A method, apparatus and equipment for identifying digital electrical meters

Abstract

This invention provides a method, apparatus, and device for identifying digital electrical meters, relating to the fields of machine vision and artificial intelligence. The identification apparatus includes an image acquisition unit, a digital display area detection unit, and a digital recognition unit. The image acquisition unit acquires an image to be detected; the digital display area detection unit performs a detection operation on the display area of ​​the image to be detected and outputs the result to the digital recognition unit; the digital recognition unit receives the display area image from the digital display area detection unit and performs a digital recognition operation. It achieves an accuracy rate of over 90% for common meters. The improved YOLOv5 detection model can simultaneously perform multiple functions, and the pruned model inference speed can be accelerated by more than 30%, while simultaneously reducing the digital meter identification process. Furthermore, this application optimizes the reading of LED meters with digital backgrounds, employing a dual-recognition fusion method for output, reducing erroneous output.

Description

Technical Field

[0001] This invention relates to the fields of machine vision and artificial intelligence, and in particular to a method, apparatus and device for identifying digital electrical meters. Background Technology

[0002] With the continuous construction of the State Grid and the development of urban rail transit, the number of substations and distribution stations is increasing, putting greater pressure on power supply departments for maintenance. Traditional manual power inspection methods are not only inefficient but also fail to guarantee inspection quality. With the country's vigorous development of smart grids, intelligent inspection systems have replaced manual labor in some automated inspection tasks. Existing automated identification of digital display meters mainly falls into two categories: traditional image recognition methods and algorithms based on deep learning models. Traditional image recognition methods first convert the image to grayscale, then perform edge detection and binarization to extract the digital contour, and finally determine the final character region after tilt correction, comparing it with a template for recognition. These methods are significantly affected by factors such as background and lighting conditions, resulting in a still low overall accuracy rate for both inspection and reading. Algorithms based on deep learning models generally include two steps: region detection and text recognition. Text recognition typically uses the common CRNN+CTC method, which, while applicable to various scenarios, suffers from low accuracy and poor practicality.

[0003] While existing digital meter reading algorithms based on object detection have addressed the issues of instability and accuracy to some extent, their accuracy is affected when the camera is tilted, as they rely solely on object detection. The recognition rate for LED digits with backgrounds drops significantly, and they can only recognize a single type of digit, exhibiting poor generalization ability. Although meter reading algorithms for multiple digit types have been researched, these methods are cumbersome and do not address camera tilt issues. Summary of the Invention

[0004] Purpose of the invention: To propose a method, apparatus, and device for identifying digital electrical meters in order to solve the aforementioned problems existing in the prior art.

[0005] Technical solution: Firstly, a method for identifying digital electrical meters is proposed, with the following steps:

[0006] Acquire the initial image of the digital electrical meter for which readings are required;

[0007] The acquired initial image is preprocessed; the preprocessing operation includes at least image scaling and image normalization to obtain a preprocessed image.

[0008] For the preprocessed image, the display area is identified and output as a rectangular prediction box surrounding the display area, which is the first display area image;

[0009] For the first display area image, the improved YOLOv5 model is used to detect the dial and regress key points;

[0010] The key points are used to perform horizontal correction on the displayed font;

[0011] Determine the type of meter, including LED meters and non-LED meters (rotary type, LCD screen type, etc.).

[0012] A. If the current meter is determined to be an LED meter, then perform dynamic threshold binarization processing on the first display area image of the LED meter to generate a binarized mask template of the first display area image.

[0013] The binarized mask template is fused with the first display area image to form a second display area image;

[0014] The SVTR_LCNet lightweight text recognition network is used to perform OCR recognition on the images of the first and second display areas, and the best of the two recognition results is output.

[0015] B. If the current meter is determined to be a non-LED meter, the display area will be input into the SVTR_LCNet lightweight text recognition network, and the output will show a result with a confidence level higher than the predetermined value, at which point the process will end.

[0016] In a further embodiment of the first aspect, YOLOv5 is a model with faster detection speed and higher lightweightness in the YOLO series. However, in this project, considering the impact of speed and multitasking, further improvements were made to the original YOLOv5 model.

