为了实现对工业产品乳液泵的缺陷检测, 本文采集泵顶、泵上端、泵下端、尾管4个角度的样本图像, 并基于深度学习中迁移学习和卷积神经网络的原理分别构建各角度的分类模型以检测缺陷样本。首先, 使用Mini-ImageNet数据集预训练网络模型。然后, 调整模型结构并加载预训练网络的参数, 并将乳液泵各角度的训练集和验证集经过图像预处理算法后输入至卷积神经网络中训练, 根据训练过程中验证集准确率的变化调整网络超参数, 得到最终网络模型。最后, 将预处理后的乳液泵测试样本输入至训练好的模型中, 检测最终模型的缺陷识别效果。最终4个角度检测准确率均在93%以上, 单样本检测用时为2.52 s, 优于传统方法。本文算法可用于设计乳液泵缺陷检测系统, 该系统能与工业结构相结合筛选出有缺陷的泵体, 也可拓展到工业其他物件的缺陷检测。

In order to realize the defect detection of lotion pump in the industrial production, the sample images at four angles: the top section of pump, the upper section of pump, the lower section of pump and the tail tube of pump were collected. The classification models of each angle were also constructed respectively to detect defect samples at various angles based on the principle of migration learning and convolutional neural network in deep learning. Firstly, the network model was pre-trained by the Mini-ImageNet data set. Then, the model structure was adjusted and the parameters of the pre-training network were loaded. The training set and validation set of each part of the lotion pump which were processed with the image pre-processing algorithm were input into the convolutional neural network for training. According to the accuracy of the verification set during the training process, the network hyperparameters were adjusted to get the final network model. Finally, the preprocessed lotion pump test samples were input into the trained model to detect the defect recognition effect of the final model. The accuracy of the final four angles of defect detection is over 93%, and the single sample detection time is 2.52 s. The algorithm is superior to the traditional algorithm. The algorithm can be used to design the lotion pump defect detection system which can combine with the industrial structure to screen out the defective pump body. It can also be extended to the defect detection of other industrial objects.