Performance Evaluation of LWIR Image Detection Using Fine-tuning of YOLOX Model

Owing to the rapid development of artificial intelligence technology, a range of data is being used for training neural networks. For example, studies using images with traditional RGB channels are predominant in the field of deep learning. Furthermore, the number of studies that employ RGB channel data is constantly increasing and they are achieving considerable performance enhancements. However, research on infrared images, including long wave infrared (LWIR), is neglected compared to RGB channel images. In this paper, we focus on LWIR data to evaluate the performance of YOLOX through a fine-tuning technique and confirm the possibility of applying pre-trained weights trained with RGB images to LWIR images. In addition to training the YOLOX model, we construct an LWIR image dataset to evaluate the performance of YOLOX. An experiment was conducted using pre-trained weights trained by RGB channel images and weights trained by our LWIR images. The results indicated clear differences in performance, achieving 3.2% and 54.2% of mean average precision (mAP), respectively. Our study confirmed that it is necessary to perform training through fine-tuning to ensure reliable performance depending on the lens performance, cooling characteristics of the infrared cameras, and wavelength band of the camera. © ICROS 2024.