[0017] Based on YOLOv5 V6.1, the Focus layer was replaced with a 3-layer 3x3 kernel convolutional module, facilitating model deployment on CPUs. The backbone network uses MobileNet-v3. MobileNet-v3 employs a lightweight attention model with depthwise separable convolutions, an inverse residual structure with linear bottlenecks, and a squeeze and excitation attention structure, using the computationally efficient and high-performance Hard-Swish activation function. To achieve multiple functionalities, a keypoint regression branch was added, inspired by RetinaFace, to regress keypoints in the digit display area. The improved model simultaneously outputs bounding boxes, digit meter categories, and keypoints. To achieve faster speeds, the model was pruned, achieving a 50% pruning rate while maintaining accuracy. Multiple optimization techniques balanced model size, detection speed, and detection accuracy. The lightweight improvements resulted in no significant change in detection accuracy on CPUs, but a substantial increase in inference speed.

[0018] In a further embodiment of the first aspect, the acquired initial image should simultaneously possess at least the following characteristics:

[0019] The image resolution is greater than the predetermined value;

[0020] The meter angle deviation is within the predetermined range;

[0021] The initial image includes pictures or videos in various formats.

[0022] In a further embodiment of the first aspect, the image scaling operation includes:

[0023] Maintain the aspect ratio of the initial image, calculate the scaling ratio of the width and height based on the model input size, and adjust the length and width of the image to multiples of 32 respectively.

[0024] In a further embodiment of the first aspect, the method for identifying digital electrical meters further includes:

[0025] Scale the long side L of the initial image to a multiple of 32, and record the scaled long side as L. S And record the scaling ratio as S;

[0026] The width W of the initial image is scaled proportionally according to the scaling ratio S;

[0027] If the width W is scaled according to the scaling ratio S, then W S If the value is also a multiple of 32, then the image scaling is complete, and the image size is now L. S ×W S ;

[0028] If the width is scaled according to the scaling ratio S, then W S If the number is not a multiple of 32, then the search term is greater than W. S And the nearest multiple of 32 The image size at this time is .

[0029] In a further embodiment of the first aspect, the process of horizontally correcting the displayed font using the key points includes:

[0030] The transmission matrix is ​​calculated using the information of the four corner points of the first display area, and the display area is corrected into a rectangular image on the same plane using transmission transformation.

[0031] The formula for the projection transformation is as follows:

[0032]

[0033] In the formula, u and v represent the left side of the original image, corresponding to the x and y coordinates of the transformed image; where ;

[0034] Transformation matrix It is divided into 4 parts, among which Represents a linear transformation; Used for translation; Generates transmission transformation;

[0035] get:

[0036]

[0037] In a further embodiment of the first aspect, the dynamic threshold binarization process includes:

[0038] Convert the first display area into a grayscale image;

[0039] Apply Gaussian filtering to the grayscale image;

[0040] Statistical analysis of grayscale images' grayscale histograms;

[0041] Smooth the grayscale histogram;

[0042] Calculate the histogram gradient change from back to front, and find the gray values ​​at the first and third changes;

[0043] Calculate the dynamic threshold, which is the average of two grayscale values.

[0044] In a further embodiment of the first aspect, the process of fusing the binarized mask template with the first display area image includes:

[0045] The mask binary image is obtained based on the calculated dynamic threshold, then a closing operation is performed on the mask, and finally the mask image is fused with the display area image to form a new display area image.

[0046] In a further embodiment of the first aspect, the SVTR_LCNet lightweight text recognition network is used to recognize digits. First, a dictionary containing the digits, letters, and symbols to be recognized is designed. The digit recognition model is trained using a training set generated by the digit synthesis method proposed in this application, combined with real samples. The LCNet network employs acceleration strategies such as MKLDNN, but considering hardware compatibility, the Hswish activation function is replaced with a ReLU activation function, reducing computational load while increasing deployment applicability. The LCNet network uses a SE module, which increases inference time on the CPU. The core idea of ​​the SE module is that the network learns feature weights based on the loss. Therefore, L1 regularization is added to the parameters of the SE module during training to make the parameters sparse. During subsequent deployment, some channels are removed using the SE module. To ensure model accuracy, the modified model is fine-tuned.

[0047] Secondly, a digital electrical meter identification device is proposed, comprising an image acquisition unit, a digital display area detection unit, and a digital recognition unit. The image acquisition unit acquires an image to be detected; the digital display area detection unit performs a detection operation on the display area of ​​the image to be detected and outputs the result to the digital recognition unit; the digital recognition unit receives the display area image from the digital display area detection unit and performs a digital recognition operation.

