Weld Defect Detection from X-ray Images using CNN

Welded joints in pressure vessels, pipelines and structures are inspected with X-ray radiography — a trained interpreter studies each film for dark rounded indications (porosity), sharp lines (cracks) or elongated misses (lack of fusion). It is slow, skilled work, and fatigue causes misses. Computer vision can pre-screen: a detection CNN learns the visual signatures of each defect type and flags candidates with boxes a human then confirms. This project builds that pre-screener on the GDXray database's weld series (88 weld radiographs in three series, with defect annotations) using Faster R-CNN — a region-proposal network finds candidate indications and a classifier names them. The inspection viewer renders boxes with class and confidence over each radiograph, aggregates batch verdicts, and reports per-class average precision.

1. The GDXray weld series (88 radiographs across 3 series with defect annotations) is parsed into COCO-style training records.
2. Images are contrast-normalized (CLAHE) and augmented with rotations, flips and contrast jitter that preserve radiographic appearance.
3. Anchors are tuned toward small, elongated boxes matching porosity clusters and lack-of-fusion indications.
4. A Faster R-CNN (ResNet-50 FPN backbone) is trained with the RPN and classifier heads jointly, using focal-style sampling for the rare crack class.
5. At inspection time, each radiograph is passed through the detector; boxes above the confidence threshold are kept after non-maximum suppression.
6. Detections aggregate per weld into an accept/reject verdict following a documented workmanship rule set.
7. The held-out split is scored with mAP@0.5, per-class AP and recall, and the viewer renders every detection for visual verification.

Tech stack:

• Python 3, PyTorch (Faster R-CNN, ResNet-50 FPN)
• torchvision detection models
• GDXray weld series (88 radiographs, 3 series)
• OpenCV, CLAHE contrast normalization
• NumPy, Matplotlib
• HTML/CSS/JavaScript inspection viewer (canvas)
• Git

Dataset & model details

• Dataset: GDXray (GRIMA X-ray database, U. de Chile) — weld series: 88 weld radiographs in 3 series (W0001: 20 complete weld samples) with defect annotations.
• Task: Defect detection + classification; input = weld radiograph, output = bounding boxes with defect class (porosity, crack, lack of fusion, inclusion) and confidence.
• Model: Faster R-CNN with ResNet-50 FPN backbone; RPN anchors tuned for small elongated indications; joint RPN + classifier training.
• Metrics: mAP@0.5 ≥ 0.82 and defect recall ≥ 0.90 (design targets); per-class AP reported. No accuracy is claimed as measured until the training run is executed for the order.

• [Faster R-CNN detector] Region-proposal network plus classifier localizes four defect types — porosity, crack, lack of fusion, inclusion — with confidence scores.
• [Radiograph inspection viewer] Web app rendering each X-ray with defect boxes, class labels, confidence and the accept/reject verdict.
• [GDXray training pipeline] Annotation parsing, anchor tuning for small elongated defects, augmentation (contrast, rotation) suited to radiographs.
• [Batch queue reporting] Per-weld defect tables, defect-type mix charts and accept/reject tallies across an inspection batch.
• [Per-class evaluation] Average precision per defect type plus recall at the operating point, with the crack class honestly reported as the hardest.
• [Confidence thresholding] Tunable operating point balancing missed defects against false calls, with the trade-off curve in the report.
• [Indication sizing] Approximate defect length estimated from box dimensions against the image scale marker for the report.

• GDXray parsing, augmentation and training pipeline
• Trained Faster R-CNN weld defect detector
• Radiograph inspection viewer (boxes, verdicts)
• Batch queue report generator
• Per-class AP evaluation and threshold analysis
• Project report PDF (background, detection theory, NDT context, results)
• PPT presentation for final review
• Viva Q&A preparation document (R-CNN family, anchors, mAP, radiographic testing)

• 88 radiographs is a small dataset — heavy augmentation and transfer learning carry the training, and the report states this limit plainly.
• 0.82 mAP is a design target for the training run, stated honestly — the report documents the actual achieved figure after training.
• The system pre-screens; final accept/reject on real welds remains a certified human inspector's decision — this is an assistive tool, not a certifying authority.
• Cracks are the hardest class (thin, rare) and will show the lowest AP — reported honestly per class.
• Only the four trained defect types are detected; unusual indications outside them are out of scope.