Noise-Robust Wagon Text Extraction Based on Defect-Restore Generative Adversarial Network

Wagon text extraction mainly depends on manual identification of relevant information, which is laborious, time consuming, monotonous and error-prone. To address this concern, we develop a two-stage wagon text extraction system based on the combination of transfer learning and defect-restore generative adversarial network (GAN). Considering the limited number of wagon images and vast computer resource required, wagon texts are first detected via refined connectionist text proposal network. In this study, we focus on text recognition and develop a generic strategy comprising an adversarial learning. The generator is made up of encoder-decoder-encoder sub-networks, enabling it to learn discriminative representations from the intermediate layer. In addition, by adding a random mask block prior to the generator and rebuilding the encoder with full convolution structure, the proposed strategy is able to significantly reduce the noise sensitivity of the encoder. The generated images are of high quality, even when the image is masked. The propose strategy yields an accuracy of 97.76% on 2682 real-world test sub-images, remarkably outperforms the prior arts. Furthermore, the proposed defect-restore GAN model performs well on synthetic images contaminated by three types of noise, which evidently verifies its robustness.