Light character wheel type camera water meter picture recognition method and system

Abstract

The invention discloses a light-weight character wheel type camera water meter picture recognition method and system, and the method is based on an ESP32 development board, and achieves the collection, digital recognition and data transmission of water meter pictures through the connection with a server. According to the method, the neural network is adopted for water meter digital recognition, and model parameters are optimized through forward propagation and back propagation algorithms. According to the invention, the development board directly uses the character wheel identification network to carry out digital identification on the current water meter picture so as to obtain the water meter reading result, the water meter picture does not need to be uploaded to a server for reading identification, and the identification efficiency of water meter reading can be greatly improved; and automatic reading can be realized only by installing the development board on the character wheel type water meter, so that the intelligent degree of the traditional character wheel type water meter can be greatly improved. The system has the characteristics of light weight and high efficiency, and is suitable for the application scene of reading automation of the intelligent water meter.

Description

Technical Field

[0001] The present invention relates to the technical field of water meter reading recognition, and specifically provides a lightweight word wheel type camera water meter picture recognition method and system. Background Art

[0002] Water meters are one of the main instruments for measuring water supply in various industries, and are applied to civil industries such as residential buildings, hotels, and restaurants, as well as a wide range of industrial fields, including mechanical water meters. Among them, mechanical water meters are roughly divided into two types: one is a pointer type water meter with multiple dials, where the readings of the pointers on multiple dials need to be read and multiplied by the range of each dial to obtain the reading result; the other is a word wheel type mechanical water meter (i.e., word wheel type water meter) with a digital display in the form of word wheels, and the current reading of the water meter is obtained by reading the numbers on each word wheel.

[0003] Regarding the reading recognition of word wheel type water meters, many image-based automatic recognition methods have been proposed in recent years. Although traditional methods can achieve certain results in the detection and recognition of word wheel type water meter readings with images as input data, they still have the following technical problems: Currently, it is usually to take pictures of the water meter through a camera installed on the word wheel type water meter, further upload the taken water meter pictures to the server, and finally use the network model deployed on the server to recognize the readings of the water meter pictures to obtain the water meter reading result. However, the server needs to recognize and process a large number of water meter pictures, which affects the recognition efficiency of water meter readings. Summary of the Invention

[0004] (1) Technical Problems to be Solved

[0005] In view of the deficiencies of the prior art, the present invention provides a lightweight word wheel type camera water meter picture recognition method and system, which can solve the above technical problems.

[0006] (2) Technical Solutions

[0007] To solve the above technical problems, the present invention provides the following technical solution: A lightweight word wheel type camera water meter picture recognition method, including the following steps:

[0008] S1. The server transmits water meter picture data to the development board, where the development board is installed on the word wheel type water meter, and a word wheel recognition network is built on the development board;

[0009] S2. The development board uses the water meter picture data to train the neural network model of the word wheel recognition network;

[0010] S3. The development board uses its camera to take pictures of the word wheel type water meter to obtain the current water meter picture;

[0011] S4. The development board uses the trained character wheel recognition network to recognize the numbers in the current water meter picture to obtain the water meter reading result;

[0012] S5. The development board transmits the water meter reading result to the server.

[0013] Preferably, the water meter picture data includes the water meter picture and the marked water meter reading.

[0014] Preferably, in step S2, the development board first locates the number and position of the character wheels on the water meter picture, further performs segmentation processing on the water meter picture to obtain individual single-character wheel pictures, and further uses the individual single-character wheel pictures to train the neural network model of the character wheel recognition network.

[0015] Preferably, in step S2, the training process of the character wheel recognition network is as follows: The development board inputs the individual single-character wheel pictures into the character wheel recognition network, performs forward propagation using the individual single-character wheel pictures to calculate the output of the model, calculates the gradient of the loss function through backpropagation, and updates the model parameters using the gradient descent method; repeat the forward propagation and backpropagation processes until the model converges.

[0016] Preferably, the character wheel recognition network includes a multi-classification prediction module and a digital regularization module. The multi-classification prediction module is used to perform preliminary digital recognition on the single-character wheel picture to obtain a preliminary digital recognition result, and the digital regularization module is used to further regularize the preliminary digital recognition result to obtain the final digital recognition result of the single-character wheel picture.

