An automated fabric defect detection system comparing three deep learning architectures on the TILDA (Tianchi Alibaba) textile defect dataset.
Manual quality inspection in textile manufacturing is slow and inconsistent. This project trains and benchmarks object detection / classification models to automatically identify defects in fabric images.
| Model | Task | Approach |
|---|---|---|
| Swin Transformer | Classification | Hierarchical vision transformer with shifted windows |
| Faster R-CNN | Object Detection | Two-stage region proposal network |
| EfficientNet | Classification | Compound scaling of depth, width & resolution |
- TILDA Textile Dataset — sourced from Alibaba Tianchi competition
- Contains fabric images labeled with defect regions across multiple defect categories
- Preprocessing: resizing, normalization, augmentation (flips, rotations)
Fabric_Defect_Detection/
├── swin_transformer.ipynb # Swin Transformer training & evaluation
├── faster rcnn.ipynb # Faster R-CNN training & evaluation
├── efficientnet.ipynb # EfficientNet training & evaluation
└── README.md
pip install torch torchvision timm albumentations- Data Loading — Load TILDA images and defect annotations
- Preprocessing — Resize, normalize, and augment images
- Model Training — Train each architecture independently
- Evaluation — Compare mAP / accuracy across models
- Analysis — Identify best-performing architecture per defect category
| Model | Metric | Score |
|---|---|---|
| Swin Transformer | Accuracy | — |
| Faster R-CNN | mAP | — |
| EfficientNet | Accuracy | — |
Results to be updated after full training runs.
- TILDA Textile Dataset — Tianchi/Alibaba
- Swin Transformer Paper (Liu et al., 2021)
- Faster R-CNN Paper (Ren et al., 2015)
- EfficientNet Paper (Tan & Le, 2019)
Yohan Gala — B.Tech Computer Engineering, KJSIT Mumbai
GitHub