Scene text recognition

• Main Author: Muhammad Afiq Osman
• Other Authors: Lu Shijian
• Format: Final Year Project
• Language: English
• Published: 2019
• Subjects: DRNTU::Engineering::Computer science and engineering
• Online Access: http://hdl.handle.net/10356/76861
• Institution: Nanyang Technological University

Scene text recognition problem has recently seen interest within the deep learning community. Solving such a problem will inevitably open paths to more exciting inventions in the future such as for robotic navigation. However, the current solutions are far from perfect and there is potentially more that could be done. In this FYP we will investigate and attempt to replicate an existing study regarding scene text recognition. The goal is to theoretically understand and experience the practical side of implementing and training a deep learning model to tackle such a problem. We would undertake the hyperparameter tuning process in search of the optimal values for the batch size, learning rate, number of epochs and the learning optimizer. Optimal values found for batch size and learning rate coincides with the common rationale. The same goes for the number of epochs where the resulting trend suggests that as the number of epochs increases the model accuracy will start to plateau. However, as for the best optimizer was found to be Adam which was different from the original’s study optimizer of Adadelta. Adadelta in fact performed much worse producing ‘nan’ test and train error on many occasions. Future recommendation for this FYP includes experimenting with the CRNN model structure used in order to deeply understand the effects of the CNN and RNN layers used.