Deep Learning for Scene Text Detection, Recognition, and Understanding

Abstract

Detecting and recognizing texts in images is a long-standing task in computer vision. The goal of this task is to extract textual information from images and videos, such as recognizing license plates. Despite that the great progresses have been made in recent years, it still remains challenging due to the wide range of variations in text appearance. In this thesis, we aim to review the existing issues that hinder current Optical Character Recognition (OCR) development and explore potential solutions. Specifically, we first investigate the phenomenon of unfair comparisons between different OCR algorithms caused due to the lack of a consistent evaluation framework. Such an absence of a unified evaluation protocol leads to inconsistent and unreliable results, making it difficult to compare and improve upon existing methods. To tackle this issue, we design a new evaluation framework from the aspect of datasets, metrics, and models, enabling consistent and fair comparisons between OCR systems. Another issue existing in the field is the imbalanced distribution of training samples. In particular, the sample distribution largely depended on where and how the data was collected, and the resulting data bias may lead to poor performance and low generalizability on under-represented classes. To address this problem, we took the driving license plate recognition task as an example and proposed a text-to-image model that is able to synthesize photo-realistic text samples. By using this model, we synthesized more than one million samples to augment the training dataset, significantly improving the generalization capability of OCR models. Additionally, this thesis also explores the application of text vision question answering, which is a new and emerging research topic among the OCR community. This task challenges the OCR models to understand the relationships between the text and backgrounds and to answer the given questions. In this thesis, we propose to investigate evidence-based text VQA, which involves designing models that can provide reasonable evidence for their predictions, thus improving the generalization ability.