基于半监督深度卷积生成对抗网络的注塑瓶表面缺陷检测模型

Injection-molded Bottle Defect Detection Using Semi-supervised Deep Convolutional Generative Adversarial Network

注塑瓶表面缺陷检测是注塑成型工艺流程中的重要环节,但生产中存在缺陷的注塑瓶样本数量相对匮乏,使得应用深度学习算法进行缺陷检测时容易产生过拟合现象。针对上述问题,文中提出并构建一种半监督(Semi-supervised)深度卷积生成对抗网络(Deep Convolutional Generative Adversarial Network,DCGAN)模型。该模型首先使用HSV(Hue Saturation Va-lue)颜色空间转换与大津算法(Otsu)对原始注塑瓶图像进行预处理得到训练集;然后组合学习任务,使得DCGAN的无监督判别器与注塑瓶表面缺陷检测的监督分类器共享卷积层参数,同时修改损失函数,在DCGAN模型的Wasserstein距离中加入交叉熵;最后使用Adam优化器进行模型训练。实验结果表明,该模型能够准确分辨具有缺陷的注塑瓶样本,分类准确率达到98.65%。与传统的机器学习算法以及采用数据增强的卷积神经网络模型相比,所提模型的分类准确率更高,且较好地避免了过拟合现象,能满足注塑瓶生产中表面缺陷的自动检测需求。

Defect detection of injection-molded bottles is an important part of injection molding.Due to the relatively few defective samples in production,the model tends to over-fit when using deep learning algorithm.In order to solve this problem,a defect detection model based on semi-supervised deep convolutional generative adversarial network(DCGAN)is proposed.Firstly,the model preprocesses the original images using HSV color space transformation and Otsu threshold segmentation methods.Then,the learning tasks are combined so that the unsupervised discriminator and the supervised classifier share convolutional parameters.At the same time,the loss function is modified,which consists of cross entropy and wasserstein distance.Finally,the model is fine-tuned using Adam optimizer.The experimental results show that the model can distinguish the defective samples,achieving an accuracy of 98.65%.Compared with traditional machine learning algorithm and CNN model with data augmentation,the proposed model avoids over-fitting.