基于改进YOLOv8注意力机制的烟头检测

Cigarette Butt Detection Based on Improved YOLOv8n Attention Mechanism

摘要: 为了解决公共场所烟头检测任务中烟头目标尺寸小、背景复杂造成传统检测模型漏检率高的问题，本文提出一种改进YOLOv8的烟头检测方法。该算法是在YOLOv8原始网络的基础上加上SE、CBAM、CAA多注意力模块，提高网络的特征提取能力，再根据烟头检测任务的特点提出了一种小目标数据增强方法和自适应损失权重调整。经过改进后的模型，在自建的烟头检测数据集上，平均精度均值mAP50达到了0.773，比原来的YOLOv8提高了2.2个百分点，漏检率由原来的32.1%降低到30.1%，R和mAP50-90分别提升了0.6和0.8个百分点，有效地提高了小目标烟头的检测性能，可以满足公共场所实时禁烟监控的应用要求。

Abstract: To address the high miss rate of traditional detection models caused by the small size of cigarette butt targets and complex backgrounds in public area detection tasks, this paper proposes an improved YOLOv8n-based cigarette butt detection method. The algorithm enhances the original YOLOv8n network by integrating multiple attention modules, including SE, CBAM, and CAA, to improve the network’s feature extraction capability. In addition, a small-object data augmentation strategy and adaptive loss weight adjustment method are proposed based on the specific characteristics of cigarette butt detection. The improved model achieved an average precision (mAP50) of 0.773 on our self-built cigarette butt detection dataset, which is 2.2 percentage points higher than the original YOLOv8. The missed detection rate dropped from 32.1% to 30.1%, and R and mAP50-90 increased by 0.6 and 0.8 percentage points respectively. This effectively boosts the detection performance for small cigarette butts, making it suitable for real-time smoking ban monitoring in public places.