[0048] Thirdly, a digital meter reading recognition device is proposed, comprising a memory device, an input device, an output device, and a processor. The memory device is used to store relevant programs and data; the input device is used to acquire the image to be detected; and the output device is used to output the digital electrical meter reading recognition result. The processor is used to execute the digital electrical meter recognition method as described in the first aspect; when the program is executed, the algorithms of each module are run to realize the digital meter reading.

[0049] Beneficial Effects: The digital electrical meter identification method, apparatus, and device proposed in this invention can achieve an accuracy rate of over 90% for common meters. The improved YOLOv5 detection model can simultaneously perform multiple functions, and the inference speed of the pruned model can be accelerated by more than 30%, while reducing the digital meter identification process. In addition, this application optimizes the reading of LED meters with digital backgrounds, adopting a dual-identification fusion method for output, reducing erroneous outputs. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of a commonly used digital meter.

[0051] Figure 2This is an example of an architecture diagram showing how digital meter readings can be applied to a real-world system.

[0052] Figure 3 This is a flowchart of a digital meter reading method provided in one embodiment.

[0053] Figure 4 This is a diagram of the device composition provided in one embodiment.

[0054] Figure 5 This is a data synthesis flowchart provided in one embodiment.

[0055] Figure 6 This is an example of image processing effect of a digital meter provided in one embodiment. Detailed Implementation

[0056] In the following description, numerous specific details are set forth in order to provide a more thorough understanding of the invention. However, it will be apparent to those skilled in the art that the invention can be practiced without one or more of these details. In other instances, certain technical features well-known in the art have not been described in order to avoid obscuring the invention.

[0057] Figure 1 The common digital meter types are shown, which have been explained in the "Background Technology" section above and will not be elaborated here.

[0058] Figure 2 An architecture 200 of a practical system applicable to the embodiments is shown. For example... Figure 2 As shown, the system is divided into application terminals 201, 202, and 203, and server terminal 204. The application terminals include, but are not limited to, various display devices such as camera-display combination devices, smartphones, and tablets, as well as various combinations thereof. Their main functions are image information input and display of method execution results. Server terminal 204 provides the hardware for computation and logical analysis of the algorithm, processes the input digital meter image information to generate recognition results, and sends them to the application terminals; it is the carrier of the algorithm model.

[0059] Communication between application terminals 201, 202, and 203 and server terminal 204 can be achieved via wired means, such as fiber optic communication, USB universal serial bus, RS232 serial port, etc.; or via wireless means, such as Bluetooth, wireless broadband (WiFi), ultra-wideband (UWB), etc.

[0060] Figure 3 A flowchart 300 of the detection, identification, and inspection method provided in the embodiment is shown, which includes the following steps:

[0061] 301: Obtain an image of the digital meter from which the reading is required. This is typically taken by the application (though remote wireless transmission is not excluded). The image quality should be clear, the meter angle should not be too large, and it should ideally be taken from the front to minimize angular distortion. Here, "image" is a broad term encompassing various image formats and videos.

[0062] 302: Image Preprocessing. This includes image scaling and image normalization. Scaling maintains the original image's aspect ratio by calculating the scaling ratio based on the model's input dimensions and adjusting the image size to a multiple of 32. This improves model inference speed while maintaining accuracy. For example, a 1000*810 original image is too large; direct inference from the original image would significantly increase processing time. Using the long side of 1000 as a reference, scaling to 640 (a scaling ratio of 0.64) results in a width of 518, which is not a multiple of 32. Therefore, a search is conducted for the nearest multiple of 32, 544, which is greater than 518. After scaling proportionally, the width is increased to 544 and padded with the average pixel value. Image normalization involves transforming the original image into a unique standard form (this standard form is invariant to affine transformations such as translation, rotation, and scaling). The normalization implemented in this case is to subtract the mean from the image and then divide by the variance, thus limiting the model input to between 0 and 1.

[0063] 303: Obtain Meter Locations. For the image of the digital meters to be detected, since the meters occupy a small proportion of the image, it is necessary to detect the meter locations first. Meters in the image are detected using an improved YOLOv5s model, and the output format is a rectangular prediction box surrounding the meter region.

[0064] YOLOv5 is a fast and lightweight model in the YOLO series, but in this project, considering the impact of speed and multitasking, further improvements were made to the original YOLOv5 model.

