CRNN-OCR-lite
CRNN-OCR-lite is a lightweight deep learning model for optical character recognition, built with Keras and TensorFlow. It is designed for word-level text recognition, including handwritten text transcription. The architecture is based on the standard CRNN model combined with CTC loss, enhanced with a spatial transformer module to correct text slope and distortion, and depthwise separable convolutions to reduce complexity and parameter count. The model uses a two-stage transfer learning approach: initial pretraining on the large synthetic mjsynth dataset, followed by fine-tuning on the handwritten IAM dataset. This produces two models: one optimized for scene text in the wild and another for handwritten text. On a validation set of 8000 images, the mjsynth model achieves a 0.09 normalized edit distance, while the IAM model achieves 0.08. The project includes training scripts, preprocessing utilities for the IAM dataset, and a prediction script for inference. It serves as a practical demonstration of building a