基于卷积神经网络的真实场景下水表读数识别

Automatic Recognition Method of Water Meter Degree in Real Scene Based on Convolutional Neural Network

摘要: 为实现真实场景下老式机械水表读数的自动识别，提出一种基于卷积神经网络的水表数字识别方法。对原始的水表图像进行预处理，通过坐标定位以及分割，获取单个字符的数据集；使用椒盐噪声、旋转角度的方法扩增数据集；而后再依次经过灰度化、滤波、局部阈值分割等操作，对图像进行细节的处理；在TensorFlow的深度学习框架下搭建LeNet卷积神经网络模型，选取卷积核为3 * 3的卷积层，将处理好的数据放入到改进后的LeNet网络模型中进行训练，最终实现字符的分类识别。实验结果表明：该方法在训练集上的准确率达到99.97%，在测试集上的准确率达到98.36%，为真实场景下水表读数的自动识别提供了可能。

Abstract: In order to realize the automatic recognition of old mechanical water meter reading in real scene, a new method of water meter digital recognition based on convolutional neural network was proposed. The original water meter image was preprocessed, and the data set of a single character was obtained by coordinate positioning and segmentation. The method of salt-and-pepper noise and rotation angle was used to amplify the data set. Then through the grayscale, filtering, local threshold segmentation and other operations, the details of the image are processed. In the deep learning framework of TensorFlow, a LeNet convolutional neural network model was built, the convolutional kernel was selected as a 3 * 3 convolutional layer, the processed data was put into the improved LeNet network model for training, and the character recognition was finally realized. Experimental results show that the accuracy of this method is 99.97% in the training set and 98.36% in the test set, which provides the possibility for the automatic recognition of water meter reading in the real scene.