Fabric Defect Detection using Deep Learning

Garment manufacturers still inspect fabric the slow way: a worker watches the roll unspool under bright light and marks defects by eye. It is tiring, inconsistent, and the defects that slip through — a hole here, a slub there — become rejected garments or customer complaints downstream. Manual inspection typically catches only the obvious flaws, and the checker cannot keep a quantitative log of defect rates per roll. This project replaces the eyeball with a camera and a convolutional neural network: a ResNet-50 model, pre-trained on ImageNet and fine-tuned on the published TILDA textile dataset, classifies each fabric frame into defect-free or one of seven defect types, while a Grad-CAM heatmap is thresholded into bounding boxes that show exactly where each defect sits. A web-based inspection console ties it together — upload a fabric photo, run the pipeline, and get labeled boxes with confidence scores plus a defect-rate dashboard across rolls. Nothing is overstated: accuracy figures are presented as design targets, and the delivered report documents the measured metrics from the actual training run.

1. You photograph a fabric roll frame (or upload a swatch image); the console resizes it to 512×512 and normalizes it exactly as the training data was prepared.
2. The ResNet-50 backbone extracts texture features; a global-average-pooling plus dense head outputs probabilities over the 8 classes (defect-free + 7 defect types).
3. For any predicted defect class, a Grad-CAM heatmap is computed from the final convolutional layer to show which image regions drove the decision.
4. The heatmap is thresholded and connected-component labeled into bounding boxes; boxes below the 0.50 confidence threshold are discarded.
5. Each surviving box is drawn on the frame with its defect-type label and confidence score, and the frame gets a PASS or DEFECTIVE verdict.
6. Frame results are aggregated per roll into a defect-rate log (defects per 100 m); rolls crossing the configurable rework threshold are flagged.
7. The dashboard renders the trend across rolls, and the training notebook regenerates all metrics — per-class precision/recall and the confusion matrix — from the order's own training run.

Tech stack:

• Python 3 · PyTorch
• ResNet-50 · transfer learning
• TILDA textile dataset
• OpenCV · Albumentations
• Focal loss · class weights
• Flask inspection app
• HTML5 · CSS · JavaScript console
• ONNX model export

Dataset & model details

• Dataset: TILDA Textile Texture Database (DFG Texture Analysis group, University of Freiburg) — approximately 3,200 fabric images (768×512) spanning 8 fabric types in 4 structural classes (C1–C4), with a defect-free class (e0) and defect classes including hole, thread error, oil spot and foreign body. Slub, misweave, snag and broken-end examples are added from a supplementary mill collection labeled under the same 7-defect taxonomy.
• Task: 8-class fabric-frame classification with localization; input = 512×512×3 RGB image, output = probability distribution over 8 classes plus Grad-CAM-derived bounding boxes.
• Model: ResNet-50 backbone (~25.6 M parameters, ImageNet-pretrained, fine-tuned) → global average pooling → Dense(256, ReLU) → Dropout(0.4) → Dense(8, softmax); trained with focal loss and class weights.
• Metrics: Macro defect recall target ~92%+ and false-alarm rate target <5% on defect-free fabric (design targets for the built-to-order training run), per-class precision/recall/F1, confusion matrix. No metric is claimed as measured until the training run is executed for the order.

• [ResNet-50 defect classifier] ImageNet-pretrained ResNet-50 fine-tuned to an 8-way softmax — defect-free plus hole, slub, misweave, broken end, oil spot, snag and thread error — with the fine-tuning schedule documented.
• [Grad-CAM defect localization] Class-activation heatmaps are thresholded into bounding boxes over the fabric frame, each labeled with defect type and confidence, so the output shows where a defect is, not just that one exists.
• [Inspection web console] Upload any fabric photo, run the full pipeline in the browser demo, and get bounding-box overlays, a ranked detection list with confidence bars, and per-frame statistics (defect count, defective area, inference time).
• [Roll batch queue] Review multiple fabric rolls in one grid; each roll carries its frame count and a PASS/DEFECTIVE verdict with defect counts for quick triage.
• [Defect-rate dashboard] Tracks defects per 100 m of fabric across rolls with a configurable rework threshold, so quality trends are visible instead of anecdotal.
• [Imbalance-aware training] Focal loss with class weights counters TILDA's heavy defect-free skew; the train/val/test split is stratified by fabric roll so no roll leaks across splits.
• [Full training notebook] Data preparation, augmentation, model definition, training loop and evaluation in one reproducible notebook, including per-class precision/recall, the confusion matrix and training curves.

• Complete training & evaluation notebook (data prep, augmentation, training, metrics)
• Trained ResNet-50 defect classifier (SavedModel + ONNX export) with preprocessing code
• Fabric inspection web console wired to the trained model
• Per-class precision/recall, confusion matrix and training-curve plots
• Project report PDF (background, CNN/transfer-learning theory, methodology, measured results)
• PPT presentation for final review
• Viva Q&A preparation document (CNNs, transfer learning, Grad-CAM, focal loss, overfitting, dataset bias)

• TILDA covers mostly patternless woven fabrics — bold printed or highly patterned textiles are a known hard case and may need extra training data.
• Detection quality depends on capture conditions: even diffuse lighting and a fixed camera distance are assumed; harsh shadows or motion blur degrade results.
• Subtle defects (snags, faint oil spots on dark fabric) are the weakest classes; the report calls out their lower recall honestly instead of hiding it.
• The ~92% recall and <5% false-alarm figures are design targets stated up front — the report documents the actually measured values after the order's training run.
• CPU inference at ~120–180 ms/frame suits roll-sample inspection, not full-speed continuous line scanning without a GPU.
• This is an educational prototype, not certified quality-control equipment — mill deployment would need calibration against the buyer's own rolls.