基于集成深度学习的玻璃缺陷识别方法

Glass Defect Recognition Method Based on Integrated Learning

摘要:

针对玻璃缺陷形态复杂多变,难以准确识别其所属类型的特点,文章提出了一种集成深度学习模型对玻璃缺陷进行识别,该模型本质上是一种稀疏编码分类器与深度卷积神经网络的结合。该模型在自编码器的基础上引进了KL距离和L1范数作为稀疏项,构成新的稀疏自编码器。并在次通过稀疏自编码器学习输入样本特征,将训练好的权值作为卷积神经网络的卷积核从而提高了识别速度。在稀疏编码阶段用L<sub>1</sub>-L<sub>2</sub>范数代替L<sub>0</sub>范数,并在KSVD上添加了判别分类能力使其更好的进行分类运算,以此提高识别准确率。实验结果表明,该方法识别准确率达到了95%,满足了工程上的应用,并有很好的鲁棒性。

Abstract:

In view of the complex shape of glass defect, it is difficult to accurately identify the characteristics of its type. This paper proposes an integrated deep learning model to identify glass defects, which is essentially a combination of sparse coding classifier and deep convolutional neural network. Based on the auto-encoder, the model introduces the KL distance and the L1 norm as sparse terms to form a new sparse auto-encoder. The model learns the features of input sample by sparse auto-encoder and uses the trained weights as the convolution kernel of the convolutional neural network to improve the recognition speed. In the sparse coding stage, L1-L2 norm is used to replace the L0 norm, and the discriminant classification ability is added to the KSVD to make it better to classify operations, thereby improving the recognition accuracy. The experimental results show that the recognition accuracy of this method is up to 95%, which satisfy the need in on-field application and it is robust.