[0017] Preferably, the multi-classification prediction module includes a convolutional layer, a fully connected layer, and a multi-classification output layer.

[0018] Preferably, the digital regularization module includes a hidden layer and an output layer.

[0019] Preferably, the loss function adopts the mean square error loss function.

[0020] To solve the above technical problems, the present invention provides another technical solution as follows: A lightweight character wheel type camera water meter picture recognition system, including a server and a development board. The development board is installed on the character wheel type water meter, and the development board is built with a character wheel recognition network. The lightweight character wheel type camera water meter picture recognition system is used to execute the above lightweight character wheel type camera water meter picture recognition method.

[0021] Preferably, the development board adopts an ESP32 development board.

[0022] (III) Beneficial effects

[0023] Compared with the prior art, the present invention provides a lightweight word-wheel type camera water meter picture recognition method and system, having the following beneficial effects: By installing a development board on the word-wheel type water meter, a word-wheel recognition network is built on the development board. The development board directly uses the word-wheel recognition network to perform digital recognition on the current water meter picture to obtain the water meter reading result, without uploading the water meter picture to the server for reading recognition, which can greatly improve the recognition efficiency of the water meter reading; and only by installing the development board on the word-wheel type water meter can automatic reading be realized, which can greatly improve the intelligence level of the traditional word-wheel type water meter. Description of the Drawings

[0024] Figure 1 It is a flowchart of the steps of the lightweight word-wheel type camera water meter picture recognition method of the present invention;

[0025] Figure 2 It is a schematic diagram of the lightweight word-wheel type camera water meter picture recognition system of the present invention;

[0026] Figure 3 It is a schematic diagram of the word-wheel recognition network of the present invention (taking a water meter with 6 word-wheels as an example);

[0027] Figure 4 It is a schematic diagram of the multi-class prediction module of the word-wheel recognition network of the present invention;

[0028] Figure 5 It is a schematic diagram of the digital regularization module of the word-wheel recognition network of the present invention;

[0029] Figure 6 It is a schematic diagram of the forward propagation of the word-wheel recognition network of the present invention;

[0030] Figure 7 It is a schematic diagram of the multi-class output layer of the multi-class prediction module of the word-wheel recognition network of the present invention;

[0031] Figure 8 It is a schematic diagram of the fully connected layer of the multi-class prediction module of the word-wheel recognition network of the present invention. Detailed Embodiments

[0032] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0033] The present invention provides a lightweight word-wheel type camera water meter picture recognition method, including the following steps:

[0034] S1. The server transmits water meter picture data to the development board, where the development board is installed on a dial water meter, and a dial recognition network is built on the development board.

[0035] It can be understood that a dial water meter includes multiple dials. By reading the numbers of each dial (i.e., readings: 0-9), the total reading of the dial water meter can be obtained. The water meter picture data includes the water meter picture and the marked water meter reading; the marked water meter reading can be specifically saved in a json file; among them, preferably, when marking the water meter reading on the water meter picture, only the total reading of the water meter needs to be marked, and it is not necessary to mark the readings of each dial. Subsequently, the development board locates the positions of each dial and decomposes the marked total reading to automatically obtain the readings of each dial, thereby reducing the workload of water meter marking and lowering the marking cost. That is, the server initializes the development board, transmits the marked water meter picture and the corresponding marked reading file to the development board, and performs model training on the development board to obtain a trained neural network.

[0036] Preferably, the above-mentioned development board uses an ESP32 development board. The cost of the ESP32 development board is relatively low. The ESP32 development board is provided with a WiFi module, a camera, etc. The ESP32 development board is connected and communicates with the server through the WiFi module; of course, the development board can also use other single-chip microcomputer terminals.

[0037] S2. The development board uses the water meter picture data to train the neural network model of the dial recognition network.

[0038] In step S2, the development board first locates the number and position of the dials on the water meter picture, further performs segmentation processing on the water meter picture to obtain each single-dial picture, and further uses each single-dial picture to train the neural network model of the dial recognition network; the development board can specifically use recognition algorithms such as yolo in the prior art to locate the dials on the water meter picture to determine the number and position of each dial on the water meter picture, and there is no excessive limitation here. Taking a dial water meter with a 6-digit water meter reading as an example, it correspondingly includes 6 dials, and each of the 6 dials corresponds to a digital reading (0-9), thus forming a 6-digit water meter reading result.

