一种轻量级的不规则场景文本识别模型  

A Lightweight Model for Irregular Scene Text Recognition

作　　者：产世兵 刘宁钟[1] 沈家全 CHAN Shi-bing;LIU Ning-zhong;SHEN Jia-quan(School of Computer Science and Technology,Nanjing University of Aeronautics and Astronautics,Nanjing 211106,China)

机构地区：[1]南京航空航天大学计算机科学与技术学院,江苏南京211106

出　　处：《计算机技术与发展》2020年第11期20-24,29,共6页Computer Technology and Development

基　　金：国家自然科学基金(61375021)。

摘　　要：场景文本识别是近年来极具挑战性的任务,不同于规则的文档文本图像,场景图像中的文本具有形态多变和弯曲等特点,识别起来很有难度。该文提出了一种轻量级的场景文本识别模型(ISTR-LW),不同于现有的场景文本识别模型具有参数量大的缺点,该模型在特征序列提取中引入了经过改变后的轻量级网络PeleeNet,不仅大幅度减少了模型的参数量,还加快了网络预测的速度;在循环网络层中获取标签分布时,引入了Dense Block模块,加快了网络训练的收敛速度;在获取最终识别结果时,引入了注意力机制,获得需要关注的重点区域,提高了模型文本识别的准确度;引入了薄板样条插值转换,通过修正不规则的文本,改善了不规则的文本识别率低的问题。ISTR-LW模型是一个端到端的文本识别模型,在Synth90K、Street View Text和ICDAR等公开数据集上进行了实验,取得了不错的效果。Scene text recognition is a challenging task in recent years.Unlike regular document text image,the text in scene image has the characteristics of changeable shape and bending,so it is quite difficult to recognize.A lightweight model for irregular scene text recognition(ISTR-LW)is proposed.Different from the existing scene text recognition model,which has a large number of parameters,we introduce the changed lightweight network PeleeNet into the feature sequence extraction of the model,which not only greatly reduces the number of parameters of the model,but also speeds up the network prediction.The Dense Block module is introduced to obtain the label distribution in the recurrent neural network,which greatly accelerates the convergence of the network.The attention mechanism is introduced to obtain the final recognition results,which improves the accuracy of model text recognition.The thin-plate spline transformation improves the low accuracy rate of irregular text by correcting irregular text.ISTR-LW model is an end-to-end text recognition model.Experiments are carried out on Synth90k,Street View Text,ICDAR and other public data sets to obtain better results.

关 键 词：场景文本识别 卷积神经网络 轻量级网络 循环神经网络 空间变换网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]