| 摘 要: 为了解决工厂烟草滤嘴纸盒码垛过程中易发生错误的问题,并提高码垛效率,提出了一种基于改进SGBM(Semi-Global Block Matching)算法与轻量化YOLOv5s网络结构相结合的识别检测与定位一体化方法。首先,采用MobileNetv3作为YOLOv5s的主干网络,以降低模型的复杂度、提高检测速率;其次,将AD-Census(Absolute Differences-Census)代价计算方法应用于SGBM算法中,以提高测距精度。实验结果显示,在自建数据集上,该方法获得的mAP(mean Average Precision)达到91.2%,帧率FPS(Frames Per Second)达到30.8,可以准确识别出纸盒的类别与位置,可为烟草滤嘴盒自动码垛机器人提供视觉系统技术支持。 |
| 关键词: 烟草滤嘴盒;SGBM算法;立体匹配;轻量化YOLOv5 |
|
中图分类号: TP391
文献标识码: A
|
| 基金项目: 国家自然科学基金资助项目(61861007,61640014);贵州省科技计划资助项目(黔科合基础-ZK[2021]一般303);贵州省科技支撑计划资助项目(黔科合支撑[2022]一般264,黔科合支撑[2023]一般096,黔科合支撑[2023]一般412,黔科合支撑[2023]一般409);贵州省教育厅创新群体项目(黔教合KY字[2021]012);中国电力建设股份有限公司科技项目(DJ-ZDXM-2022-44);贵大引进人才项目(贵大人基合字(2014)08号) |
|
| Lightweight YOLOv5-based Detection and Localization Technique for Tobacco Filter Tip Boxes |
|
ZHANG Fang, WANG Xiao, XU Linghua, JIN Qiao
|
(School of Electrical Engineering, Guizhou University, Guiyang 550025, China)
2964151489@qq.com; xwang9@ gzu.edu.cn; lhxua@gzu.edu.cn; qjin168@fomail.com
|
| Abstract: To address the issue of frequent errors during the stacking process of tobacco filter tip boxes in factories and improve stacking efficiency, this study proposes an integrated recognition, detection, and localization method combining an improved SGBM (Semi-Global Block Matching) algorithm with a lightweight YOLOv5s network architecture. Firstly, MobileNetv3 is adopted as the backbone network of YOLOv5s to reduce model complexity and enhance detection speed. Secondly, the AD-Census (Absolute Differences-Census) cost computation method is integrated into the SGBM algorithm to improve ranging accuracy. Experimental results on a self-built dataset demonstrate that the proposed method achieves a mean Average Precision (mAP) of 91.2% and a frame rate of 30.8 FPS (Frames Per Second), enabling accurate identification of box categories and positions. This approach provides technical support for vision systems in automatic stacking robots for tobacco filter tip boxes. |
| Keywords: tobacco filter tip boxes; SGBM algorithm; stereo matching; lightweight YOLOv5 |
