Material Classification Using Non-supervised Segmentation and Transfer Learning

This work presents a new way of classifying materials within a scene. In this work, we propose a way to combine machine learning techniques such as unsupervised methods, learning transfer, and supervised methods. We use a clustering technique to segment the areas of interest from the distribution of the pixels in a texture image. Extracting characteristics from the images in conjunction with transfer learning techniques allows us to extract high-level characteristics in interest areas. Interest and use the knowledge previously learned; once we obtain the image’s characteristics, we introduce this input to the convolutional neural network to make it easier to learn these materials.