[0039] In step S2, the training process of the dial recognition network is as follows: the development board inputs each single-dial picture into the dial recognition network, uses each single-dial picture for forward propagation to calculate the output of the model (i.e., the dial recognition network), calculates the gradient of the loss function through backpropagation, and uses the gradient descent method to update the model parameters; repeat the forward propagation and backpropagation processes until the model converges.

[0040] Preferably, the character wheel recognition network adopts a VSDRNN (very simple digital recognization neural network) deep neural network. The character wheel recognition network includes a multi-classification prediction module and a digital regularization module. The multi-classification prediction module is used to perform preliminary digital recognition on a single character wheel image to obtain a preliminary digital recognition result, and the digital regularization module is used to further regularize the preliminary digital recognition result to obtain the final digital recognition result of the single character wheel image.

[0041] Preferably, the multi-classification prediction module includes a convolutional layer, a fully connected layer, and a multi-classification output layer. Specifically, the multi-classification prediction module may include three convolutional layers, three fully connected layers, and a multi-classification output layer. The processing process of the multi-classification prediction module is as follows: Input a single character wheel image with a size of 30*30. After 3 convolutional operations with a convolutional kernel size of 3*3 (conv3*3), a feature map with a size of 24*24 is obtained. Further, the feature map is flattened into a vector with a length of 576 through a Flatten layer, and a vector with a length of 180 is obtained through the first fully connected layer (Linear) and activated by a relu activation function. Then, it is activated through the second fully connected layer to obtain a vector with a length of 45. The output of the last fully connected layer is connected to the multi-classification output layer. The multi-classification output layer is an output layer including 20 neurons, and each neuron corresponds to a category. The multi-classification output layer uses a softmax function. It can be understood that the multi-classification prediction module of the character wheel recognition network VSDRNN is a single-channel neural network with parameters of float. The multi-classification prediction module is a 20-classification network for recognizing the numbers of single character wheel images. The preliminary digital recognition result, that is, the prediction value of 20-classification, is obtained through the multi-classification prediction module. The 20-classification includes the complete display readings of 10 categories from 0 to 9 and the other 10 categories of "semi-character" readings. "Semi-character" means that the character wheel stays between two numbers. Of course, the multi-classification output layer can also include 40 neurons to be further subdivided into 40-classification, or it can be other numbers of classifications, which are not limited here too much.

[0042] The digital regularization module includes a hidden layer and an output layer. Through the digital regularization module, the "semi-character" readings that may appear in the preliminary digital recognition results obtained by the multi-classification prediction module are further digitally regularized and recognized to improve the recognition accuracy of the water meter readings. Taking a dial water meter with 6 dials as an example, 6 x (which can be 20, 40 or other quantities) classified numbers are input into the digital regularization module, and the output of the digital regularization module is the water meter reading of 10 classifications (0-9). Preferably, the number of the above-mentioned hidden layers is multiple. The hidden layers are stacked layer by layer to gradually extract more and more abstract features. Through multiple hidden layers, more complex features are learned to enhance the prediction ability of the model. In addition, the digital regularization module may also include a fully connected layer for connecting the preliminary digital recognition results of each single dial picture.

[0043] It can be understood that after the output of the forward propagation calculation model of the dial recognition network, the difference between the model output and the true value is measured by the loss function. The gradient of the loss function with respect to the parameters of each layer is calculated layer by layer starting from the output layer through backpropagation. Further, the gradient descent method is used to update the model parameters. The core idea of backpropagation is to transmit the error from the output layer to the input layer layer by layer through the chain rule and calculate the gradient of the parameters of each layer. Then, the parameters are updated to minimize the error between the prediction result and the true value and improve the model accuracy.

[0044] Preferably, the above loss function adopts the mean squared error loss function (MSE), which is a loss function in machine learning and deep learning and is mainly used for regression tasks. It measures the difference between the model prediction value and the true value and quantifies the error by calculating the average of the squared differences between the prediction value and the true value. The loss function is shown in the following formula (1) (taking a dial water meter with 6 dials as an example):

[0045]

[0046] Where is the true value of the i-th digit (i.e., the reading corresponding to the dial), and y i is the predicted value of the i-th digit.

