Object Classification Using Spectral Images and Deep Learning

Spectral images contain valuable information across the electromagnetic spectrum, which provides a useful tool for classification tasks. Most of the traditional machine learning algorithms for spectral images classification such as support vector machine (SVM), k-nearest neighbor, or random forest required complex handcrafted features extraction of the data, in contrast with these approaches deep learning-based methods realize the feature extraction automatically. This paper proposes a procedure to classify spectral images with a Convolutional Neural Network (CNN) approach which consists in the experimental acquisition of two datasets, medicines and honey, the pre-processing of the raw data, the design of the (CNN) and finally the classification results performed by the designed CNN. The results of the first simulation of the proposed CNN-Med show accuracy in the validation set of up to 97.3% for the medicines dataset compared with 94.6% ResNet-18 architecture accuracy and 89.2% AlexNet architecture accuracy. The results of the proposed CNN-Honey show an accuracy, by patches, in the validation set of up to 92.11% for the honey dataset compared with 86.84% ResNet-18 architecture accuracy.