Abstract
Thermal imaging technologies have traditionally spanned various industries, the military, healthcare, and increasingly in autonomous driving applications. Given the recent advancements of thermal imaging cameras, newer applications of these cameras have spawned interest in research of thermography, given its advantage over normal cameras as they are invariant to various lighting conditions even in total darkness. This thesis explores utilizing advances in deep learning techniques in the thermal domain. Thermography in recent years have employed traditional image processing methodologies for tasks such as classification, detection and object tracking. Recent advancements in deep learning has shown its superiority in accuracy, albeit at a cost of higher computing power which is widely available today. The advances in deep learning have also spawned new tasks in computer vision, namely style transfer techniques and image generation to name a few. Traditional deep learning computer vision tasks e.g. classification, detection and segmentation have been trained on widely available datasets which are RGB images in the visible spectrum. This is a common issue in computer vision and also one of the thesis’s focus which looks at domain adaptation techniques to aid in thermography. In light of that problem, this thesis also explores self-supervised learning in which intrinsic features of images are learnt without explicit label supervision. These methods will be presented on currently available data as well as the FLIR thermal dataset which will be the primary thermal dataset used for demonstrating the efficacy of these models.