一种基于改进Faster RCNN的易拉罐印刷缺陷检测方法

A Method for Detecting the Printing Defects of Easy Open Can Based on Improved Faster RCNN

针对现有易拉罐印刷缺陷检测方法对小尺度缺陷检测难、漏检率高、检测精度低等问题,本研究提出一种基于改进Faster RCNN的易拉罐印刷缺陷检测方法。首先,以Faster RCNN的检测框架为基础,选取改进的VGG16模型作为特征提取网络,提取出卷积特征图;其次,针对缺陷目标的小尺度结构特点,通过改进RPN网络生成更具表征能力的缺陷目标候选框;最后,对缺陷目标候选区域进行分析,通过数学形态算法将缺陷目标从背景中分割出来,实现对目标区域的缺陷识别和形态提取。实验结果表明,本研究检测方法可准确且完整地提取缺陷目标,在易拉罐罐体印刷缺陷数据集上平均准确率达到94.78%,与现有的目标缺陷检测算法相比,识别性能更优、智能化程度更高,对提升易拉罐智能化生产具有现实意义。

To solve the problems of difficult detection of small-scale defects,high leakage rate and low detection accuracy in existing cans’printing defect detection methods,an improved method basedon Faster RCNN for cans’printing defect detection was proposed in this study.Firstly,the improved VGG16 model was selected as the feature extraction network to extract the convolutional feature map based on the detection framework of Faster RCNN.Secondly,according to the small-scale structure of the defective target,the RPN network was improved to generate more characterization ability of the defective target candidate frame.Finally,the defective target candidate areas were analyzed,the defect target was segmented from the background by the mathematical morphological algorithm to achieve defect identification and morphological extraction of the target region.The experimental results showed that the defect target detection method designed can accurately and completely extract the defect targets,and the average accuracy reached 94.78%on the printed defect data set.Compared with the existing target defect detection algorithm,the detection method in this study has better recognition performance and higher intelligence,which is of practical significance to enhance the intelligent production of cans.