基于PHAM-YOLO网络的卷烟纸燃烧线检测方法

PHAM-YOLO Network-based Detection Method of Combustion Line of Cigarette Paper

关键词:  卷烟纸  燃烧线检测  YOLO  并行混合注意机制

Key Words:cigarette paper  combustion line detection  YOLO  parallel hybrid attention mechanism

摘要:为实现卷烟纸燃烧时燃烧线的准确识别，构建了常见应用场景下的卷烟纸燃烧线数据集。针对检测背景复杂、多目标、燃烧线尺度不一且形状各异的难题，将并行混合注意机制嵌入了YOLO v5主干网络，构建了PHAM-YOLO网络模型用于卷烟纸燃烧线的检测。采用特征金字塔快速池化、边界盒回归等方法提升了卷烟纸燃烧线的定位准确性。结果表明，对于卷烟纸燃烧线数据集， PHAM-YOLO网络检测平均精度均值、精度和召回率分别为99.0%、99.8%和99.0%，其中平均精度均值比原始模型提高了5.0%，高于其他类型的目标检测方法。

Abstract:To determine the combustion line of cigarette paper， the dataset for cigarette combustion line detection was construct from common scenarios. To address the challenges of detecting complex backgrounds， multiple targets， varying scales， and shapes of combustion lines， a PHAM （parallel hybrid attention mechanism） was embedded into the YOLO v5 （you only look once， version 5） backbone network， and PHAM-YOLO was constructed for detecting multiple targets with varying scales and shapes in complex backgrounds. In addition， a spatial pyramid pooling fast （SPPF）， the boundary box regression （BBR） module were introduced to improve the accuracy of combustion line positioning. The results showed that the proposed PHAM-YOLO network achieved the average precision mean （mAP）， precision（P） and recall （R） of 99.0%， 99.8%， and 99.0%， respectively， where mAP was improved by 5.0% compared to the original model and higher than other types of target detection methods.