Synthesizing Scene Text Images for Recognition with Style Transfer
Haoran Liu, Anna Zhu · 2019
Majority of the existing datasets for tasks like scene text recognition only include a few thousand images with a very limited words and characters. Therefore, it cannot meet the need of the typical deep learning based text recognition approaches. In the same time, although the standard synthetic datasets usually comprise millions of scene text images, which means that the data distribution of the small target datasets cannot be learned very well. We propose a word image generating method called Synth-Text Transfer Network to solve these problems. It has the capability of estimating and simulating the distribution of target datasets. Synth-Text Transfer Network utilizes a style transfer approach to synthesis images with arbitrary text content with preserving the texture of the referenced style image in the target dataset. The large amount of synthesized images can help to alleviate the overfitting problem and improve the accuracy in latter scene text image recognition tasks. In addition, our proposed method is flexible and fast, which has a relatively fast speed among regular style transfer approaches.