小样本深度学习方法实现LED TV屏缺陷检测

Defect inspection for LED TV screen using deep learning method on small sample datasets

摘要:

为实现当前工业4.0时代电子类企业智能制造的全过程，引入机器视觉完成产品缺陷检测，用于解决缺陷问题多样性在算法能力的不足。首先对已标注小样本数据集通过深度学习得到初始特征模型，接着针对该特征模型施以迁移学习方法用以实现LED TV的检测，并将已检测样本进一步用于增量学习完成模型参数的修正，最后采用全连接神经网络FCNet (Fully Connected Neural Network)完成分类，探讨了一种运用机器视觉实现LED TV的光学屏检技术；并给出了检测样品作为补充的样本数据集增量学习模型。实践表明，本文提出的方法能进一步提升工业机器人智能制造阶段自动化检测的准确率，最终实现工业生产的柔性和智能化水平，并为机器视觉的应用提供示范。

Abstract:

In order to better implement the whole process of intelligent manufacturing for electronic enterprises in industry 4.0 era, the machine vision is introduced to address the diversity problem of deficiency on defects in the ability of algorithms. Firstly, through depth learning for the small labeled sample datasets a feature model is obtained. After that, we train the feature model for implement the LED screen inspection by transfer learning. Meanwhile, using incremental learning the parameters of model are corrected step by step. Finally, the FCNet (Fully Connected Neural Network) is used to implement the classification. This paper discussed machine vision to complete LED TV screen detection, and the incremental learning model of the sample datasets is given as a continuous supplement. Many experiments show that the deep learning could further improve the accuracy on automatic inspection, also enhance the flexibility and intelligence level on industrial production, and expand the application of machine vision to provide demonstration.