/OCR_detection_IC15

OCR detection for ICDAR2015, which is based on FOTS, the precision is 80.6%.

Primary LanguageC++

OCR detection for ICDAR2015, which is based on FOTS detection algorithm.

Introduction

This project is a pytorch implementation of fots detection for OCR, and we focus on achieved the detection algorithm only:paper link-FOTS: Fast Oriented Text Spotting with a Unified Network. The code is created by Ning Lu originally, and we would like to appreaciate to his contributions.

What we are doing and going to do

  • Change some code to make the project work.
  • Add PSPnet model to experiment, but is not work effectively(the another project that we have doing: code).
  • Support visdom.
  • Support pytorch-0.4.1 or higher.

Benchmarking

We benchmark our code thoroughly on the dataset: ICDAR2015, using network architecture: resnet50. It's worth noting that, the project had used the multi-scale to train network and haven't done the skill of OHEM. Below are the results:

1). ICDAR2015 (scale=512):

model #GPUs batch size lr Recall Precision Hmean
Res-50 1080Ti 4 1e-3 69.72% 80.09% 74.54%

Preparation

First of all, clone the code

git clone https://github.com/Vipermdl/OCR_detection_IC15

prerequisites

  • Python 3.6
  • Pytorch 0.4.1
  • CUDA 8.0 or higher

Data Preparation

  • ICDAR 2015: Please download the dataset in the folder in your project named dataset, you can refer to any others. After downloading the data, creat softlinks in the folder data/.

Compilation

Install all the python dependencies using pip:

pip install -r requirements.txt

Train

Try:

python train.py

Test

If you want to evlauate the detection performance, simply run

python eval.py

Below are some detection results:

Authorship

This project is equally contributed by Ning Lu and DongLiang Ma, and many others (thanks to them!).