Material classification technology based on Convolutional neural　networks

摘要

The contact measurement techniques are typically used in the field of object material classification. It has a lot ofdisadvantages, such as the complex operation and time-consuming. In this paper, a new non-contact object materialidentification method based on Convolutional neural networks (CNNs) and polarization imaging is proposed. Firstly, therelationship between the complex refractive index of object and the polarization information is simulated, and then thestructure of the CNNs is constructed according to the specific conditions of the polarization imaging system. Theaccuracy of the identification method is measured by repeated test using 7 materials. The experimental results show thatthe CNNs model can quickly realize the object material classification with the polarization images, and the classificationaccuracy is above 92%.