Improving text detection by generating images with curved text instances

Abstract

Modern text detection algorithms rely on deep neural networks, which are trained on labeled datasets to achieve high performance. Despite the increasing popularity of text detection, accurate detection of text in natural images remains a challenging problem due to variations in text size, shape, color, and font. In particular, curved text instances present a unique challenge for detection algorithms, yet they are seldom found in existing text detection datasets. In this paper, we present an approach to improve curved text detection performance by generating synthetic images with curved text instances and polygon bounding regions as annotations. We train a deep neural network-based text detector on these synthetic images and evaluate its performance on test sets. Our findings highlight the importance of utilizing diverse and realistic datasets for training robust text detection systems.