A Comparison of the VGG19, InceptionV3, NASNetMobile, and ResNet50 Architectures for Object Classification in Thermal Images

Identifying objects within thermal images is of paramount importance across diverse applications, employing sophisticated machine learning algorithms for precise object localization. Thermal imaging cameras excel at capturing the infrared radiation emitted by objects, enabling visibility through obstacles such as smoke, fog, and darkness. Navigating through thermal images to discern objects mirrors the challenges encountered in traditional visual image analysis. This paper focuses on the development of a convolutional neural network model designed specifically to address multiple classification challenges using contrast-enhanced thermal images processed with CLAHE (Contrast Limited Adaptive Histogram Equalization). The precision, F1-score, and recall of popular architectures, namely VGG19, ResNet50, InceptionV3, and NASNetMobile, are rigorously evaluated on a curated selection of contrast-enhanced CLAHE images. The findings reveal varying levels of accuracy across these models, with VGG19 achieving a notable 97%, InceptionV3 and NASNetMobile at 95%, and ResNet50 registering an accuracy of 94%. This research present significant insights into the utilization of transfer learning methodologies for the classification of thermal imagery.