Scholarlink

<p>Detecting small objects in UAV-based maritime surveillance is challenging due to extreme target scales, dynamic sea clutter, and stringent onboard computing budgets. This paper presents DEIM-SCCH, a lightweight detector that jointly optimizes feature enhancement, cross-scale fusion, and frequency-preserving downsampling for real-time small-object perception. The detector incorporates three specialized design modules: a Spatial–Channel Enhancement Unit (SCEU) that couples anisotropic directional filtering with adaptive channel attention to amplify weak targets against cluttered backgrounds; a C3k2 dual-branch interaction module that replaces conventional top-down fusion with split–transform–merge pathways, improving multi-scale feature aggregation without excessive parameters; and HWDown, a Haar wavelet-based downsampling operator that decomposes feature maps into frequency subbands before spatial reduction, retaining edge and texture details that stride-based subsampling discards. On the SeaDronesSee, DEIM-SCCH achieves 84.0% AP@0.5 and 53.4% AP@0.5−0.95 with only 4.42M parameters and 7.9 GFLOPs, outperforming the DEIM (DETR with Improved Matching for Fast Convergence) baseline by 4.5 percentage points AP@0.5−0.95 and all YOLO nano variants with comparable computational cost. Deployed on a Jetson Orin Nano under TensorRT FP16, it runs at 51.72 FPS with 8.65 FPS/W energy efficiency, indicating its suitability for real-time inference.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deim-scch: an efficient lightweight detector for real-time small object perception

  • 期刊论文

  • 发表日期：

    2026年7月12日

查看全文
  • Yanqin Zhao,
  • Xinyi Tian

摘要

Detecting small objects in UAV-based maritime surveillance is challenging due to extreme target scales, dynamic sea clutter, and stringent onboard computing budgets. This paper presents DEIM-SCCH, a lightweight detector that jointly optimizes feature enhancement, cross-scale fusion, and frequency-preserving downsampling for real-time small-object perception. The detector incorporates three specialized design modules: a Spatial–Channel Enhancement Unit (SCEU) that couples anisotropic directional filtering with adaptive channel attention to amplify weak targets against cluttered backgrounds; a C3k2 dual-branch interaction module that replaces conventional top-down fusion with split–transform–merge pathways, improving multi-scale feature aggregation without excessive parameters; and HWDown, a Haar wavelet-based downsampling operator that decomposes feature maps into frequency subbands before spatial reduction, retaining edge and texture details that stride-based subsampling discards. On the SeaDronesSee, DEIM-SCCH achieves 84.0% AP@0.5 and 53.4% AP@0.5−0.95 with only 4.42M parameters and 7.9 GFLOPs, outperforming the DEIM (DETR with Improved Matching for Fast Convergence) baseline by 4.5 percentage points AP@0.5−0.95 and all YOLO nano variants with comparable computational cost. Deployed on a Jetson Orin Nano under TensorRT FP16, it runs at 51.72 FPS with 8.65 FPS/W energy efficiency, indicating its suitability for real-time inference.