Scene Text Localization and Recognition for Images of Serial Numbers and Odometer Readings

Abstract

This thesis investigates deep learning models and methods to create a full scene text extraction system. The system is composed of two main parts, a localization network and a recognition network, with the recognition network being the main focus. The localization network is a segmentation network that localizes the region in an image containing text. Once this region is identified the recognition network predicts the text within the image. In addition to investigating these models, we look at data processing, data generation, and model prediction processing techniques to improve the system’s robustness and make the learning processes easier.