Balanced Synthetic Data for Accurate Scene Text Spotting

Previous approaches for scene text detection or recognition have already achieved promising performances across various benchmarks. There are a lot of superior neural network models to choose from to train the desired classifiers. Besides concentrating on designing loss functions and neural network architectures, number and quality of dataset are key to using neural networks. In this paper we propose a new method for synthesizing text in natural scene images that takes into account data balance. For each image we obtain regions normal based on depth and regions information. After choosing a text from text resource, we blend the text in the original image by using the homography matrix of original region contours and mask contours where we put text directly in. Especially, the text source is obtained by a specific loss function which reflects the distances of current characters' distribution and target characters' distribution. Text detection experiments on standard dataset ICDAR2015 and augmented dataset demonstrate that our method of balanced synthetic dataset gets an 84.5% F-score which achieves 2% increase than the result of standard dataset and is also higher than synthetic dataset without balance. Training on balanced synthetic datasets achieves great improvement of text recognition than on some public standard recognition datasets and also performs better than synthetic datasets without balance.