CRNN for Text Recognition

This repository provides an implementation of a Convolutional Recurrent Neural Network (CRNN) for sequence modeling tasks, such as text recognition. The model leverages a Convolutional Neural Network (CNN) for feature extraction and a Long Short-Term Memory (LSTM) network for sequence modeling, making it well-suited for tasks that require handling variable-length sequences (e.g., OCR with CTC loss).

Features

• CNN Feature Extractor: A series of convolutional and pooling layers to extract high-level features from input images.
• LSTM Sequence Modeler: An LSTM processes sequential features extracted from the CNN.
• CTC Loss Compatibility: The network output includes an extra "blank" label for compatibility with Connectionist Temporal Classification (CTC) loss.
• Checkpointing: Automatically saves model checkpoints after each epoch for resuming training later.
• Customizable Architecture: Parameters like LSTM hidden size, number of layers, and number of classes can be easily adjusted.

Requirements

• Python 3.6 or higher
• PyTorch
• Torchvision

Additional dependencies (e.g., NumPy) might be required based on your dataset and training pipeline.

Installation

1. Clone the repository:

   git clone https://github.com/yourusername/CRNN-Text-Recognition.git
   cd CRNN-Text-Recognition

2. Create and activate a virtual environment (optional but recommended):

   python -m venv venv
   source venv/bin/activate  # On Windows use: venv\Scripts\activate

3. Install the required packages:

   pip install torch torchvision
   # If using additional dependencies, list them here.

Usage

Dataset & Data Loaders

• Data Preparation:
  Make sure you have defined your data loaders (train_loader and test_loader) to feed training and testing data into the model.
• Character Mapping:
  Define the char_to_index mapping that converts characters to indices. The length of this mapping determines the number of classes for the model.

Model Training

The main training loop is provided in the script. Key points include:

• Model Initialization:
  The model is defined as CRNN(num_classes+1), where the extra class is reserved for the blank token used in CTC loss.

• Checkpoint Loading:
  To resume training from a saved checkpoint, set the load variable to True. The checkpoint is then loaded from checkpoint.pth.

• Running Training:
  The training loop runs for a preset number of epochs (default is 100). After each epoch, training loss, test loss, and test accuracy are printed, and a checkpoint is saved.

To start training, simply run:

python main.py