Computer Vision

Airport X-ray Prohibited-Item Detection

Fine-tuned YOLOv8m on the OPIXray benchmark (8,885 images, 5 knife-type classes across 3 occlusion levels). Evaluated on 1,776 held-out test images. Recall-first metric choice for a safety-critical setting. Includes a controlled YOLOv8s vs YOLOv8m ablation, ONNX export (opset 20), resolution-agnostic inference API, and Docker container.

Key figures

0.924 mAP50

1,776 Test images

5 Classes

What it does

— mAP50 of 0.924; recall ≥0.80 on 4 of 5 classes on the held-out 1,776-image test set
— Ablation: YOLOv8s → YOLOv8m gave +6.6 mAP50, concentrated on the rarest/most-occluded class (+15 AP50)
— Class imbalance handled with mosaic, copy-paste, and mixup augmentation
— ONNX export (opset 20); inference API with schema validation, failure handling, structured logging
— Recall-first evaluation for safety-critical context; GDPR-compliant image retention policy documented