|
基于PP-YOLOE的烟草甲虫无锚框小目标检测 |
Anchor-Free Small Object Detection of Tobacco Beetle Based on PP-YOLOE |
投稿时间:2023-12-01 修订日期:2024-02-28 |
DOI: |
中文关键词: 烟草虫检测 PP-YOLOE-SOD 注意力机制 数据增扩 |
英文关键词: Tobacco beetle detection PP-YOLOE-SOD Self-Attention Data augmentation |
基金项目: |
|
摘要点击次数: 73 |
全文下载次数: 0 |
中文摘要: |
烟草产业受到烟草甲虫和烟草粉螟的严重危害。烟草虫体积小,标注数据少,且不易获取,这对于传统的数据驱动的目标检测算法提出挑战。为了实时监测烟草虫的种群和分布情况,本文提出一种基于PP-YOLOE的烟草虫小目标检测方法PP-YOLOE SOD(PP-YOLOE Small Object Detection)。该方法在PP-YOLOE的特征金字塔网络(Feature Pyramid Network,FPN)中引入注意力机制,以更好地关注全局上下文信息和高层语义之间的相互依赖关系,从而实现精确的小目标检测。此外,本文还通过引入回归负半轴来减少回归小目标时的噪声,并通过一系列数据增扩方法对小样本条件下烟草虫小目标的进行精准检测。实验证明,基于PP-YOLOE-SOD的烟草虫检测方法在烟草虫数据集上取得了94.6%的检测准确率,优于目前主流的目标检测方法。 |
英文摘要: |
Tobacco industry is severely threatened by tobacco beetles and tobacco moths. Tobacco beetles are characterized by small size and limited labeled data, which pose challenges to conventional data-driven object detection algorithms. In order to monitor the population and distribution of tobacco beetles in real-time, this paper proposes a small object detection method for tobacco beetles, called PP-YOLOE SOD (PP-YOLOE Small Object Detection), based on PP-YOLOE. We introduce a Self-Attention module into the Feature Pyramid Network (FPN) of PP-YOLOE to better capture global contextual information and dependencies between high-level semantics, thus achieving accurate detection of small objects. Additionally, this paper reduces noise in regressing small objects by introducing a regression negative axis and achieves accurate detection of tobacco beetles in few-shot conditions through a series of data augmentation methods. Experiments demonstrate that the proposed method based on PP-YOLOE-SOD achieves an 94.6% detection accuracy on the tobacco beetle dataset, surpassing mainstream object detection methods. |
View Fulltext
查看/发表评论 下载PDF阅读器 |
关闭 |