On Thermal Image Classification with ANN and Similarity Based Classifiers

The objective of the work is to evaluate a set of classifiers based on similarity of patterns, in the context of human thermal images, and for face and emotion recognition. The processing and classification of such images is important for various difficult environments, i.e., media with less lightning, including here the actual pandemic context. The classifiers use various distances to evaluate the similarity among patterns. The classification objective is to recognize the person and the state, from a set of three values, i.e., normal, happy, and sad. Three solutions were considered. The first one uses an artificial neural net, in a standard multilayer feedforward structure. The second one uses the discrete cosine transform (DCT) for feature selection. The last one uses the entropy of the processed images for feature selection. A classifier based on minimum distance is used for the last two solutions. The best results are obtained with DCT based features. The methods could be extended to other types of images, as ultrasound or time-frequency based images.