基于目标检测的药品外观识别

Drug Appearance Recognition Based on Object Detection

关键词:  目标检测  YOLOv4  药品外观识别  GhostNet  注意力机制  双向特征金字塔网络

Key words:  target detection  YOLOv4  drug appearance recognition  GhostNet  coordinate attention  Bi-directional Feature Pyramid Network

摘要:

目的 在药品调剂过程中，利用计算机视觉技术识别药品容易受到光照、角度以及包装等因素的影响，会产生较大的识别误差。因此，本文提出了一种用于药品外观识别的目标检测算法(YOLOv4-GhostNet-CMB)。方法 首先，该算法使用GhostNet结构重新设计 YOLOv4的骨干特征提取网络；其次，在Ghost 模块中融合CA注意力机制，沿着水平和垂直方向聚合特征，增强模型对药品的精确定位能力；最后，通过引入Bi-FPN 特征金字塔结构与新主干相连，并新增了一个特征图输出，加强特征的提取，增强药品的识别率。结果 YOLOv4-GhostNet-CMB 算法平均准确率可达到92.31%，与YOLOv4算法相比提升了4.49%。结论 本方法能够有效识别药品，且模型大小仅有150 M。

Abstract:

OBJECTIVE In the process of drug dispensing, using computer vision technology to identify drugs is vulnerable to the influence of lighting, angle, packaging and other factors, which will produce large identification errors. Therefore, this paper proposes an object detection algorithm for drug appearance recognition(YOLOv4-GhostNet-CMB). METHODS Firstly, the algorithm redesigned the backbone feature extraction network in YOLOv4 by using GhostNet. Secondly, the CA attention model was brought into the Ghost module, aggregate features along horizontal and vertical directions to enhance the precise positioning of drugs. Finally, Bi-FPN feature pyramid structure was introduced to connect with the new backbone, and added a feature graph output which could enhance feature extraction and improved the detection accuracy of drugs. RESULTS The experimental results show that the average detection accuracy of YOLOv4-GhostNet-CMB algorithm reached 92.24%, which was a significant improvement of 4.49% compared with YOLOv4 algorithm in term of detection accuracy. CONCLUSION The model size is only 150 MB, nd this algorithm can effectively identify drugs.