基于Sauvola算法和神经网络的图像自适应二值化方法  

Adaptive Image Binarization Method Based on Sauvola Algorithm and Neural Network

作　　者：李艺杰 邹坤霖 孙炜[1,2,3] 王耀南 梁桥康[1,2,3] LI Yi-je;ZOU Kun-lin;SUN Wei;WANG Yao-nan;LIANG Qiao-kang(College of Electrical and Infommation Engineering,Hunan University,Changsha 410082,China;National Engineering Iabornatory for Robot Vision Perception and Control Technology,Changsha 410082,China;Hunan Key Laboratory of Inelligent Robot Technology in Electronic Manufacturing,Changsha 410082,China).)

机构地区：[1]湖南大学电气与信息工程学院,湖南长沙410082 [2]机器人视觉感知与控制技术国家工程实验室,湖南长沙410082 [3]电子制造业智能机器人湖南省重点实验室,湖南长沙410082

出　　处：《测控技术》2020年第8期62-69,75,共9页Measurement & Control Technology

基　　金：国家自然科学基金项目(61673163);湖南省科技计划项目(2017XK2102);湖南省重点实验室开放基金(IRT2018003)。

摘　　要：二值化处理是文档图像OCR识别中非常重要的一个环节。针对目前主流Sauvola二值化算法存在参数需要人为确定、效果受设定的参数影响等问题,提出了基于Sauvola算法和神经网络的图像自适应二值化方法。在针对Sauvola算法的参数进行分析的基础上,阐述了开窗大小r与参数k对二值化结果的影响。使用SWT算法实现了开窗大小r的计算,将二值化视为对像素点的二分类问题,根据Sauvola算法的思想提取窗内的灰度均值、灰度标准差及当前像素点灰度值作为特征向量,并搭建全连接神经网络进行训练,训练后的神经网络可为每个窗生成特定的参数k,并判定各个像素点的二值化结果。实验结果表明,相较于使用人工精确调节的全局参数k的Sauvola算法,所提出的方法具有更好的自适应性和二值化效果,且具有较好的实用性和普适性。Binarization is very important for image OCR recognition.The parameters of the existing binarization methods based on Sauvola need to be determined artificially,and the effect is affected by the set parameters.An adaptive image binarization method based on Sauvola algorithm and neural network is proposed.On the basis of analyzing the parameters of Sauvola algorithm,the influences of window size and parameters on the binarization result were expounded.The SWT(Stroke Width Transform)algorithm was used to calculate the window size.The binarization was regarded as the binary classification problem of pixel points.According to the ideas of Sauvola algorithm,the gray mean value,gray standard deviation and current pixel gray value in the window were estracted as the feature vectors,and a fully connected neural network was built for training.The trained neural network can generate specific parameters for each window and determine the binarization result of each pixel.The experimental results show that the proposed method has better adaptability,binarization effect,practicability and universality than the Sauvola algorithm that adjusts global parameters artificially.

关 键 词：二值化处理 Sauvola算法 图像识别 神经网络 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]