基于深度学习的任意形状场景文字识别

Arbitrary shape scene text recognition based on deep learning

摘要:

场景文字识别的一个具有挑战性的方面是处理具有扭曲或不规则布局的文字.尤其是侧视文字和曲线文字在自然场景中较为常见，且难以识别.本文提出了一个带有灵活矫正功能的注意力增强网络,将其用于任意形状场景文字识别.此网络由基于卷积神经网络的文字矫正网络和基于注意力增强的识别网络两部分组成.矫正网络自适应地将输入图像中的文字进行矫正，降低识别难度，使基于注意力增强的序列识别网络直接根据矫正后的图像预测字符序列.整个模型可以进行端到端的训练，训练只需要图像和相应的文字真实标签.在各种公开数据集上进行了广泛的实验，包括SVT、ICDAR 2003和CUTE80等数据集，验证了此网络具有优异的性能.

Abstract:

One of the challenging in scene text recognition is to deal with distortions or irregular layout. Especially, perspective text and curved text are common in natural scenes and are difficult to recognize. In this paper, we propose an attention enhanced network with flexible rectification function for Arbitrary shape scene text recognition. The network consists of a text rectification network and an attention enhanced recognition network. The rectification network adaptively rectifies the text in the input image to reduce the difficulty recognition. The recognition network is an attention enhanced sequencetosequence model that predicts a character sequence directly from the rectified image. With end to end training approach, only images and corresponding text labels are required. Extensive experiments have been conducted on a variety of open datasets, including SVT, ICDAR 2003 and CUTE80, and the experimental results shows the proposed network has excellent performance.