Adversarial Rectification Network for Scene Text Regularization

Scene text recognition with irregular layouts is a challenging yet important problem in computer vision. One widely used method is to employ a rectification network before the recognition stage. However, most previous rectification methods either did not consider recognition information or were integrated into end-to-end recognition models without considering rectification explicitly. To overcome this issue, we propose an adversarial learning-based rectification network that integrates transformation (from irregular texts to regular texts) with recognition information into a unified framework. In this framework, we optimize the rectification network with an extended Generative Adversarial Network that competes between rectifier and discriminator, together with the results of a recognizer. To evaluate the rectification performance, we generated a regular-irregular pair set from the benchmark datasets, and experimental results show that the proposed method can achieve significant improvement on the rectification performance with comparable recognition performance. Specifically, the PSNR and SSIM are improved by 0.81 and 0.051, respectively, which demonstrates its effectiveness.

Duke Scholars

Author Kaizhu Huang DKU Faculty