The code organization is WIP at the moment. Will update the repo after it is ready to be checked in.
A project to perform transliteration on commercial signboard images.
The project involves 3 stages:
- Text Detection: To detect text regions in signboard images
- Text Recognition: To decipher the hindi text in the regions detected in Step 1
- 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:
-
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. -
proj_station_detection_od.ipynb
: Used to finetune a Faster-RCNN model to detect text regions in Signboard images -
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. -
project_station_td_recognition.ipynb
: Used to train a Text Recognition systems using an encoder decoder model inspired by the paperShow and Tell
. -
Text Transliteration: [WIP]
Kushagra Pandey / @kpandey008