xiaofengShi/CHINESE-OCR

What it solves

This project provides a pipeline for detecting and recognizing Chinese text in natural scene images, handling challenges like varying text orientation and variable-length text strings.

How it works

The system uses a three-stage network architecture:

1. Text Orientation Detection: A VGG16-based classifier that detects if text is rotated by 0, 90, 180, or 270 degrees.
2. Text Region Detection: A CTPN (CNN+RNN) network that identifies the bounding boxes of horizontal text areas.
3. End-to-End Recognition: A CRNN (CNN+GRU/LSTM+CTC) network that recognizes the actual characters within the detected regions without needing to segment individual characters.

Who it’s for

Researchers and developers interested in Optical Character Recognition (OCR) for Chinese text, specifically those looking for historical implementations using TensorFlow 1.x, Keras, or PyTorch.

Highlights

• Multi-stage Pipeline: Combines orientation detection, region localization, and character recognition.
• C-T-C Loss: Employs Connectionist Temporal Classification (CTC) to recognize variable-length text without manual character-level segmentation.
• Flexible Frameworks: Provides training code for both Keras and PyTorch.
• Comprehensive Dataset Guide: Includes a curated list of large-scale public datasets for training Chinese text detection and recognition models.