Learning Deep Spatial-Spectral Features for Material Segmentation in Hyperspectral Images

In this study, we propose a two-stage method for material segmentation in hyperspectral images. The first stage employs a Convolutional Neural Network (CNN) to predict the material label at individual pixels. The second stage further refines the segmentation by a fully-connected Conditional Random Field (CRF) framework. For the first stage, we experimented with two different network architectures. One is a network with five convolutional layers and three fully connected layers trained on small patches to predict the label of the central pixel of each patch. The other is an encoder-decoder architecture trained on larger image regions to predict the label of every pixel in a region. In the fully connected CRF, the unary term is aimed to respect the predicted label by the CNN while the pairwise term models the label compatibility between two pixels based on their PCA features. Experimental results demonstrate that the two proposed variants are able to outperform several existing methods quantitatively.