Real-time Scene Text Detection with Differentiable Binarization

Introduction

The Differential Binarization (DB) Algorithm is one of the cutting-edge approaches to effectively detect curved text.

• Improved Text Detection: The algorithm excels at accurately identifying text within images, even when it's curved or distorted.
• Accurate Text Recognition: It paves the way for more precise text recognition, ensuring that the text is correctly extracted and understood.

Environment

The dependencies are listed in requirements.txt. Please install and follow the command below:

pip install -r requirements.txt

Data preparation

Please download the ICDAR2015 and TotalText dataset and set up the folder structure:

dataset/icdar2015
|test
|____|gt
|____|______gt_img_1.txt
|____|______gt_img_2.txt
|____|images
|____|______img_1.jpg
|____|______img_2.jpg
|train
|____|gt
|____|______gt_img_1.txt
|____|______gt_img_2.txt
|____|images
|____|______img_1.jpg
|____|______img_2.jpg

Training

Before training, please modify configurations in src/configs/det_icdar2015.yml

python -m src.train

Evaluation

python -m src.evaluate

Prediction

python -m src.predict --image_path <path_to_image>

Example:

python -m src.predict --image_path images/example.jpg

test1.png	test2.png

Experiments

Export format	image size	mAP	mAP_50	mAP_75	Inference time (RTX3060)	learning rate
Pytorch - ResNet18	736x736	0.36	0.65	0.36	0.003s	0.0005
TorchScript - ResNet18	736x736	0.36	0.65	0.36	0.0018s	0.0005
Pytorch - ResNet50	736x736	0.40	0.70	0.40	0.003s	0.007
TorchScript - ResNet18	736x736	0.40	0.70	0.40	0.004s	0.0007

Reference

• https://github.com/MhLiao/DB
• Real-time Scene Text Detection with Differentiable Binarization