Increasing the Model Classification Accuracy of Thermal Images

Temperature measurement sensors are today found in numerous applications, from the automotive industry, surveillance, navigation, fire detection, and rescue missions to medicine. Some of the applications utilize a completely automated approach where the system has to detect the presence of the object on the thermal image without additional user input or tweaking. A high level of automation is enabled by specialized deep machine learning models that analyze camera data as input and provide localization and classification of the objects in the scene. This work describes the development of a lightweight backbone deep learning model for the classification of the objects in the thermal image, with the final goal to be used on computationally limited embedded systems. The model is firstly trained on the publicly available Teledyne FLIR Thermal Dataset, and then using transfer learning retrained on dataset specifically crafted for the in-house developed thermal module based on Flir Lepton. A proposed lightweight model with 9MB in size is based on Resnet14, and trained from scratch on the HPC system providing 96.4% accuracy on test dataset.