基于语义分割和卷积神经网络的数显表识别算法研究

Research on digital display meter recognition algorithm based on semantic segmentation and convolutional neural network

针对传统数显表读数识别算法适用性差、抗噪能力弱等问题,提出了一种适用于小数据集的基于语义分割和卷积神经网络(CNN)的数显表读数识别算法。该算法通过融合残差网络的U-net实现数字区域定位,利用卷积神经网络实现数字识别。采用mnist数据集预训练模型,使用真实表盘数字图片进行微调,建立适用于多种类、有背景噪声条件下的数显表识别模型。利用家用水表图片构建的测试数据集对算法进行验证。实验结果表明,数字区域定位分割结果的平均IoU为99.76%,160张水表读数识别准确率为100%,单张图片识别用时350.59 ms,满足工程应用需求。

To solve the problems of poor applicability and weak noise immunity of traditional digital display meter reading recognition algorithms,a reading recognition algorithm of digital display meter based on semantic segmentation and Convolutional Neural Network(CNN)for small data sets is proposed.The algorithm realizes digital region location by fusing U-net of residual network,and uses picture classification algorithm CNN for digital recognition. The mnist data set is used to pre-train the model,and then the real dial digital pictures are used to fine tune,so as to establish the recognition model of digital display meter under the condition of multiple kinds and background noise. The test data set constructed by household water meter pictures is used to verify the algorithm. The results indicate the average IoU of digital region positioning and segmentation experiment is 99.76%,the recognition accuracy of 160 water meter reading is 100%,and the recognition time of single picture is 350.59 ms,which meets the needs of engineering application.