Script-Identifier

Scene Text Script Classification using Traditional Machine Learning Techniques

Overview

Script-Identifier is a Pattern Recognition and Machine Learning (PRML) course project developed as part of CSL2050. The objective of the project is to identify the script/language of scene text images using traditional machine learning techniques. The system is trained and evaluated on the Bharat Scene Text Dataset, and the focus is on comparing classical ML algorithms with various handcrafted and deep feature extractors.

This project adheres to the CSL2050 guidelines, emphasizing rigorous evaluation, failure case analysis, and comprehensive deliverables including code, report, demo, and presentation.

---

Project Team

Developed by Team Lexiconauts
Under CSL2050 - Pattern Recognition and Machine Learning

---

Features

• Classification of scene text into scripts such as Hindi, Tamil, Bengali, etc.
• Traditional ML models: SVM, KNN, ANN, Decision Trees, Logistic Regression, etc.
• Feature extraction techniques: HOG, SIFT, ResNet, VGG, Vision Transformer (ViT)
• Dimensionality reduction: PCA, LDA
• Visualizations: t-SNE, PCA, class distributions, and decision boundaries
• FastAPI based interfaces for inference and demo
• Modular, well-documented codebase with YAML-based configuration

---

Repository Structure

.
├── config/              # Configuration files for experiments
├── dataset/             # Data loaders and transformations
├── models/              # ML models and backbones
├── tools/               # Training and inference scripts
├── utils/               # Plotting and utility functions
├── Visualisation/       # Visualization scripts
├── data/                # Placeholder for processed data and latents
├── frontend/            # Next.js frontend interface
├── main.py              # Entry point for inference
├── ui.py                # UI
├── fastapi_server.py    # REST API with FastAPI
├── requirements.txt     # Dependencies
└── README.md            # Repo documentation

---

Installation

Prerequisites

• Python 3.10+
• CUDA-enabled GPU (for faster training)
• pip / conda for package management

Setup

git clone https://github.com/AurindumBanerjee/Script-Identifier.git
cd Script-Identifier
pip install -r requirements.txt

---

How to Run

Refer to the Logistic Regression ReadMe.md for an example of how to run models. Every other model made also runs similarly.

Use the following code from root to run the model. [Follow path structure from root, separated via '.' instead of '']

python -m model_folder.model_file

To run the fastapi server

python fastapi_server.py 

For UI

cd frontend
npm install
npm run dev

---

Dataset

We use the Bharat Scene Text Dataset containing scene text in 13 Indian scripts.

Download the dataset from the official Bharat Scene Text Dataset and put the dataset folder in data directory as data/recognition

.
└── data/recognition/
    ├── train/        # Training data directory with subfolders for each class (language)
    │   ├── english/
    │   ├── gujarati/
    │   ├── hindi/
    │   └── ...       # Additional language/class folders
    │
    ├── test/         # Test data directory with subfolders for each class (language)
    │   ├── english/
    │   ├── gujarati/
    │   ├── hindi/
    │   └── ...       # Additional language/class folders
    │
    ├── train.csv     # CSV file mapping training images to their labels
    └── test.csv      # CSV file mapping test images to their labels

In the data, we specifically use the Cropped Word Recognition Set. Data splits are provided for training and testing.

Language	Train	Test
Assamese	2,623	1,343
Bengali	4,968	1,161
English	28,778	8,113
Gujarati	1,956	693
Hindi	14,855	4,034
Kannada	2,241	693
Malayalam	2,408	567
Marathi	3,932	1,045
Odia	3,176	1,022
Punjabi	8,544	2,560
Tamil	2,041	507
Telugu	2,227	482
Others	20600	-
Total	77,749	22,220

---

Deliverables

As per CSL2050 project requirements:

Component	Status
Mid-Project Report	✅ Submitted
Final Report	✅ Submitted
GitHub Repository	✅ Updated
Project Page	✅ Web Page
Web Demo (Gradio)	✅ Created
Spotlight Video	✅ Submitted
Minutes of Meetings	✅ Maintained

---

Citation

@misc{BSTD, title = {{B}harat {S}cene {T}ext {D}ataset}, howpublished = {\url{https://github.com/Bhashini-IITJ/BharatSceneTextDataset}}, year = 2024, }

---

License

This project is licensed under the MIT License.