当前位置>主页 > 期刊在线 > 计算机技术 >

计算机技术21年23期

基于改进 YOLOv4 算法的铝材表面缺陷识别方法研究
栾明慧,李松松,李晨,王宇恒,郭忠宇
(大连海洋大学 信息工程学院,辽宁 大连 116023)

摘  要:文章针对铝材表面缺陷识别原始算法精度低与提取突出特征能力弱的问题,提出一种改进的 YOLOv4 算法。首先,为提高对小目标缺陷的检测能力,改进了多尺度预测,增强更浅层的细粒度特征信息融合;其次,对铝材标注数据样本采用K-means 聚类,获取更适合缺陷目标的先验框。实验结果表明,在检测速度基本不变的前提下,改进 YOLOv4 算法的平均精度达到 95.02%,比原始的 YOLOv4 算法提高了 1.42%,比 YOLOv3 提高了 2.34%,比 Faster R-CNN 提高了 11.48%。


关键词:铝材表面缺陷;YOLOv4;多尺度预测;K-means 算法



DOI:10.19850/j.cnki.2096-4706.2021.23.025


基金项目:国家自然科学基金资助项目 (51778104);辽宁省渔业厅资助项目(201723)


中图分类号:TP391.4                                       文献标识码:A                                     文章编号:2096-4706(2021)23-0096-05


Research on Aluminum Surface Defect Identification Methods Based on Improved YOLOv4 Algorithm

LUAN Minghui, LI Songsong, LI Chen, WANG Yuheng, GUO Zhongyu

(School of Information Engineering, Dalian Ocean University, Dalian 116023, China)

Abstract: Aiming at the problems of low accuracy of the original algorithm for aluminum surface defect recognition and weak ability to extract prominent features, an improved YOLOv4 algorithm is proposed in this paper. First, in order to improve the defect detection ability of small targets, improve multi-scale prediction and enhance the fusion of shallower fine-grained feature information; secondly, use K-means clustering for aluminum labeling data sample to obtain priori frame that is more suitable for defect target. The experimental results show that on the premise of basically unchanged detection speed, the average accuracy of the improved YOLOv4 algorithm reaches 95.02%, which is 1.42% higher than the original YOLOv4 algorithm, 2.34% higher than YOLOv3 and 11.48% higher than Faster R-CNN.

Keywords: aluminum surface defect; YOLOv4; multi-scale prediction; K-means algorithm


参考文献:

[1] 彭向前 . 产品表面缺陷在线检测方法研究及系统实现 [D]. 武汉:华中科技大学,2008.

[2] 昊贵芳,徐科,杨朝霖 . 钢板表面质量在线监测技术 [M]. 北京:科学出版社,2010:102.

[3] 吴晓君,唐婷,张林,等 . 基于机器视觉技术的表面缺陷在线检测系统设计 [J]. 自动化与仪表,2016,31(4):72-76.

[4] 徐镪,朱洪锦,范洪辉,等 . 改进的 YOLOv3 网络在钢板表面缺陷检测研究 [J]. 计算机工程与应用,2020,56(16):265-272.

[5] FERGUSON M,AK R,LEE Y T,et al. Automatic localization of casting defects with convolutional neural networks [C]//2017 IEEE International Conference on Big Data (Big Data). Boston:IEEE,2017:1726-1735.

[6] 方叶祥,甘平,陈俐 . 金属表面缺陷检测的改进 YOLOv3 算法研究 [J]. 机械科学与技术,2020,39(9):1390-1394.

[7] BOCHKOVSKIY A,WANG C Y,LIAO H Y M. YOLOv4: Optimal Speed and Accuracy of Object Detection [J/OL].arXiv: 2004.10934 [cs.CV].(2020-04-23).https://arxiv.org/abs/2004.10934.

[8] REDMON J,FARHADI A. YOLO9000:Better,Faster, Stronger [C]//2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR).Honolulu:IEEE,2017:6517-6525.


作者简介:栾明慧(1996.12—),女,满族,辽宁朝阳人,硕士研究生在读,主要研究方向:检测技术与自动化装置;通讯作者: 李松松(1973.10—),女,汉族,辽宁凌源人,教授,硕士生导师, 博士,主要研究方向:超声无损检测技术、无损检测信号及图像处理 技术。