基于迁移学习的塑件外观缺陷柔性检测方法

Appearance defects detection method for plastics based on transfer learning

卷积神经网络大数据与尺寸标注少、大数据与强计算之间的矛盾限制了塑件外观检测自动化的发展,迁移学习方法通过共享网络结构、特征参数等方法,可以在样本数量较少的情况下快速训练新的模型,有效缓解上述矛盾。考虑塑件外观缺陷种类繁多,但外观缺陷特征基本一致,基于此,提出了共享模型中低维特征参数的柔性外观缺陷检测方法,该方法首先通过卷积神经网络提取外观缺陷的抽象特征,训练一个目标检测模型,在需要检测类似缺陷时,将该模型最后一层重新初始化后作为预训练模型,获得识别该缺陷特征的经验知识,最后通过少量样本对重新初始化后的模型进行微调,快速训练得到一个新的检测模型。

The contradiction between big data and less labeling, big data and strong calculation in convolutional neural networks limited the development of visual inspection automation. The transfer learning method could quickly alleviate the above contradiction by sharing the network structure, characteristic parameters and other methods, and training the new model quickly in a small sample. Considering the variety of defects on the parts, the common appearance defects were basically the same. Based on this, a flexible appearance defect detection method for low-dimensional characteristic parameters in the shared model was proposed. Firstly, the abstract features of appearance defects were extracted by convolutional neural network, and a target detection model was trained. When similar defects need to be detected, the last layer of the model was re-initialized as a pre-training model,so as to obtain the empirical knowledge of identifying the defects. Finally trained the re-initialized model with a small number of samples, and got a new detection model quickly.