ZoomingCOD: Zooming-like behavior for Camouflage Object Detection based on Cross-axis Perception and Adaptive Receptive Field

Feng Xiao

Abstract: Camouflaged object detection (COD) plays a critical role in various applications such as wildlife monitoring, and security surveillance, where accurately identifying targets hidden within complex environments is essential. In COD tasks, the large-scale variation of camouflaged objects often leads to situations where traditional multi-scale methods struggle to capture sufficient contextual information for both small and large objects. Additionally, compared to other detection tasks, the boundaries of camouflaged objects are typically more ambiguous and harder to distinguish, further complicating the process. Although existing methods ZoomNet, enhance contextual information by extracting multi-scale information, they fail to adaptively adjust the scale of required context based on the size of camouflaged objects, limiting the effectiveness of the models. To address the issue of significant size and shape variations of camouflaged targets and their varying demands for contextual information, we propose a novel COD method based on the zooming experience of human vision. The model effectively enhances the fuzzy boundary information of the target to be identified through a novel axial perception module and combines the adaptive receptive field selection module to realize a dynamic multi-scale feature adjustment mechanism. We also optimize computational resources with the Mamba state-space model, reducing the global computation through a selective computation mechanism, thus enhancing overall computational efficiency. Through extensive validation on multiple public benchmark datasets and our own custom dataset, our method achieves a 13.5% performance improvement over SOTA wF metrics, demonstrating ZoomingCOD's superior performance in COD tasks.