Welding Defect Detection

About Model

This model is a YOLOv8-based object detection model developed using the Aargus DIY Visual Inspection Tool for automated visual inspection of welding operations. The model detects and classifies welding regions into three categories: Bad Weld, Good Weld, and Defect. It uses the YOLOv8m architecture and transfer learning from pretrained weights, followed by task-specific training on a labelled welding inspection dataset. The development workflow includes dataset ingestion in YOLO format, image preprocessing and augmentation, model training, hyperparameter experimentation, validation and performance evaluation. Augmentation techniques include mosaic, blur, median blur, grayscale conversion and CLAHE. The trained model can identify welding regions and provide bounding boxes, class labels and confidence scores for detected objects. This enables automated visual inspection and can support quality-control workflows in manufacturing and industrial environments. On the reported held-out test set containing 126 images and 301 annotated instances, the model achieved 0.736 precision, 0.699 recall, 0.717 F1-score, 0.734 mAP@50 and 0.535 mAP@50–95. The model is intended as an AI-assisted visual inspection system and should be validated against deployment-specific data before being used for production quality or safety decisions.