[0047] The backpropagation process of the dial recognition network is as follows 1-10 (calculating the gradient of the loss function with respect to the parameters layer by layer from the output layer to the input layer and updating):

[0048] 1. Gradient of the loss function with respect to the output layer: First, calculate the partial derivative (i.e., the gradient) of the loss function with respect to the predicted value y i of the output layer, as shown in the following formula (2):

[0049]

[0050] This step is the starting point of backpropagation, representing the error between the predicted value and the true value.

[0051] 2. Gradient of the loss function with respect to the hidden layer parameters: Next, calculate the partial derivative of the loss function with respect to the hidden layer parameter w ij as shown in the following equation (3):

[0052]

[0053] where x i is the input of the hidden layer, and is the partial derivative of the output layer with respect to the weight.

[0054] 3. Gradient of the loss function with respect to the hidden layer input: Calculate the partial derivative of the loss function with respect to the hidden layer input x i as shown in the following equation (4):

[0055]

[0056] This step transfers the error from the output layer to the hidden layer.

[0057] And so on, the gradient of the loss function loss with respect to the output of the multi-class prediction module of the character wheel recognition network VSDRNN (i.e., the predicted values of the above 20-classification) can be obtained.

[0058] 4. Gradient of the loss function with respect to the 20-class output: The multi-class prediction module of the character wheel recognition network VSDRNN is a 20-class network for identifying the readings of a single character wheel. Calculate the partial derivative of the loss function with respect to the 20-class output z i as shown in the following equation (5):

[0059]

[0060] where, when j = i, when j ≠ i,

[0061] 5. Gradient of the loss function with respect to the fully connected layer parameters: Calculate the partial derivative of the loss function with respect to the fully connected layer parameter w ij as shown in the following equation (6):

[0062]

[0063] where r i is the input of the fully connected layer.

[0064] 6. Gradient of the loss function with respect to the output of the activation layer (i.e., corresponding to the above relu activation function): Calculate the partial derivative of the loss function with respect to the output r i as shown in the following equation (7):

[0065]

[0066] 7. Gradient of the loss function with respect to the input of the activation layer: Calculate the partial derivative of the loss function with respect to the input \(x\) i of the activation layer as follows:

[0067] When \(x\) i > 0,

[0068] When \(x\) i < 0,

[0069] And so on, obtain the gradient of the loss function \(loss\) with respect to the 576 fully connected layer, which is the gradient of the 24*24 convolutional layer.

[0070] 8. Gradient of the loss function with respect to the convolutional kernel parameters: The gradient of the loss function with respect to the convolutional kernel parameters of the third layer of the character wheel recognition network: It is obtained by performing a convolution operation on the second convolutional layer with the gradient of the third convolutional layer as the convolutional kernel. The corresponding formula is shown in Equation (8):

[0071] Equation (8): Conv2d(26, 26, 24, 24, second convolutional layer, gradient of the third convolutional layer, output result).

[0072] The following uses Equation (9), that is, a 3*3 convolution operation, as an example to explain Equation (8) above:

[0073]

[0074] where \(a\) is the second convolutional layer, \(w\) is the convolutional kernel, and \(z\) is the third convolutional layer.

[0075] Using the definition of convolution, it is easy to obtain the following Equation (10):

[0076] \(z\) 11 = a 11 w 11 + a 12 w 12 + a 21 w 21 + a 22 w 22

[0077] \(z\) 12 = a 12 w 11 + a 13 w 12 + a 22 w 21 + a 23 w 22

[0078] z 21 = a 21 w 11 + a 22 w 12 + a 31 w 21 + a 32 w 22

[0079] z 22 = a 22 w 11 + a 23 w 12 + a 32 w 21 + a 33 w 22 (10)

[0080] Next, calculate the gradient through each w. According to the above equation (10), the following equations (11)-(14) are further obtained:

[0081]

[0082] The above four equations (11)-(14) can be expressed in the form of a matrix convolution as follows in equation (15), that is:

[0083]

[0084] That is, the gradient of the third-layer convolution kernel parameters can be obtained.

