Hyperspectral fruit and vegetable classification using convolutional neural networks

The classification of different types of fruits and vegetables is a difficult task, since many types are quite similar in color and shape. In this study, we show an easy way to classify hyperspectral images with state of the art convolutional neural networks pre-trained for RGB image data. A small, custom dataset of hyperspectral images was recorded from staged but realistic scenes. With this dataset, an ImageNet pre-trained convolutional neural network was fine-tuned to obtain a classifier. An additional data compression layer has been added to be able to classify the hyperspectral images with the RGB pre-trained network. To isolate the benefit of increased spectral resolution for the classification, the same analysis was also performed with pseudo-RGB images calculated from the hyperspectral images. The results show that the hyperspectral image data increases the average classification accuracy from 88.15% to 92.23%. The approach can easily be extended to other applications.