Skip to content
MORAN: A Multi-Object Rectified Attention Network for Scene Text Recognition, implemented in tf.keras
Python
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.

Files

Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
.vscode
docs
moran
tests
.gitignore
LICENCE
README.md
poetry.lock
pyproject.toml

README.md

MORAN: A Multi-Object Rectified Attention Network for Scene Text Recognition

The implementation is under active development. Beware of bugs and breaking changes!

This repository contains a tf.keras implementation of MORAN, a neural network with rectified attention for general scene text regonition outperforming all current state-of-the-art approaches with the lexicon-free mode of operation.

The codebase is heavily inspired by the Pytorch implementation provided by the original authors.

Table of Contents

  1. Technical Details

  2. Examples & Limitations

Technical Details

The model consists of two parts: MORN (Multi-Object Rectification Network), which learns the offsets used for text rectification, and ASRN (Attention-based Sequence Recognition Network), a CNN-LSTM encoder coupled with an attention-based decoder.

The figure below shows the overall structure of the model:

MORN (Multi-Object Rectification Network)

The architecture of MORN is given in the table below:

ASRN (Attention-based Sequence Recognition Network)

The architecture of ASRN is given in the table below:

Examples & Limitations

You can’t perform that action at this time.