Skip to content

Latest commit

 

History

History
71 lines (53 loc) · 2.23 KB

README.md

File metadata and controls

71 lines (53 loc) · 2.23 KB
Raw

shinTB

Abstract

A python package for use Textboxes : Image Text Detection Model

implemented by tensorflow, cv2

Textboxes Paper Review in Korean (My Blog) : shinjayne.github.io/textboxes

shintb : useable textboxes python package (Source codes are in here)

svt1 : Street view Text dataset. can use with shintb.svt_data_loader.SVTDataLoader when training Textboxes model

config.py : (NECESSARY) configuration of model building and training with shinTB

main.py : simple example useage of shinTB package

svtexample

Dependancies

  1. python Version: 3.5.3
  2. numpy Version: 1.13.0
  3. tensorflow Version: 1.2.1
  4. cv2

How to use

  1. Clone this repository to your local.
  2. You will use shintb python package and config.py for building and training your own Textboxes model.
  3. svt1 gives us training / test data.
  4. Open new python file.
  5. Import config.config and shintb.
from config import config
from shintb import graph_drawer, default_box_control, svt_data_loader, runner
  1. Initialize GraphDrawer,DefaultBoxControl,SVTDataLoader instance.
graphdrawer = graph_drawer.GraphDrawer(config)

dataloader = svt_data_loader.SVTDataLoader('./svt1/train.xml', './svt1/test.xml')

dbcontrol = default_box_control.DefaultBoxControl(config, graphdrawer)

  1. GraphDrawer instance contains a tensorflow graph of Textboxes.

  2. DefaultboxControl instance contains methods and attributes which is related to default box.

  3. SVTDataLoader instance loads data from svt1.

  4. Initialize Runner instance.

runner = runner.Runner(config, graphdrawer, dataloader, dbcontrol)
  1. Runner uses GraphDrawer,DefaultBoxControl,SVTDataLoader instance.
  2. If you want to train your Textboxes model, use Runner.train(). Every 1000 step, shintb will save ckpt file in the directory you set in config.py.
runner.train()
  1. If you want to validate/test your model, use Runner.test()
runner.test()
  1. After training, if you want to detect texts from one image use Runner.image().
runner.image(<your_image_directory>)