基于矫正网络的场景文本识别应用与研究

Application and research of scene text recognition based on correction network

场景文本在文字识别(Optical Character Recognition,OCR)领域一直是个难题,因此受到学术界的广泛关注。场景文本通常包括透视文本、弯曲文本、定向文本等。目前大多深度学习方法都不能够很好的识别这些不规则的文本,特别是严重变形的文本。针对上述问题,本文提出了一种迭代思想的矫正网络用于场景文本的识别,这种网络是一种端到端无需额外字符级注释的可训练网络。该矫正网络通过迭代细化的方式,逐步达到最优矫正。其中参数变换采用薄板样条(Thin Plate Spline,TPS)参数变换,自适应的进行图像变换,进而提高后序识别网络的识别性能。通过在大量公共数据集上进行的实验,证明了本文方法的有效性,特别是在不规则文本上的实验,证明了该方法有着较好的鲁棒性和准确性。

Scene text has always been a difficult problem in the field of Optical Character Recognition(OCR),so it has been paid much attention in academic circles.Scene text usually includes perspective text,curved text,oriented text,etc.At present,most deep learning methods are not able to recognize these irregular texts,especially severely distorted texts.To solve the above problems,this paper proposes an iterative correction network for scene text recognition,which is an end-to-end trainable network without additional character level annotation.The correction network reaches the optimal correction step by step through the iterative refinement method.Thin Plate Spline(TPS) parameter transformation is used to transform the image adaptively,thus improving the recognition performance of the sequential recognition network.Experiments on a large number of public data sets demonstrate the effectiveness of the proposed method,especially on irregular text.It is proved that the proposed method has good robustness and accuracy.