微型电脑应用

提出了一种基于改进的LeNet-5卷积神经网络的识别方法。通过加装摄像头和通信线路的方式，实时采集图像信息，并对图像进行部分预处理。引入Gabor滤波器、ReLU-Softplus函数、SVM分类器等优化传统LeNet-5模型，并根据图像数据的不均衡性，运用Grid Loss函数优化CNN网络，由此，实现燃气表自动化识别方法的构建。在Caffe深度学习框架下进行实验测评，结果表明该方法整体的识别准确性高达99.60%、整个样本集及单幅字码的训练总时间均优于其他识别方法，且对于不完整表码字的识别准确率也达到了99.21%。

This paper proposes a recognition method based on improved LeNet-5 convolutional neural network. The image information is collected in real time by adding cameras and communication lines, and the image is processed. The traditional LeNet-5 model is optimized by introducing Gabor filter, ReLU-Softplus function and SVM classifier, and according to the imbalance of image data, the CNN network is optimized by using Grid Loss function, so as to realize the construction of gas meter automatic recognition method. The experimental evaluation under the Caffe deep learning framework shows that the overall recognition accuracy of this method is 99.60%. The total training time of the whole sample set and single code is better than other recognition methods, and the recognition accuracy of incomplete table codes is 99.21%.