Convolutional Neural Network Based Thermal Image Classification

Abstract

Classification of Thermal Images has been extensively used for its significant applications in many fields. There are many problems with the visible spectrum like object shadows, clothes or the body of a human being matches the background and different lighting conditions. These limitations are overcome by using thermal imaging. Each and every object emits heat (Infrared energy) according to its temperature. Normally the hotter object emits more radiation than the colder one. As all objects have a mostly different temperature so thermal camera detects them and these objects get appear as distinct objects. In the start, thermal imaging was used by the military for detection, recognition, and identification of enemy personnel and equipment. Nowadays it is extensively used in the detection of face, self-driving car, detection of pedestrians and it also has application in the field of environmental work that is monitoring for energy conservation and pollution control. This research presents a novel study for the classification of thermal images using convolutional neural networks (CNN). The research focused on developing a framework that detects multiple thermal objects using CNN. Developed a framework based on deep learning Inception v3 model; work with thermal images that are captured by Seek Thermal and FLIR. For training and testing of the model two datasets are used that include three classes’ cat, car, and man. For the FLIR dataset the highest accuracy achieved is 98.91% and for Seek thermal dataset highest accuracy achieved is 100%. A comparison of the proposed framework with some other CNN models (DenseNet, MobileNet, and YOLOv4), with a customized CNN model and with a conventional model is also presented. The results of the proposed framework and comparison with other models prove that the proposed framework is effective for the classification of thermal images.