Efficient Object Detection and Classification of Heat Emitting Objects from Infrared Images Based on Deep Learning

Abstract

Object detection from infrared (IR) images recently attracted attention of researches. There are several techniques that can be performed on images in order to detect objects. Deep learning is an efficient technique among these techniques as it merges the feature extraction in the classification process. This paper presents a deep-learning-based approach that detects whether the image includes a certain object or not. In addition, it considers the scenario of object classification that has not been given attention in the literature for IR images. The importance of multi-object classification is to maintain the ability to discriminate between objects of interest and trivial or discarded objects in the IR images or image sequences of very poor contrast. The suggested deep learning model is based on Convolutional Neural Networks (CNNs). Two scenarios are included in this study. The first scenario is to detect a single object from an IR image. The second one is to detect multiple objects from IR images. Both scenarios have been studied and simulated at different Signal-to-Noise Ratios (SNR) on self-recoded as well as standard IR images. The proposed scenarios have been tested and validated by comparison with the traditional approach based on Histogram of Gradients (HoG) technique that is popularly considered for object detection. Moreover, a comparison with other state-of-the-art methods is presented. Simulation results reveal that the HoG approach may fail with IR images due to the low contrast of these images, while the proposed approach succeeds and achieves an accuracy level of 100 % in both studied scenarios.