https://github.com/watersink/Character-Segmentation

Character Segmentation using TensorFlow
https://github.com/watersink/Character-Segmentation

Last synced: 3 months ago
JSON representation

Character Segmentation using TensorFlow

• Host: GitHub
• URL: https://github.com/watersink/Character-Segmentation
• Owner: watersink
• Created: 2018-06-22T03:19:30.000Z (almost 6 years ago)
• Default Branch: master
• Last Pushed: 2018-06-22T03:38:16.000Z (almost 6 years ago)
• Last Synced: 2024-01-16T02:48:08.553Z (5 months ago)
• Language: Python
• Size: 74.8 MB
• Stars: 27
• Watchers: 5
• Forks: 9
• Open Issues: 1
• Metadata Files:
  • Readme: README.md

Lists

• awesome-ocr - watersink/Character-Segmentation
• awesome-ocr - watersink/Character-Segmentation

README

        
# Character Segmentation

Segment characters and spaces in one text line,from this paper [Chinese English mixed Character Segmentation as Semantic Segmentation](https://arxiv.org/pdf/1611.01982.pdf)

## dependencies

tensorflow1.3,or 1.4

> 

python3

## differences from the paper

the paper set the label of the space to 1,others 0.But that is not hommizate,because the space between two characters is many pixes,the network is hard to distinguish which is 1,which is 0,even though it can work.Here we change to set the characters to 1,spaces to 0.

## architecture of the network

![image]( ./other/1.png)

## Heuristic Rules for balanced_Binary_CrossEntropy

![image]( ./other/2.png)

## make training images and labels

    python3 make_train_images.py

## train

    python3 train_char_seg.py

![image]( ./other/3.png)

## test

    python3 test_char_seg.py

![image]( ./other/4.png)

## other_things

you can choose first make traing images and then use these maked images to train ,or training and making at the same time.all you need to do is change below codes in data_generator.py

> 

    enqueuer = GeneratorEnqueuer(generator_on_the_fly(**kwargs), use_multiprocessing=False)

    #enqueuer = GeneratorEnqueuer(generator_from_folder(**kwargs), use_multiprocessing=False)