Paper detail

Learning Robust Feature Representations for Scene Text Detection

Scene text detection based on deep neural networks have progressed substantially over the past years. However, previous state-of-the-art methods may still fall short when dealing with challenging public benchmarks because the performances of algorithm are determined by the robust features extraction and components in network architecture. To address this issue, we will present a network architecture derived from the loss to maximize conditional log-likelihood by optimizing the lower bound with a proper approximate posterior that has shown impressive performance in several generative models. In addition, by extending the layer of latent variables to multiple layers, the network is able to learn robust features on scale with no task-specific regularization or data augmentation. We provide a detailed analysis and show the results on three public benchmark datasets to confirm the efficiency and reliability of the proposed algorithm. In experiments, the proposed algorithm significantly outperforms state-of-the-art methods in terms of both recall and precision. Specifically, it achieves an H-mean of 95.12 and 96.78 on ICDAR 2011 and ICDAR 2013, respectively.