Textile Weave Defect Inspection using CNN

Fabric mills still inspect cloth the slow way — an operator watches meters of fabric glide past under bright lights, and misses defects when attention fades. The defect types are visually subtle: a missing weft thread, a small oil stain, a slub in the yarn. Machine vision is the natural fix, but training from scratch needs thousands of defective samples mills don't share. The practical route is transfer learning: a ResNet-50 pretrained on ImageNet already understands texture, and fine-tuning on the TILDA textile database (3,200 images across 8 classes — one defect-free plus seven defect types) teaches it fabric-specific faults. This project builds that pipeline — tiling high-resolution fabric images, classifying and localizing defects per tile, aggregating to roll-level grades, and rendering an inspection UI with boxes, confidence scores and shift batch reports.

1. Fabric images from the TILDA database (3,200 images, 8 classes) are tiled into 512×512 patches with 256-pixel stride; defect bounding boxes are mapped into tile coordinates.
2. Tiles are augmented (rotations, flips, brightness jitter) to simulate lighting variation on the inspection line.
3. A ResNet-50 backbone (ImageNet-pretrained, progressively unfrozen) with a detection head is fine-tuned with Adam.
4. At inspection time, each frame is tiled identically, tiles are classified, and non-maximum suppression merges overlapping boxes.
5. Tile detections aggregate per roll: defect count, defect-area percentage and the worst defect class determine the grade.
6. The inspection app renders boxes with class + confidence on the fabric image and the batch view summarizes the shift.
7. The held-out TILDA split is scored with per-class F1, mAP@0.5 and false-alarm rate for the report.

Tech stack:

• Python 3, PyTorch (ResNet-50, detection head)
• torchvision (pretrained weights, transforms)
• TILDA textile database (3,200 images, 8 classes)
• DAGM 2007 (textured-surface extension)
• NumPy, Matplotlib
• HTML/CSS/JavaScript inspection app (canvas)
• Git

Dataset & model details

• Dataset: TILDA (DFG Texture Analysis group) — 3,200 fabric images across 8 classes (1 defect-free + 7 defect types: hole, oil stain, thread error, slub, missing weft, knot, nep); extended with DAGM 2007 textured-surface defect classes.
• Task: Defect detection + classification; input = 512×512 fabric tile, output = bounding boxes with defect class and confidence.
• Model: ResNet-50 backbone (ImageNet-pretrained, progressively unfrozen) + detection head; Adam optimizer, focal loss for class imbalance.
• Metrics: mAP@0.5 ≥ 0.88 and false-alarm rate ≤ 3% (design targets); per-class F1 reported. No accuracy is claimed as measured until the training run is executed for the order.

• [Tile-based inspection] Fabric frames are split into 512×512 tiles with stride overlap so defects on tile borders are never missed.
• [Defect localization] Bounding boxes with class labels and confidence scores drawn on every defective tile in the inspection view.
• [ResNet-50 transfer learning] ImageNet-pretrained backbone fine-tuned on TILDA, with the backbone-freezing schedule documented for the viva.
• [Roll grading] Tile results aggregate to roll-level verdicts — first quality vs seconds — with defect-area percentage.
• [Batch shift reports] Per-roll defect counts, defect-type mix and grade distribution across a production shift.
• [False-alarm control] Tunable confidence threshold balancing missed defects against false stops, with the trade-off curve in the report.
• [DAGM 2007 extension] Additional textured-surface defect classes from the DAGM 2007 set broaden the model beyond plain weaves.

• TILDA tiling, augmentation and training pipeline
• Trained ResNet-50 defect detector with preprocessing
• Inspection web app (boxes, confidence, roll grading)
• Batch shift-report generator
• Per-class evaluation and threshold trade-off analysis
• Project report PDF (background, transfer-learning theory, methodology, results)
• PPT presentation for final review
• Viva Q&A preparation document (CNNs, transfer learning, mAP, focal loss)