The research problem was identified and defined by the external industry, GSK Global Pty Ltd. After several sessions with direct feedbacks from GSK Global, the draft software has been developed and approved by our industry partner. GSK provided funding for this collaboration.
Methodologies used for developing the software tool:
This is a classic scene text recognition problem. There were two main challenges in this project. They are –
1.Real-time imaging data analysis, feature extraction, pattern recognition using machine learning and post-processing of the extracted data.
2.Processing the whole computational analysis locally in hand-held mobile device
This project has implemented algorithms from several open-source repositories and journals. The data processing pipeline for the implementation of this imaging data analysis broadly used following methods:
•We used CRAFT model developed by Baek, Lee, Han, Yun, and Lee (2019) for alpha-numeric text detection
•For sequencing model, we used an End-to-End Trainable Neural Network for Image-based Sequence Recognition, CRNN developed by Shi, Bai, and Yao (2016).
•For Feature extraction, we used Resnet model He, Zhang, Ren, and Sun (2016).
•For Word and character level Sequence labelling, we used LSTM based model developed Ma and Hovy (2016) .
•We used sequence learners based on recurrent neural networks (RNNs) developed by Graves, Fernández, Gomez, and Schmidhuber (2006)
•For Scene Text Recognition Model, we used training model developed by J. Baek et al. (2019)
•Finally, we used a customized pattern recognition model based on ISO 6346.
| Period | 04 Aug 2022 → 30 Jan 2023 |
|---|
| Work for | GSK Global Pty Ltd |
|---|
| Degree of Recognition | National |
|---|
- image processing
- data analysis
- text recognition
- data management
- SDG 11: Sustainable Cities and Communities