TEA-DETR: A Texture-Edge Augmented Framework for Military Camouflaged People Detection

Military Camouflaged People Detection (MCPD) aims to detect military personnel in natural environments, facing challenges from indistinct boundaries, deceptive textures, and limited high-quality datasets. Existing camouflaged object detection (COD) methods struggle with edge blur issues when distinguishing camouflaged soldiers from complex backgrounds. We propose Texture-Edge Augmented DEtection TRansformer (TEA-DETR) to address edge ambiguity through three innovations: (1) a multi-scale edge enhancement backbone extracting multi-resolution edge information with reduced computational overhead, (2) a Gabor-MaxPool Texture Enhancement Module utilizing Gabor filters' orientation selectivity and adaptive pooling for discriminative texture extraction while suppressing background noise, and (3) a feature fusion module integrating edge and texture features for comprehensive representations enhancing boundary attention. We develop a high-quality MCPD dataset providing realistic military scenario benchmarks. Experimental results show TEA-DETR achieves superior performance with 97.5\% mAP50 on CPDD and 92.4\% mAP50 on HCPD, significantly outperforming existing object detectors.