ETD

• Computer Vision
• Deep Learning
• Image Preprocessing
• Object Detection
• Parcel Damage Detection
• YOLO11m

This thesis investigates automated parcel damage detection using deep learning-based object detection. The objective is to evaluate how preprocessing and training configurations affect the performance of a YOLO11m model for detecting and localizing different types of parcel damage. The study uses 1,723 annotated images containing five damage classes: crushed, hole, scratched/cracked, shrunken-corner, and tear/ripped. The images are converted to grayscale to reflect the intended application and to investigate detection performance without color information.
Several experimental configurations are evaluated, including a baseline model, custom crop and augmentation, default YOLO augmentation, optimized augmentation, increased image resolution, and a combination of custom preprocessing with optimized augmentation. A Seg-Fix version of the dataset is also evaluated to examine the effect of dataset preparation on detection performance. The experiments are assessed using precision, recall, mAP@50, and mAP@50--95.
The experimental results are used to determine how the investigated preprocessing, augmentation, image resolution, and dataset configurations influence YOLO11m performance. The study identifies the most effective configuration among the evaluated approaches and demonstrates the potential of YOLO11m for automated parcel damage detection in logistics applications.