/signboard-transliteration

A project to perform transliteration on commercial signboard images

Primary LanguageJupyter NotebookMIT LicenseMIT

The code organization is WIP at the moment. Will update the repo after it is ready to be checked in.

Signboard transliteration

A project to perform transliteration on commercial signboard images.

The project involves 3 stages:

  1. Text Detection: To detect text regions in signboard images
  2. Text Recognition: To decipher the hindi text in the regions detected in Step 1
  3. Text transliteration: To convert the hindi text recognized in Step2 to English.

The project contains 4 ipython notebooks written in PyTorch. The notebooks can be described as follows:

  1. proj_station_detection_preprocess.ipynb: Used to correct bounding box annotations provided by the dataset. This is a preprocessing step before the object detection step.

  2. proj_station_detection_od.ipynb: Used to finetune a Faster-RCNN model to detect text regions in Signboard images

  3. project_station_td_preprocess.ipynb: Used to tokenize the text labels provided in the Text Recognition dataset. This is a preprocessing step before the Text Recognition step.

  4. project_station_td_recognition.ipynb: Used to train a Text Recognition systems using an encoder decoder model inspired by the paper Show and Tell.

  5. Text Transliteration: [WIP]

Author

Kushagra Pandey / @kpandey008