Hyper-spectral image classification for wood recognition


Hyperspectral Image (HSI) classification amounts to classify images that contain a multitude of spectral bands. In the H2I project we have been investigating how Convolutional Neural Networks (CNNs) can be adapted to perform HSI classification. In this lighting talk we present a novel way of viewing the HSI through a simple data format transformation and the new design of the network training strategy. With minor modification for the lightweight CNN based classifier Cifar10, the proposed approach enables the network’s ability to exploit the information between the different spectral bands. The classifier is evaluated extensively, using different strategies, on a dataset for wood recognition. Obtained results in terms of accuracy and training time prove that the proposed approach is lightweight, simple to train, and effective.