[0085] 9. Gradient of the loss function with respect to the convolutional layer: The gradient of the loss function with respect to the second convolutional layer: The third convolutional layer gradient is padded with 2 to obtain a convolutional layer of size 28*28. The third convolutional kernel is flipped, and then a convolution operation is performed to obtain it. The corresponding formula is shown in equation (16):

[0086] Equation (16): Conv2d(28, 28, 3, 3, padded third-layer convolutional layer gradient, flipped third-layer convolutional kernel, output result).

[0087] The following uses equation (17), that is, a 3*3 convolution operation as an example to explain the above equation (16):

[0088]

[0089] Among them, a is the second convolutional layer, w is the convolutional kernel, and z is the third convolutional layer.

[0090] Using the definition of convolution, it is easy to obtain the following equation (18):

[0091] z 11= a 11 w 11 + a 12 w 12 + a 21 w 21 + a 22 w 22

[0092] z 12 = a 12 w 11 + a 13 w 12 + a 22 w 21 + a 23 w 22

[0093] z 21 = a 21 w 11 + a 22 w 12 + a 31 w 21 + a 32 w 22

[0094] z 22 = a 22 w 11 + a 23 w 12 + a 32 w 21 + a 33 w 22 (18)

[0095] Next, calculate the gradients through each a. For example, for the gradient of a11, since in the 4 equations a11 only has a product relationship with z11, the following equation (19) is obtained:

[0096]

[0097] where δ is the gradient of the third convolutional layer.

[0098] For the gradient of a12, since in the 4 equations a12 only has a product relationship with z12 and z11, the following equation (20) is obtained:

[0099]

[0100] Similarly, the following equation (21) can be obtained:

[0101]

[0102] The above formulas (19)-(21) can be represented in the form of a matrix convolution, as shown in the following formula (22):

[0103]

[0104] In the above formula (22), to conform to the gradient calculation, a circle of 0s is filled around the error matrix. At this time, after flipping the convolution kernel and convolving it with the gradient error of backpropagation, the gradient error of the previous time is obtained.

[0105] 10. Parameter update: After calculating the gradients of all parameters through the above steps 1-9, use the gradient descent method to update the parameters of the character wheel recognition network, as shown in the following formula (23):

[0106]

[0107] Among them, η is the learning rate, which is used to control the step size of parameter update.

[0108] It can be understood that the core idea of backpropagation is to transmit the error from the output layer to the input layer layer by layer through the chain rule and calculate the gradients of the parameters of each layer; by continuously iteratively updating the parameters, the loss function is gradually reduced, thereby improving the prediction accuracy of the model. The backpropagation process of the character wheel recognition network VSDRNN of the present invention covers the complete gradient calculation from the output layer to the convolutional layer, ensuring that the model can be efficiently trained and inferred on a resource-constrained single-chip microcomputer.

[0109] In addition, in other embodiments, the character wheel recognition network can also adopt other types of neural network architectures in the prior art, and no excessive restrictions are imposed here.

[0110] S3. The development board uses its camera to take pictures of the dial water meter to obtain the current water meter picture.

[0111] Specifically, the camera can adopt OV2640 or other low-power cameras. Specifically, the development board can set a hardware timer. For example, the photographing task is triggered once every 1 hour. After the timer is triggered, the development board calls the connected camera to take pictures of the current water meter.

[0112] S4. The development board uses the trained character wheel recognition network to recognize the numbers in the current water meter picture to obtain the water meter reading result. It can be understood that the water meter reading result is obtained by combining the digital recognition results of each single dial picture.

[0113] S5. The development board transmits the water meter reading result to the server.

[0114] In addition, as Figure 2As shown in the figure, the present invention also provides a lightweight word-wheel type camera water meter picture recognition system, including a server 1 and a development board 2. The development board 2 is installed on the word-wheel type water meter 3. The development board 2 constructs a word-wheel recognition network. This lightweight word-wheel type camera water meter picture recognition system is used to execute the above-mentioned lightweight word-wheel type camera water meter picture recognition method. For specific details, refer to the above content and no further elaboration will be made here.

