基于PP-PicoDet的半自动标注烟丝异物检测研究

【目的】应用半自动标注技术和轻量级目标检测模型，实现烟丝异物实时检测模型的快速开发和落地应用。【方法】在少量标注的样本上使用picodet-s-320训练基础模型，基于该模型生成对剩余数据样本的预测标注，人工矫正预测标注结果I后，.使用picodet-l-640训练基于全量数据的最终应用模型。【结果】数据标注效率提升223%，迭代周期缩短到1d，训练模型m APoU=o5达到1.000，测试集上漏检率为0%、误检率0.3%，部署模型推理速度提高291%，试运行期间无漏检，多检误报控制在0.23次/批。【结论】采用基于PP-PicoDet的半自动标注技术，大幅缩短算法的开发和迭代周期，模型的预测速度、精度、通用性好，可实现制丝工序烟丝异物的实时检测处置。

Semi-automatic annotation technology for foreign materials detection in cut tobacco based on PP-PicoDet

  Objective  This study aims to achieve rapid development and application of real-time detection model of cut tobacco foreign materials by using semi-automatic annotation technology and lightweight object detection model.  Methods  The base model was trained using picodet-s-320 on a small labeled samples, predictive annotations on the remaining data samples were generated based on the model. After manually correcting the prediction annotation results, picodet-l-640 was used to train a final application model based on the full amount of data.  Results  The data annotation efficiency increased by 223%, the iteration cycle was shortened to 1 day, trained model $ {mathrm{m}mathrm{A}mathrm{P}}_{IoU=0.5} $ reached 1.000. The missed detection rate on the test dataset was 0%, false detection rate was 0.3%, the inference speed of the deployed model was increased by 291%, there was no miss detection during test run, and the multichecking and-false alarm rate was controlled at 0.23 per batch.  Conclusion  Based on the semi-automatic annotation technology of PP-PicoDet, the development and iteration cycle of the algorithm is greatly shortened. It also has solid prediction speed, accuracy and versatility, which can realize the real-time detection and disposal of cut tobacco foreign materials during cigarette processing