🌧️ Weather Invariant OCR

An End-to-End Pipeline for Deraining → Detection → OCR

Live processing of rainy scene text (Desktop/Jetson compatible)

🚀 Key Features

• Multi-Stage Pipeline:

  • 🌧️ Rain Removal: Restormer model
  • 🔍 Text Detection: CRAFT detector with geometric filtering
  • 🔠 Text Recognition: PARSeq transformer OCR
  • 🤖 LLM Enhancement: Optional Gemini 2.0 Flash post-processing

• Multiple Interfaces:

  • 🖥️ Web UI (Streamlit) for static images
  • 🎥 Live Processing (Jetson/OpenCV)

📂 Project Architecture

🛠️ Installation

Prerequisites

• Python 3.11 (3.6.9 for Jetson)
• NVIDIA Jetson Nano (Optional for edge deployment)

Setup

git clone https://github.com/anshul45-github/Weather-Invariant-OCR
cd Weather-Invariant-OCR

# Create environment (Recommended)
python -m venv venv
source venv/bin/activate  # Linux/MacOS
.\venv\Scripts\activate   # Windows

# Install core dependencies
pip install -r requirements/base.txt  # or requirements/jetson.txt for edge devices

# Download pretrained models
bash scripts/download_models.sh

💻 Usage Options

1. Web Interface (Streamlit)

cd src/core
streamlit run ../../frontend/app.py

Features drag-n-drop processing and side-by-side comparisons

2. Command Line

# Single image processing
cd src/core
python pipeline.py

# Live camera processing (Jetson)
cd src_jetson/core
python text_detection_live.py

Tested by adding synthetic rain to ICDAR-2013 dataset

📝 Academic Context

Developed for IC201P Design Practicum at IIT Mandi
Team:

• Anshul Mendiratta
• Bhupesh Yadav

🪧 Live Demo Files

Google Drive