Script-Identifier

Script-Identifier is a PRML course project under CSL2050 at IIT Jodhpur. It identifies the script of text in scene images using traditional machine learning methods with a focus on explainability and performance.

• 13 Indic scripts supported
• Feature extractors: HOG, SIFT, ResNet, VGG, ViT
• Models: LR, SVM, KNN, ANN, Naive Bayes, Decision Trees, XGBoost
• PCA, LDA for reduction
• t-SNE, PCA plots for visualizations
• Dataset: Bharat Scene Text

Images are passed through feature extractors, then classified with traditional ML. Dimensionality reduction and plotting tools visualize separability and behavior.

Example: Logistic Regression decision boundary (HOG + PCA):

• config/ – YAML experiment configs
• dataset/ – PyTorch dataset setup
• models/ – ML models and backbones
• tools/ – Training & evaluation scripts
• utils/ – Helpers and plotters
• Visualisation/ – PCA, t-SNE, boundary plots
• frontend/ – Next.js UI
• main.py, ui.py, fastapi_server.py

# Run model
python -m models.logistic.LogBackBone

# Train ANN on ViT
python -m tools.ann_train

# API server
python fastapi_server.py

# Frontend UI
cd frontend
npm install
npm run dev
      

• Reports: ✅ Midterm & Final
• Video: ✅ Spotlight Submitted
• Demo: Live App
• GitHub: Repo
• Page: Project Page

By Team Lexiconauts – CSL2050, IIT Jodhpur (Spring 2025).

Thanks to Bhashini team for the dataset.