结合模板匹配和深度神经网络的电能表信息识别

Information recognition method for electricity meter based on template matching and deep neural network

针对真实拍摄电能表图像受光照、污渍及拍摄角度等影响给识别带来的挑战,提出一种结合模板匹配和深度神经网络的电能信息识别方法。利用SIFT特征与模板库中图像进行匹配来获得待测电能表的类型(即高压或低压电能表),利用边缘信息和Hough变换提取出准确的电能表屏幕区域,进一步借助所匹配的标准模板标定信息获得待测图像的屏幕示数及示数标签区域;在此基础上,利用等间隔分割和对标签区域是否存在标签的二分类判定网络来实现示数数字的分割和标签识别,利用数字识别网络识别出示数。所提方法充分利用了模板标定信息,将复杂条件下的示数检测变为简单有效的等距分割,将标签识别由复杂的文本检测和识别任务变为简单高效的二值检测任务,因而具有更好的鲁棒性。实验结果证明了该方法的有效性。

Aiming at the challenge of recognition difficulties has been rising for electricity meter information including readings and labels because electricity meter image taken is susceptible to light,stains,and shooting angles,this paper proposes an information recognition method for electricity meter based on template matching and deep neural network.Firstly,the type of electricity meter including high-voltage and low-voltage energy meter is determined by using SIFT features to match the meter image with template images.Then,the screen area of the meter is accurately extracted by using edge information and Hough transformation.Furthermore,the on-screen readings and labels area of the meter are obtained respectively with aid of matching calibration information of the standard template.On this basis,segmentation tasks of the readings area and labels area are finished by utilizing equal space segmentation method and binary model respectively.Finally,the readings of the meter are recognized by running a digital recognition network.The proposed method makes full use of the template calibrated information in advance and solves the readings detection problem under complex conditions by a simple and effective equidistant segmentation process,and changes complex text recognition to a simple and efficient binary detection task.Therefore,it has better robustness to recognize the reading and text information of electricity meter.Experimental results verify the effectiveness of the proposed method.