基于CNN的水表指针读数识别及STM32实现方案设计

为了提高卷积神经网络对于水表指针读数识别的准确率,同时实现将卷积神经网络移植到STM32单片机中运行,使用了包含2 913张水表指针图片的数据集对GoogLeNet和ResNet-18进行迁移学习和测试,其中GoogLeNet的测试集准确率为89.37％,ResNet-18的测试集准确率为93.24％.借鉴于ResNe...

Recognition of water meter pointer reading based on CNN and design of STM32 implementation scheme

In order to improve the accuracy of convolutional neural network for water meter pointer reading recognition and realize the operation of convolutional neural network transplanted into STM32 microcontroller, A data set containing 2913 water meter pointer pictures was used for transfer learning and testing of GoogLeNet and ResNet-18. The accuracy of GoogLeNet test set was 89.37%, and that of ResNet-18 test set was 93.24%. Based on the jumping connection idea of Resnet-18 model, the method of feature fusion of high and low levels is used. On the premise that the size of receptive field remains unchanged, the 7×7 large convolution kernel is replaced by 3 3×3 small convolution kernels in series to reduce the number of network parameters, reduce the depth of the network, and speed up the convergence of the network during training. Then, a convolutional neural network model with higher accuracy and faster convergence for water meter pointer reading is proposed. The test set accuracy of this model is 95.11%. In order to overcome the difficulty of extremely limited storage resources of STM32 microcontroller and further reduce the network size and the number of network parameters on the condition of ensuring high accuracy, the test set accuracy of the designed model is 91.51%. The training process is completed using MATLAB deep learning toolbox on PC, and the generated onnx model is only 948KB in size. The running footprint of RAM is 437.14KB.