基于深度卷积神经网络及迁移学习的纸病分类方法研究

Paper Defects Classification Based on Deep Convolution Neural Network and Transfer Learning

本课题结合迁移学习策略,设计了一种适用于纸病图像的小样本深度卷积神经网络分类器。首先冻结VGG16网络卷积层的前7层卷积层参数,微调后面的卷积层,完成纸病特征的提取;其次改进用于分类的全连接层,使其满足纸病分类的需求;最后在训练过程中采用迁移学习策略,提高训练效率。结果表明,该方法能够提高纸病识别效率及精度,并进一步加强纸病识别功能。

In this study,combined with the transfer learning strategy,a small sample deep convolution neural network classifier suitable for paper disease images was designed.Firstly,the parameters of first 7 convolution layers of the VGG16 network convolution layers were frozen,and the following convolution layers was fine-tuned to complete the feature extraction of paper defects.Secondly,the fully connection layers for classification were improved to meet the needs of paper defects classification.Finally,transfer learning strategy was adopted in the training process to improve the efficiency.The results demonstrat that this method could improve the efficiency and accuracy of paper defects recognition,and further strengthen the paper defects recognition function.