Thermal imager performance evaluation with recorded infrared images: Triangle Orientation Discrimination (TOD) classifier versus You Only Look Once (YOLO)

Thermal imagers with aided or automatic target recognition (ATR) are becoming increasingly important in military systems. Today, there are no standard performance evaluation techniques for thermal imagers with respect to algorithms, such as ATR, as data users. As a possible option, classifier models for triangle orientation discrimination (TOD) have been proposed. These models are small and fast, thus allow to process large datasets during the lab measurement. In a recent study, TOD classification models have been compared against “You Only Look Once” (YOLO), which is a state-of-the-art framework for object detection, with respect to their performance decrease under camera-related image degradations. Similar performance drops were found by impairing the respective databases by equivalent simulated noise, blur and subsampling. As a follow-up, we present in this paper evaluation results for TOD and YOLO model performance on recorded infrared images. An experimental setup consisting of a high-resolution infrared scene projector is used to project images of respective databases subsequently onto thermal imagers. The recorded images are evaluated by TOD and YOLO models and performance metrics such as precision and accuracy are calculated. Based on these metrics, different imagers are ranked for both models respectively and these rankings are compared. Finally, implications of incorporating recorded image data in the model training are discussed.