Keywords

1 Introduction

Intelligent meter investment and operation and maintenance cost is low, fast, high efficiency, wide application prospects. The camera reading table mainly includes character positioning, character segmentation and character recognition. The method based on area aggregation runs faster, but is easily affected by low image resolution and noise [1]. The target recognition algorithm matches feature vectors through improved feature matching algorithm, which improves the accuracy but decreases the recognition speed. Template matching based on euler number groups the template images and matches the target image with the template with the same euler number. At the same time, the above methods are easily affected by environmental factors and are not robust.

To improve the performances of classification, several fusion methods have been developed by combining different feature extraction and classification methods. CNN-ELM classifier has been proposed by Zeng, Y. et al. [6] for traffic sign recognition and achieved human level accuracy. Guo, L. et al. [7] have proposed Hybrid CNN-ELM with improved accuracy and validate its performance on MNIST dataset. Gurpinar, F. et al. [8] have replaced ELM to Kernel ELM for classification that minimizes the mean absolute error. Face features have been extracted from a pre-trained deep convolutional network. Kernel extreme learning machines are used for classification. Yoo, Y. et al. [9] proposed a novel fast learning architecture of CNN-ELM. Their core architecture is based on a local image (local receptive field) version of the ELM adopting random feature learning. Weng, Q. et al. [10] combines ELM classifier with the CNN-learned features instead of the fully connected layers of CNN to land-use classification.

The method proposed in this paper combines the characteristics of the classical network structure to build a new type of convolution neural network M_CNN (Modified_CNN) suitable for water meter image recognition. The experiment shows that it can not only recognize the word wheel figure of water meter but also recognize the digital image. And not easily affected by environmental factors. The rest of the paper is organized as follows: Sect. 2 describes the classical neural network used in this paper, Sect. 3 describes the network structure, Sect. 4 is the experiment, and finally the conclusion.

2 Related Work

The classical neural network based on CNN includes LeNet5, AlexNet, GoogLeNet, etc. As an early CNN, LeNet5 adopted ReLU activation function, added Dropout layer and expanded the amount of training data. However, AlexNet has only eight layers. The first five layers are convolution layer, and the last three layers are full connection layer. The network depth is not deep enough, and the discarding pool layer is prone to over-fitting.

GoogLeNet injected the modularization idea into CNN and used the convolution of 1 × 1 to carry out the elevating dimension. Its sparse connection structure improved the adaptability of the network to a variety of complex images, but it took too long to apply it to the calculation of character recognition.

The first four layers of M_CNN proposed in this paper can be regarded as a module. In the future, the structure of the module remains unchanged and only some parameters such as convolution kernel are changed. M_CNN is a convolutional neural network specially constructed for water meter characters, which USES the characteristics and ideas of classic CNN for reference, and more suitable for the use of camera meter reading scene.

M_CNN is improved on the basis of CNN_ON_MNIST [1, 6] and CNN_ON_CIFAR [3] to build a model that can recognize both the number and the small wheel. CNN_ON_CIFAR can effectively identify images in the cifar-105 dataset, which is built for 10 common objects in nature. This network can be used to identify the word wheel image of water meter. CNN_ON_MNIST can effectively recognize handwritten digits, this network can effectively recognize water meter character image.

By combining the two models, a new neural network M_CNN is constructed, which can recognize the word wheel image and character image of water meter at the same time. By using the self-built water meter training set and test set, the training, testing and generalization ability is better.

3 Proposed Model

As shown in Fig. 1, M_CNN draw lessons from the modular design idea, the two convolution layer, one of the biggest pooling layer, a dropout as a module, the network use a total of three such module, the first module number convolution kernels is 32, the second module has 64 convolution kernels, the third module has 128 convolution kernels, the convolution kernel module size is 3 × 3, pooling nucleus is 2 × 2, drouput ratio is 0.25, finally after a flattening all connections, regularization operation again suppress a layer of neurons activate prevent fitting, Finally complete connection layer can output a 128 - dimensional vector, softmax layer will this vector as input, and then calculate the test images respectively belong to the probability distribution of the 10 class, softmax using cross entropy loss function (cross-entro-py) to calculate each forecast loss value size, commonly used loss value calculation function and hinge loss [4], but the hinge loss calculation value is without calibration, it is difficult to decide on all the classes. However, the value of cross-en-tropy can be used to determine all classes, which will give the prediction probability of each class and finally output the prediction result.

Fig. 1.
figure 1

Architecture of M_CNN.

Full size image

4 Experiment Results

4.1 Experimental Setup

This experimental platform is Intel i7-8700 processor, 16G of memory, GTX1080 graphics card, 4G video memory dell workstation, using Tensorflow as the back-end Keras platform to build M_CNN, and training and verification on the data set.

