Perception Stage, we utilize an attention-based backbone to locate objects. In the Structural Feature Enhancement Stage, to accumulate rich search features, we perform cross-layer feature fusion and enhancement operations on the feature maps, considering aspects such as texture, edges, morphology, and semantics. In the Feature Fusion Stage, we employ a Structural Feature Fusion module to smooth exploration and accomplish structural information interaction, precisely separating highly similar foreground and background.