改进YOLOV3算法在零件识别中的应用  

Application of Improved YOLOV3 Algorithm in Part Recognition

作　　者：宋栓军[1,2] 侯中原 王启宇 倪奕棋 黄乾玮 SONG Shuanjun;HOU Zhongyuan;WANG Qiyu;NI Yiqi;HUANG Qianwei(College of Mechanical and Electrical Engineering,Xi′an Polytechnic University,Xi′an 710600,China;Key Laboratory of Modern Intelligent Textile Equipment,Xi′an Polytechnic University,Xi′an 710600,China)

机构地区：[1]西安工程大学机电工程学院,西安710600 [2]西安工程大学现代智能纺织装备重点实验室,西安710600

出　　处：《机械科学与技术》2022年第10期1608-1614,共7页Mechanical Science and Technology for Aerospace Engineering

基　　金：国家自然科学基金青年科学基金项目(61701384);西安市科技计划先进制造业技术攻关项目(21XJZZ0016)。

摘　　要：针对工业生产中小型零件存在漏检及识别率不高等问题,结合深度学习基本理论,提出一种改进YOLOV3网络的零件目标识别算法。该算法首先在零件特征融合结构信息中增加了一个特征尺度,进一步融合深层网络与浅层网络的特征信息,以便更好融合零件的位置信息和语义信息。为克服YOLOV3算法中使用K-means聚类对初值不稳定的缺点,根据不同零件类别宽高比,采用K-means++算法对Anchor框重新进行了聚类。最后,在自制常见的六种零件的数据集上,通过实验对该算法进行了验证。结果表明,所提出的改进算法识别效果优于YOLOV3识别效果,在目标识别检测中具有准确率高的优势。Aiming at the problems of missed detection and low recognition rate of small and medium-sized parts in industrial production, combined with the basic theory of deep learning, an improved YOLO(You Only Look Once)V3 network part target recognition algorithm is proposed. The algorithm first adds a feature scale to the part feature fusion structure information, and further fuses the feature information of the deep network and the shallow network to better integrate the position information and semantic information of the part. In order to overcome the shortcoming of using K-means clustering in YOLOV3 algorithm that the initial value is unstable, according to the aspect ratio of different parts category, K-means++ algorithm is used to re-cluster the Anchor box. Finally, the algorithm was verified by experiments on data sets of the six self-made common parts. The results show that the recognition effect of the improved algorithm proposed is better than that of YOLOV3, and it has the advantage of high accuracy in target recognition and detection.

关 键 词：深度学习 YOLOV3 特征融合 零件识别 

分 类 号：TP273.5[自动化与计算机技术—检测技术与自动化装置]