Ultralytics YOLO26: PCB defect training and TensorRT export

This tutorial uses Ultralytics Platform to train a YOLO26 model for PCB defect detection, then covers predictions, model export and hosted deployment. The dataset contains almost 25,000 images and about 53,000 annotations across six defect classes, including missing holes, mouse bites and spurious copper. The presenter reviews class distribution, dataset splits and annotated examples before starting training through the platform interface.

The demonstrated run uses 30 epochs and roughly 20,000 training images. The speaker reports about 30 minutes of training and approximately 97% precision and recall. These are results from this run, rather than guarantees for other inspection systems. The walkthrough also examines loss curves and a confusion matrix, then suggests using a vision-language model to verify lower-confidence detections.

For local AI deployment, the export workflow includes TensorRT, a target GPU setting and a downloadable .pt model. The presenter recommends edge processing for fast production lines. A separate hosted CPU deployment exposes an API with Python, JavaScript and curl examples; requests require an endpoint URL and API key. Responses include defect classes, confidence scores and locations. The tutorial demonstrates the export controls, but does not walk through installing or running the exported model on an edge device.