基于FasterR-CNN改进的数粒机系统-包装工程.PDF

第39 卷 第9 期 包 装 工 程 2018 年5 月 PACKAGING ENGINEERING ·141 · 机械与过程控制 基于 Faster R-CNN 改进的数粒机系统 胡安翔，李振华 （山东大学控制科学与工程学院，济南 250061 ） 摘要： 目的 解决目前数粒机只能计数不能同时分拣残损药粒的问题。方法 设计以 Faster R-CNN 深度 神经网络为核心的药粒数粒机系统。在原有的数粒机基础之上，更换 CCD 线阵相机为面阵相机，以满 足图像采集的需求，进一步使用图像分割和多线程技术加快图像处理速度。最终通过训练好的 Faster R-CNN 网络检测出目标并分拣。结果 经过测试集的验证，正常药粒识别率达到了95.47% ，残损药粒识 别率达到了97.94% ，单幅图像处理达到了65 ms 的实时速度。结论 该方法在传统的计数基础上很好地 融合了先进的深度学习技术，实现了目标的自动分拣。 关键词：数粒机；面阵相机；Faster R-CNN ；药粒分拣 中图分类号：TB486 ；TP391 文献标识码：A 文章编号： 1001-3563(2018)09-0141-05 DOI ：10.19554/j.cnki.1001-3563.2018.09.025 Improved Capsule Counting Machine Based on Faster R-CNN HU An-xiang , LI Zhen-hua (School of Control Science and Engineering, Shandong University, Jinan 250061, China) ABSTRACT: The work aims to solve the problem that the current capsule counting machine can only count capsules and cannot sort damaged capsules at the same time. A capsule counting machine with the Faster R-CNN deep neural net- work as the core was designed. On the basis of the original capsule counting machine, the CCD line-array camera was re- placed by area-array camera to meet the demand of image acquisition, and the image segmentation and multi-thread technology were further used to speed up the image processing speed. Finally, the target was detected and sorted through the well trained Faster R-CNN network. After verification of the test set, the identification rate of normal capsule reached 95.47%, the identification rate of damaged capsule reached 97.94%, and the single image processing reached the real-time speed of 65 ms. The proposed method prop