An Efficient Algorithm for Object Detection in Thermal Images using Convolutional Neural Networks and Thermal Signature of the Objects

Abstract

In recent years, thermal imagery techniques have gained even more relevance because of its application and uses in various domains starting from agriculture to electrical engineering and going all the way to military and surveillance to name a few. In this paper, the use of thermal imagery and deep learning techniques for object detection has been explored. The objects chosen are commonly available entities such as: cat, dog, man and a car. The dataset consists of thermal images and objects are detected on the basis of different heat signatures and further these images are used to train a Convolutional Neural Network (CNN) based model. Also, model training is done using idea of transfer learning and pre-trained models to evaluate and compare performance metrics against the exiting Keras based transfer learning algorithms which are utilized here. The best model achieved an average accuracy of 91.94%. The results were also verified against a test dataset.