Evaluation of machine learning methods for detection andclassification of objects in IR images

This thesis investigates the effectiveness of deep learning-based object detection models for identifying vehicles in infrared (IR) imagery, with a focus on combat vehicles from an aerial perspective. IR imaging offers advantages in low-visibility conditions but presents challenges such as sparse texture, inconsistent thermal signatures, and limited public datasets. A custom dataset was assembled by combining several public IR image sets, amounting to around 40,500 annotated images, with the split 7/2/1 between training, validation and test sets. The images in the training set were then augmented to increase the quantity and variation of the training images. The study compares traditional and state-of-the-art detectors, including SSD, Faster R-CNN, Mask R-CNN, RetinaNet, YOLOv11, YOLOv12, YOLO-MIF, and YOLO-Phantom. Each model was trained and evaluated using metrics such as precision, recall, F1-score, mean average precision (mAP), and inference speed (FPS). Results show that while YOLO-MIF achieved the highest detection accuracy, YOLO-Phantom offered the best trade-off between speed and performance, making it suitable for deployment in real-time, resource-constrained applications. These findings provide insights into selecting and optimizing object detection models for IR-based aerial surveillance and reconnaissance systems deployed on edge devices.