The thesis addresses the problem of mine detection in thermal images. We demonstrate the advantages of thermal images compared to color images for mine detection. We analyze the only publicly available dataset of thermal mine images, highlight its shortcomings, and propose a derived dataset that addresses these issues. We enhance the existing mine detection method using YOLOv8 and then propose a new approach to solving the problem through defect detection using the SegDecNet architecture. We compare the effectiveness of the YOLOv5, YOLOv8, and SegDecNet architectures on the original and derived datasets, showing that our method achieves better results in all cases. We also highlight the limitations of comparing results on the derived dataset and suggest improvements to the dataset and potential extensions of the SegDecNet architecture for object detection.
|
