An End-to-End Basis-Spline Based Text Spotting Network for Scene Text
Ningli Xu, Zhichao Lian · 2020
Scene text spotting is considered as an essential step for many applications in text-based visual question answering (VQA), such as auto scoring system, image caption retrieval and keyword spotting (KWS). The performance of a text-based VQA system is largely based on its text detection accuracy. Since irregular text accounts for large proportion in outdoor text-based VQA system, traditional text spotting algorithms for handling regular text has been disappointing. The recent text detection algorithms are mainly divided into two categories. They are regression-based methods and segmentation-based methods. Regression-based methods cannot deal with perfectly with the text of various layout while segmentation-based methods require much computation cost to post-process the pixel-wise prediction result. Thus, we propose a basis-spline based text spotting model for robust irregular text spotting. We propose a basis-spline text representation method with few parameters to represent irregular text layout. With this novel basis spline curve representation, our model can precisely fit the text layout of various types. Finally, we propose a new text-based VQA dataset including images collected from a real-world computer operator qualification exam. Based on this dataset, we implement an auto scoring system and other state-of-the-art recognition algorithms that automatically provides a score corresponding to a student's feedback given a set of standard answers.