Scholarlink

<p>Deploying high-precision polyp detectors on endoscopic hardware remains hindered by resource constraints of embedded systems. To resolve these barriers, this study introduces G-YOLO, a framework optimized for efficient colorectal polyp detection. To explicitly handle specular reflections and weak mucosal boundaries, we design an Asymmetric Feature Reconstruction Module (AFRM) utilizing a dual-pathway architecture with asymmetric convolutions to independently extract global context and fine-grained edges, suppressing illumination artifacts without heavy computational overhead. Additionally, the backbone is restructured with efficient Ghost modules to eliminate redundancy, while the Convolutional Block Attention Module is sequentially integrated to optimize channel-spatial weight hierarchies. Evaluated across three benchmarks, G-YOLO successfully bridges the gap between lightweight topologies and diagnostic precision. Compared to the standard YOLOv5s baseline, our network constrains the calculation scale to a lean 11.8 GFLOPs while accelerating inference velocity up to an ultra-fast 378 frames per second (FPS) under server processing and maintaining 8.4 FPS on a low-power edge platform. Concurrently, G-YOLO sustains superior accuracy with an mAP@0.5 of 96.9% on CVC-ClinicDB and a leading Recall of 80.2% on the challenging ETIS-Larib sequence, confirming its exceptional deployment potential for edge-based diagnostic devices.</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

G-YOLO with Asymmetric Feature Reconstruction for Efficient Colorectal Polyp Detection

  • 期刊论文

  • 发表日期：

    2026年6月26日

查看全文
  • Xiaxiang Su,
  • Xin Lu,
  • Rongjiang Tang,
  • Fengli Shen

摘要

Deploying high-precision polyp detectors on endoscopic hardware remains hindered by resource constraints of embedded systems. To resolve these barriers, this study introduces G-YOLO, a framework optimized for efficient colorectal polyp detection. To explicitly handle specular reflections and weak mucosal boundaries, we design an Asymmetric Feature Reconstruction Module (AFRM) utilizing a dual-pathway architecture with asymmetric convolutions to independently extract global context and fine-grained edges, suppressing illumination artifacts without heavy computational overhead. Additionally, the backbone is restructured with efficient Ghost modules to eliminate redundancy, while the Convolutional Block Attention Module is sequentially integrated to optimize channel-spatial weight hierarchies. Evaluated across three benchmarks, G-YOLO successfully bridges the gap between lightweight topologies and diagnostic precision. Compared to the standard YOLOv5s baseline, our network constrains the calculation scale to a lean 11.8 GFLOPs while accelerating inference velocity up to an ultra-fast 378 frames per second (FPS) under server processing and maintaining 8.4 FPS on a low-power edge platform. Concurrently, G-YOLO sustains superior accuracy with an mAP@0.5 of 96.9% on CVC-ClinicDB and a leading Recall of 80.2% on the challenging ETIS-Larib sequence, confirming its exceptional deployment potential for edge-based diagnostic devices.