CRNN

Introduction

CRNN is the abstraction of Convolutional Recurrent Neural Network. It is basically a CNN followed by a RNN. The CNN is actually doing the encoding work, and the RNN is actually doing the decoding work. CRNN can be used to do text recognition. In this process, CRNN takes an image as input and outputs texts contained in the image.

The paper where CRNN comes from is An End-to-End Trainable Neural Network for Image-based Sequence Recognition and Its Application to Scene Text Recognition. The original code of this paper is bgshih/crnn. The codes in this repository are basically built on meijieru/crnn.pytorch.

Dependency

• PyTorch 1.3.0
• lmdb 0.98
• Pillow 6.2.1
• a GPU

Demo

A demo program can be found in demo.py.

A trained model is provided in Baidu Netdisk. You should download it first and put it in trained_model/.

Then launch the demo by:

python demo.py

Example image:

Expected output:

load input image...
load trained model...
predicted label: developer

Result

More results are listed below.

Image:

Output:

load input image...
load trained model...
predicted label: discussion

Image:

Output:

load input image...
load trained model...
predicted label: reception

The model is trained on the dataset of Robust Reading Competition's Born-Digital Images (Web and Email) Challenge Task 1.3. The model can achieve around 70% accuracy on the validationset.

Train a new model

1. Construct a new dataset using create_dataset.py.

2. Execute:

    python train.py --trainset_path PATH_TO_TRAINSET --validationset_path PATH_TO_VALIDATIONSET
   

   The model will be saved every epoch. The model will be saved into directory model/.