基于改进YOLO11的小目标检测

打开文本图片集

中图分类号：TP391.41 文献标识码：A

YOLO11-based Algorithm for Small Object Detection with Perspective Rectification

LIU Yong，BAI Mengfei， LI Xinxin， YAN Mingtao，FU Chaoxing （Collge of Mechanical and Electrical Engineering，Qingdao University，Qingdao ，China）

Abstract： Aiming at the insufficient positioning accuracy of tiny targets in the robotic arm workspace under perspective distortion， scale variation， and occlusion interference， a lightweight rotational detection framework fused with Gold-YOLO is proposed to achieve fast and accurate perception of tiny objects in the workspace. Based on YOLOll-obb，the framework introduces the Gather-and-Distribute mechanism of Gold-YOLO to reconstruct the Neck module，while maintaining consistent parameters （2.7M） ，input resolution （ 640×640） ，and training configuration （400 epochs on the RTX-307O platform）. Experimental results on a self-built dataset of 900 images demonstrate that the improved model achieves a mAP50 of 95.8% ， representing increases of 0.2% compared with the original YOLOll-obb. The single-frame inference time is less than 76ms ，and the recall rate for occluded targets is enhanced from 95.3% to 95.8% . With high precision and real-time performance，the proposed framework provides a reliable visual basis for robotic arm path planning and adaptive grasping.

Keywords： rotation detection； YOLOll-obb; recall rate; adaptive grasping

近年来，智能制造对机械臂的自主决策和精细操作提出了更高要求，视觉感知成为机械臂获取复杂环境信息的核心手段[1-2]。（剩余7414字）