基于YOLOv5和生成对抗网络的塑料标签缺陷检测

塑料标签物的缺陷检测与识别是工业过程控制和质量控制的关键；为了克服现有塑料标签缺陷检测方法的局限性,使用了单阶段目标检测模型YOLOv5对其瑕疵进行实时检测与分类；此外,为解决由于样本缺陷数量不足造成的模型识别准确率低等问题,采用了一种基于Defect-GAN的生成对抗网络对小样本进行数据增强和扩增；该方法通过模拟缺陷生成和缺陷图像重建的过程,可以高效合成大量具有高保真度和多样性的缺陷样本,尤其适用于形状不规则、分布随机且尺寸不同的瑕疵生成；实验结果表明,通过使用扩增数据集训练目标检测器,并对网络的超参数进行优化,可以显著提高目标检测器的准确率和精度,其平均精度mAP可达99.5%；此外,为了模拟该方法在实际生产中的应用场景,设计并定制了一台半自动的图像采集机械平台用于采集圆柱样品表面的印刷标签,以及一个自主开发的图像处理和统计分析软件用于样本采集、图像处理及统计分析；该方法和平台可以很容易地推广并应用到其他工业质量控制和缺陷检测系统中。

Industrial Defect Detection of Plastic Labels Based on YOLOv5 and Generative Adversarial Networks

The defect detection and identification of plastic label is the key of industrial process control and quality control. In order to overcome the limitations of the plastic label defect detection methods, the single-stage target detection model YOLOv5 was used to detect and classify the defects in real time. In addition, in order to solve the problem of low model recognition accuracy due to insufficient number of sample defects, a Defect-GAN based generative adversarial networks is used for data augmentation and amplifying the data of small samples. By simulating the process of defect generation and defect image reconstruction, the method can efficiently synthesize a large number of defect samples with high fidelity and diversity. It is especially suitable for defect generation with irregular shape, random distribution and different size. The experimental results show that the accuracy of the detector can be significantly improved by using the amplified data set to train the detector and optimizing the network hyperparameters, and the mean Average Precision(mAP) of the detector can reach 99.5%. In addition, in order to simulate the application scenario of this method in actual production, a semi-automatic image acquisition machine platform was designed and customized for collecting printed labels on the surface of cylindrical samples, and a self-developed image processing and statistical analysis software was used for sample collection, image processing and statistical analysis. The method and platform can be easily extended and applied to other industrial quality control and defect detection systems.