摘要
针对复杂工业巡检场景中小目标因尺度小、特征弱、易受背景干扰而产生的漏检率高和定位精度不足问题,在YOLOv8n网络基础上提出一种小目标检测优化方法。该方法在特征提取阶段引入上下文感知的CoTM模块,以增强局部特征与上下文语义的关联;在特征融合阶段嵌入R-CBAM模块,通过残差注意力机制强化关键区域响应;在回归阶段采用Focaler-IoU损失函数替代传统CIoU,以提升模型对小目标样本的关注能力。实验结果表明,改进模型在自建工业巡检小目标数据集上取得了更优的检测性能,在烟雾、火焰、跑冒滴漏及人员异常行为等检测任务中,相较于原YOLOv8n模型Recall提升5.3%,mAP@50提升7%,具有较好的有效性与鲁棒性。
To address the issue of high miss-detection rates and insufficient localization accuracy of small objects in complex industrial inspection caused by small target scale,weak features,and susceptibility to background interference,we propose an optimized small target detection method based on the YOLOv8n network.This method introduces a context-aware CoTM module in the feature extraction stage to enhance the correlation between local features and contextual semantics.An R-CBAM module is embedded in the feature fusion stage to strengthen responses in critical regions via a residual attention mechanism.Focaler-IoU loss replaces the conventional CIoU in the regression stage to improve the models ability to focus on smallobject samples.Experimental results show that the improved model achieves superior detection performance on a self-built industrial inspection small target dataset,with a 5.3%increase in Recall and a 7%improvement in mAP@50 compared to the original YOLOv8n model,validating the effectiveness and robustness in detection of smoke,flame,leakage,and abnormal human behaviors.
作者
张旭龙
诸德伟
刘爽
杨忠
ZHANG Xulong;ZHU Dewei;LIU Shuang;YANG Zhong(Jinling Institute of Technology,Nanjing 211169,China;Nanjing TetraBOT Electronic Technology Co.Ltd.,Nanjing 210012,China)
出处
《金陵科技学院学报》
2026年第1期8-14,共7页
Journal of Jinling Institute of Technology
基金
江苏省高等学校自然科学基金重大项目(23KJA480002)。
