不规则场景文本的识别方法

Recognition Method of Irregular Scene Text

场景文本识别是近年来极具挑战性的任务,针对自然场景中密集文本形态多变和弯曲导致识别困难的特点,提出面向场景图像文本的矫正与识别方法。首先利用矫正网络矫正场景文本,然后将校正后的文本输入到双分支网络模块提取图像特征进行融合,双分支模块分别利用InceptionV2和可变卷积代替普通卷积,获取不同的感受域范围,并增加调整卷积核的方向向量,促使卷积核形态更贴近文本形状,最后通过嵌入注意力的双向门控循环单元获取不同权重的文本语义信息。实验结果表明,该模型在ICDAR2013、ICDAR2015和CUTE80数据集上取得了令人信服的结果,尤其在弯曲变形文本CUTE数据集上表现最好,准确率高达89.54%,相比传统方法提高了近1.9%,说明该模型可有效识别扭曲变形的文本信息。

Scene text recognition is a very challenging task in recent years,aiming at the difficulty of recognition due to the variability and curvature of dense text in natural scenes.A correction and recognition method for scene image text is proposed.First,use the correction network to correct the scene text,and then input the corrected text into the dual-branch network module to extract image features for fusion.The dualbranch module uses InceptionV2 and variable convolution instead of ordinary convolution to obtain different ranges of receptive fields.Increase and adjust the direction vector of the convolution kernel to make the shape of the convolution kernel closer to the shape of the text.Finally,the textual semantic information with different weights is obtained through the two-way gated loop unit embedded in the attention.Experimental results show that the model has achieved convincing results on the ICDAR2013,ICDAR2015,and CUTE80 datasets,especially on the CUTE dataset of curved texts,with an accuracy rate of up to 89.54%,which is nearly improved compared to traditional methods.1.9%,the model can effectively identify distorted text information.