ANALYSIS OF OPTIMIZER AND HYPERPARAMETER INFLUENCE ON YOLO IN THERMAL LANDMINE DETECTION

This paper investigates the impact of optimizer choice and hyperparameter tuning on the performance of the YOLO deep learning model for landmine detection in thermal images. The aim of this work is to study the effect of different optimizers and parameter configurations on model accuracy and training stability. The object of the study is the process of detecting landmines in thermal imagery using deep neural networks. A dataset of thermal landmine images annotated in YOLO format was used for training. The experiments were conducted with the YOLOv11n architecture initialized with pre-trained weights. The varied parameters included the optimizer (SGD or Adam), learning rate, and batch size. Each model was trained for 50 epochs, and performance was evaluated using mAP, precision, and recall metrics. The study provides a comparative analysis of the influence of Adam and SGD optimizers on the accuracy and stability of YOLO when trained on a limited dataset of thermal landmine images. The results suggest that, given appropriate configuration, SGD is capable of achieving performance competitive with adaptive methods, despite their popularity. The experiments also confirm the feasibility of achieving high detection accuracy even with a relatively small dataset. All configurations achieved high mAP values. The Adam optimizer enabled a faster initial reduction in loss functions, whereas SGD provided smoother and more stable training dynamics. The highest precision and recall were obtained in the experiment with SGD at a learning rate of 0.01 and batch size of 64, making this configuration the most promising for further research. The findings on optimizer and hyperparameter selection can be applied to improve the efficiency of automated thermal image analysis systems based on unmanned aerial vehicles, contributing to safer and faster detection of explosive hazards.