Fabric Classification Using OpenCV

A computer-vision system that identifies the fabric type of a textile image using Local Binary Pattern (LBP) texture features. It trains an SVM (RBF) and a Random Forest, keeps the better one, and serves it through a Streamlit app with per-class probabilities.

Results on the Ten Fabrics Dataset (TFD), 10 fabric types, 594 held-out images:

Model	Test accuracy	Macro F1
SVM (RBF, C=10)	100.0%	1.000
Random Forest (200 trees)	100.0%	1.000

Developed as an Industrial-Oriented Mini Project for the B.Tech in Computer Science and Engineering (AI & ML).

---

Features

• Texture-based classification from a compact 26-dimensional LBP histogram
• SVM (RBF) and Random Forest compared on the same stratified split, with the better model saved automatically
• Reproducible results from fixed random seeds
• Streamlit app showing the predicted fabric, its confidence, and probabilities for every class

Working Pipeline

1. Load: images are read from one folder per fabric type and resized to 128×128
2. Extract features: grayscale conversion, then a uniform LBP (radius 3, 24 points) turned into a normalized 26-bin histogram
3. Split: stratified 80:20 train/test split (2,375 / 594 images)
4. Train: SVM (RBF, C=10, γ=scale) and Random Forest (200 trees)
5. Select: the higher-accuracy model is saved as fabric_classifier_model.pkl, with label_encoder.pkl
6. Predict: the Streamlit app applies the same resizing and LBP extraction to an uploaded image

Results

Test set (stratified 80:20 split)

Both models classify all 594 test images correctly.

Robustness checks

TFD patches are cut from one scan per fabric, with 10% overlap between neighbouring patches. A random split can therefore place near-duplicate patches in both train and test. To check that the result isn't caused by that leakage, the models were also evaluated in two stricter ways:

Evaluation	SVM	Random Forest
5-fold stratified cross-validation	99.8% ± 0.2	99.97% ± 0.07
Region hold-out: train on the first 80% of each fabric's patches, test on the last 20% (a separate area of the scan)	100.0%	100.0%
Region hold-out, reversed: test on the first 20%	99.3%	100.0%

Accuracy stays at or above 99% when whole regions of each fabric are unseen during training. So the result reflects real texture differences, not memorized overlapping patches.

Note: each TFD class is a single fabric from one factory scan, which makes this an easy benchmark. Expect lower accuracy on photos of other fabrics, or under different lighting or scale. See Future Improvements.

Dataset

Ten Fabrics Dataset (TFD): 2,969 scanned 256×256 textile patches across 10 fabric types, labelled 001–010. It was originally built for defect detection, and about 25% of the patches contain defects; they are kept here as normal samples of their fabric type.

Type	001	002	003	004	005	006	007	008	009	010
Images	154	136	144	396	110	345	760	329	367	228

The dataset is not included in this repository. Download it from Kaggle and extract it so the folder sits next to model.py:

TFD Textile Dataset/
├── 001/
├── 002/
│   ...
└── 010/

Technologies Used

Language: Python

Libraries: OpenCV · NumPy · scikit-learn · scikit-image · Streamlit · Pillow · Joblib

Installation

git clone https://github.com/bsaaketh/Fabric-Classification-Using-OpenCV.git
cd Fabric-Classification-Using-OpenCV
pip install -r requirements.txt

Train the model

python model.py

This prints accuracy and a classification report for both models, then saves fabric_classifier_model.pkl, label_encoder.pkl, fabric_svm_model.pkl and fabric_rf_model.pkl.

Run the app

streamlit run app.py

Project Structure

Fabric-Classification-Using-OpenCV/
├── TFD Textile Dataset/     # not included, see Dataset
├── app.py                   # Streamlit interface
├── model.py                 # LBP features, SVM + RF training, model selection
├── requirements.txt
├── assets/                  # README images
└── README.md

Applications

Textile manufacturing · automated quality control · inventory management · e-commerce product tagging

Future Improvements

• Test on photos of fabrics outside TFD, under varied lighting and scale
• Add named material classes (cotton, denim, silk, wool…) from a multi-source dataset
• Combine LBP with GLCM, Gabor and color features, or move to CNN transfer learning
• Use TFD's defect labels to add defect detection