Character acquisition is completed through raspberry PI and macro camera, and character samples are shown in Fig. 2. USB HD macro camera is adopted with 8 million pixels and the maximum resolution is 640 × 480. The word wheel sample is shown in Fig. 3. The image segmentation is processed by OpenCV, and five spinner code plates and four decimal dial are cut out.

Fig. 2.
figure 2

Sample images.

Full size image
Fig. 3.
figure 3

Correlation between accuracy and rescaling factor.

Full size image

The collected digital and pointer samples are divided into 10 categories from 0 to 9. The training set and test set are divided according to 4:1, and are independent of each other without intersection.

For the half-word problem of numeric characters, as shown in Fig. 3, the number shown in the upper part is 0, that is, the number that has been displayed completely recently is read according to the scroll direction of the word wheel, and the specific reading is read out in combination with the reading of the dial pointer. For the roulette half-word problem, according to the empirical table reading method, it is classified into the number of the last time, and then read more accurate Numbers according to the following X0.1, X0.01 and X0.001 dial. This is done for each dial. The last dial is almost always a pointer to an integer because it is the smallest unit indicator.

M_CNN model parameters include activation function, pooling method, pooled core size, discard ratio, classifier, loss function, optimizer, etc. See Table 2 for specific parameters.

Table 1. Data set specification.
Full size table
Table 2. Parameters specification.
Full size table

4.2 Results and Analysis

Based on the self-built data set and training set, the traditional neural network and M_CNN were used for training test respectively, and the results were shown in Table 3 and Table 4. The loss changes and accuracy changes of CNN_ON_MNIST training are shown in Fig. 4. The loss value increases in the test set, indicating that the network basically has no fitting and cannot be effectively identified. The change when CNN_ON_CIFAR is used is shown in Fig. 5. The loss fluctuation on the test set is large, indicating that the trained network cannot be well fitted. Overall, M_CNN showed better performance in terms of loss and accuracy.

Table 3. Comparison of accuracy between traditional neural network and M_CNN.
Full size table
Table 4. Comparison of loss between traditional neural network and M_CNN.
Full size table
Fig. 4.
figure 4

Correlation between accuracy and activation function.

Full size image
Fig. 5.
figure 5

Correlation between accuracy and optimizer.

Full size image

Accuracy and loss values are compared as follows, respectively showing the accuracy and loss values of CNN_ON_MNIST, CNN_ON_CIFAR, ResNet and M_CNN. By comparison, M_CNN has better performance in accuracy and loss values. Although ResNet is a classic classification network, it does not perform well when both the word wheel and the character image are needed, and the training time is much longer than M_CNN.

4.3 Experimental Comparison

In order to solve the imbalance between sample data, the method of data expansion was proposed. It was found from the actual data collection that not every sample had the same number, which would interfere with the learning process.In order to balance the number of samples and avoid too few samples in the training process and poor fitting, re-scaling factor was added in the sample pretreatment.The experiment shows that when the rescaling factor is set as 1/150, the accuracy rate is the highest (See Fig. 4).

Activation function is actually the function relation between the output of upper node and the input of lower node. The common activation function includes Sigmoid, tanh, ReLU, elu and softplus function. The Sigmoid function in deep neural network causes the gradient to disappear when the gradient reverse transmission, which has been less used. Use a different activation function in the penultimate full connection layer. The last full connection layer is a classifier, all other things being equal. In order to avoid chance, each activation function was repeated three times, and the accuracy rate was averaged. It is found that the activation function of elu is slightly better than that of relu.

The optimizer is used to minimize the loss function, for example, SGD, RMSprop, Adagrad, AdaDelta, Adam. When SGD randomly selects the gradient, noise will be introduced, so that the direction of weight updating is not necessarily correct, and the problem of local optimal solution is not solved. RMSprop is often used to train circular neural networks. The effects of Adagrad, AdaDelta and Adam on the model were compared. Experimental results show that the AdaDelta optimizer is more accurate than others. Adadelta is an extended version with greater robustness. Instead of accumulating all the past gradients, it adjusts the learning rate according to the updated movement window. So despite many updates, Adadelta continues to learn.

5 Conclusion

Based on the classic CNN_ON_MNIST and CNN_ON_CIFAR, this paper proposes MC_CNN which can identify water meters. Based on the raspberry PI platform, it can be read and identified by external hardware and software, without the need to transform the inside of the water meter. Compared with the traditional magnetic induction reading can effectively avoid the problem of magnetic disappearance. Compared with pattern matching, it can reduce the complexity of image preprocessing and the error of manual parameter selection. Compared with other neural networks, accuracy and loss values are also better. The optimal rescaling factor, optimizer and activation function were found by experiments. Mature development can be further extended to the identification of electricity meters, heating meters.