Training deep network models for accurate recognition of texts in scenes

• Main Author: See, Yu Xiang
• Other Authors: Lu Shijian
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision
• Online Access: https://hdl.handle.net/10356/147992
• Institution: Nanyang Technological University

Scene Text Recognition is a challenging research task in the domain of computer vision for many years due to dynamic conditions of text in natural scenes. The emergence of deep learning solutions created new possibilities and has also shown significant progress and performance by playing a role in various vision-based applications. In this paper, deep network models will be implemented by emulating a state-of-the-art neural network architecture that utilize image-based sequence recognition for scene text recognition. Fine-Tuning of hyperparameters includes type of optimizers, learning rate, batch size and number of epochs to obtain the best configurations of the model for deployment by measuring against benchmark datasets. After doing so, the model’s configuration of using Adam optimizer was found to be performing better than the AdaDelta optimizer which was mentioned in the original paper. A text recognition program is also built to demonstrate the functionality of the trained model in real-time scenario. Further recommendation for this project includes exploring different methodologies to provide an end-to-end model capable of performing text detection and recognition on curved texts in scene images.