[0115] Compared with the prior art, the present invention provides a lightweight word-wheel type camera water meter picture recognition method and system, which have the following beneficial effects: (1) By installing a development board on the word-wheel type water meter, the development board constructs a word-wheel recognition network. The development board directly uses the word-wheel recognition network to perform digital recognition on the current water meter picture to obtain the water meter reading result, without uploading the water meter picture to the server for reading recognition, which can greatly improve the recognition efficiency of the water meter reading; (2) And only by installing the development board on the word-wheel type water meter can automatic reading be realized, which can greatly improve the intelligent level of the traditional word-wheel type water meter; (3) The word-wheel recognition network includes a multi-classification prediction module and a digital regularization module. The digital regularization module learns more complex features of the water meter picture through multiple hidden layers, improving the model's recognition and prediction ability for the word-wheel type water meter reading. The digital regularization module further performs digital regularization recognition on the possible "half-character" readings in the preliminary digital recognition result obtained by the multi-classification prediction module to improve the recognition accuracy of the water meter reading. The digital recognition accuracy of the word-wheel recognition network of the present invention can reach more than 99%; (4) The backpropagation process of the word-wheel recognition network covers the complete gradient calculation from the output layer to the convolutional layer, ensuring that the model can be efficiently trained and inferred on the resource-constrained development board, thereby ensuring efficient and accurate completion of the word-wheel type water meter reading recognition; (5) Realize the intelligent recognition of the word-wheel type water meter reading on the low-cost and limited-resource ESP32 development board, greatly reducing the equipment cost; (6) It is not necessary to label the readings of individual word-wheels of the water meter, only the total reading of the water meter needs to be labeled, and the labeling workload is only 10-20% of that of ordinary intelligent recognition algorithms, reducing the labeling cost.

[0116] It should be noted that the term "including", "comprising" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "including a..." does not exclude the existence of another identical element in the process, method, article or device including the said element.

[0117] Although embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. The image recognition method of a lightweight word-wheel type camera water meter, characterized in that It includes the following steps: S1. The server transmits water meter picture data to the development board, where the development board is installed on a dial water meter, and the development board constructs a dial recognition network; S2. The development board uses the water meter picture data to train the neural network model of the dial recognition network; S3. The development board uses its camera to take pictures of the dial water meter to obtain the current water meter picture; S4. The development board uses the trained dial recognition network to perform digital recognition on the current water meter picture to obtain the water meter reading result; S5. The development board transmits the water meter reading result to the server.

2. The lightweight character wheel type camera water meter picture recognition method according to claim 1, wherein: The water meter picture data includes a water meter picture and the marked water meter reading.

3. The lightweight character-wheel type camera water meter image recognition method according to claim 2, wherein: In step S2, the development board first locates the number and position of the dials on the water meter picture, further performs segmentation processing on the water meter picture to obtain individual single-dial pictures, and further uses each of the single-dial pictures to train the neural network model of the dial recognition network.

4. The lightweight character-wheel type camera water meter picture recognition method according to claim 3, characterized in that: In step S2, the training process of the dial recognition network is as follows: The development board inputs each of the single-dial pictures into the dial recognition network, performs forward propagation using each of the single-dial pictures to calculate the output of the model, calculates the gradient of the loss function through backpropagation, and updates the model parameters using the gradient descent method; repeat the forward propagation and backpropagation processes until the model converges.

5. The lightweight character-wheel type camera water meter picture recognition method according to claim 4, characterized in that: The dial recognition network includes a multi-classification prediction module and a digital regularization module. The multi-classification prediction module is used to perform preliminary digital recognition on the single-dial picture to obtain a preliminary digital recognition result, and the digital regularization module is used to further perform digital regularization on the preliminary digital recognition result to obtain the final digital recognition result of the single-dial picture.

6. The lightweight character wheel type camera water meter picture recognition method according to claim 5, characterized in that: The multi-classification prediction module includes a convolutional layer, a fully connected layer, and a multi-classification output layer.

7. The lightweight character wheel type camera water meter picture recognition method according to claim 6, characterized in that: The digital regularization module includes a hidden layer and an output layer.

8. The lightweight character-wheel type camera water meter image recognition method according to claim 4, wherein: The loss function uses the mean square error loss function.

9. Lightweight character wheel type camera water meter image recognition system, characterized in that: It includes a server and a development board. The development board is installed on a dial water meter, and the development board constructs a dial recognition network. The lightweight dial water meter picture recognition system is used to execute the lightweight dial water meter picture recognition method according to any one of claims 1-8.

10. The lightweight character wheel type camera water meter picture recognition system according to claim 9, characterized in that: The development board uses an ESP32 development board.

Images