Classical/modern methods and large language model approach in fever detection from thermal images

In this thesis, various methods were compared for detecting COVID-19 and similar febrile diseases using thermal imaging. The analyses conducted with Haar Cascade (HC), Deep Learning (DL), and Large Language Model (LLM) evaluated the performance of these methods in the fields of thermal imaging, face recognition, and data analysis. The results provided significant findings, particularly in single face detection and the use of thermal images in healthcare applications. The HC method showed successful performance with a 75% accuracy rate in thermal images containing single faces, while this rate dropped below 50% in images with multiple faces. This method demonstrated higher accuracy rates in RGB images, but exhibited limitations in complex thermal images. The potential of the HC method, especially in low contrast and low-resolution images, was noteworthy. Deep learning models stood out in thermal image processing with high accuracy rates. Specifically, a 100% accuracy rate was achieved in thermal images containing single faces. However, the success rate dropped to 0% in images with multiple faces. The model provided effective results in both normal and side-angle images, but revealed the need for further improvements in more complex scenarios. Expanding the dataset and applying transfer learning techniques could enhance the performance of this model. LLM proved to be a powerful tool in interpreting temperature analysis after face detection. The model demonstrated moderate performance with a 62.5% accuracy rate in normal face detection. However, its performance significantly decreased in images with faces captured from different angles or containing multiple faces. Although the natural language processing capabilities of LLM allowed for better analysis of the data obtained from thermal images, face detection still needs to be supported by Haar Cascade or Deep Learning methods. Keywords: Thermal imaging, Deep learning, Haar Cascade, Large Language Model (LLM)