Kutistettujen hyperspektrikuvien luokittelija

Convolutional neural networks have been successfully used in previous studies to classify medical hyperspectral images. Hyperspectral images are typically classified using semantic segmentation, where each pixel in the image is given a class based on its spectrum. With the help of semantic segmentation, it is possible to see exactly where there is disease or damage in the tissue. Hyperspectral images can also be classified as a whole, in which case a hyperspectral image is given one class based on its spectral properties. However, no previous research has been done on the classification of entire hyperspectral images. The goal of the research was to find out how hyperspectral images can be classified as a whole instead of semantic segmentation. The material of the study was the previously collected lesion material. The work sought and implemented a neural network architecture for the classification of whole hyperspectral images. In addition, the work investigated how the spatial reduction of hyperspectral images affects the classification accuracy. The neural network performed poorly in the classification of hyperspectral images. The classification accuracy improved when the size of the images was reduced spatially. The study gave indications that when classifying whole hyperspectral images, the spatial size should be small in order to maintain good classification accuracy.