[0065] Based on YOLOv5 V6.1, the Focus layer was replaced with a 3-layer 3x3 kernel convolutional module, facilitating model deployment on CPUs. The backbone network uses MobileNet-v3. MobileNet-v3 employs a lightweight attention model with depthwise separable convolutions, an inverse residual structure with linear bottlenecks, and a squeeze and excitation attention structure, using the computationally efficient and high-performance Hard-Swish activation function. To achieve multiple functionalities, a keypoint regression branch was added, inspired by RetinaFace, to regress keypoints in the digit display area. The improved model simultaneously outputs bounding boxes, digit meter categories, and keypoints. To achieve faster speeds, the model was pruned, achieving a 50% pruning rate while maintaining accuracy. Multiple optimization techniques balanced model size, detection speed, and detection accuracy. The lightweight improvements resulted in no significant change in detection accuracy on CPUs, but a substantial increase in inference speed.

[0066] 304: Detect the display area using an improved YOLOv5s model and provide corner information for the display area. Using the meter location information obtained in the previous step, the meter is cropped out. Then, the optimally trained improved YOLOv5s model is used to detect the digital display area. The output format is a rectangular prediction box surrounding the text area and corner information of the display area.

[0067] The improved YOLOv5s plays two roles in the overall method: 1. To complete the detection of the digital display area; 2. To generate corner information for each digital display area.

[0068] 305: Data Display Area Correction. This mainly involves calculating the transmission matrix using the information from the four corner points of the display area. Transmission transformation is then used to correct the display area into a rectangular image on the same plane, reducing interference from viewing angle changes on OCR recognition. Transmission transformation refers to using the collinearity of the perspective center, image point, and target point, and rotating the projection surface around the trace line by a certain angle according to the transmission rotation formula. This disrupts the original projection ray beam while maintaining the unchanged projection geometry on the projection surface. Transmission transformation projects the image onto a new viewing plane, also known as projection mapping. The general transformation formula is:

[0069]

[0070] The left side of the original image corresponds to the coordinates of the transformed image. ;That Transformation matrix It can be divided into 4 parts. These represent linear transformations, such as scaling, shearing, and rattotion.

[0071] Used for translation. This results in a transmission transformation. Therefore, affine transformations can be understood as special forms of transmission transformations. Rewriting the previous transformation formula yields:

[0072]

[0073] Therefore, given the points corresponding to the transformation, the transformation formula can be derived. Conversely, a specific transformation formula can also yield the transformed image.

[0074] 306: Determining whether the meter is an LED meter is mainly for preparing for the next step.

[0075] 307: Dynamic threshold binarization is performed on the LED display area. Its main function is to generate a binarized mask template for the display area, used to shield the LED background text from affecting recognition. Image binarization involves setting the grayscale value of points in the image to 0 or 255, thus presenting the entire image in a distinct black and white effect. Selecting a suitable threshold will more effectively separate the background and the numbers; this application designs a method for obtaining a dynamic threshold based on the grayscale histogram gradient. Because the LED digital meter display area has the number 8 as a background, issues such as camera and lighting can cause the display area to be close to the background, leading to misidentification of the numbers. To solve this problem, this application proposes a method for obtaining a dynamic threshold for image binarization. First, the obtained digital display area is converted into a grayscale image, and the grayscale histogram of the display area is calculated. Then, a maximum value filter is performed using a kernel of length 3. This process is repeated 10 times to smooth the grayscale histogram. Then, the pixel value between the penultimate peak and the penultimate peak is obtained using the gradient method, which is the threshold for image binarization. This value is used for image binarization, and then morphological opening is performed on the binarized image. This method can effectively shield the background digits from interfering with the displayed digits.

[0076] 308: Obtain the fused image of the mask and the display area; obtain the binary image of the mask based on the calculated dynamic threshold, then perform a closing operation on the mask, and finally fuse the mask image with the display area image to form a new display area image. This operation mainly masks the numbers or characters in the background area, improving the accuracy of OCR recognition.

[0077] 309: Process the original LED display area and the merged LED display area in parallel to form a 2×H×W×3 array, where H refers to the height of the display area and W refers to the width of the display area.

[0078] 310: OCR recognition using an improved SVTR_LCNET network. The improved SVTR_LCNET network is constructed as follows: First, a dictionary containing the numbers, letters, and symbols to be recognized is designed; then, a digit recognition model is trained using a training set generated by the digit synthesis method proposed in this application, combined with real samples. The LCNet network employs acceleration strategies such as MKLDNN, but considering hardware compatibility, the Hswish activation function is replaced with the ReLU activation function, reducing computational load while increasing deployment applicability. The LCNet network uses an SE module, which increases inference time on the CPU. The core idea of ​​the SE module is that the network learns feature weights based on the loss. Therefore, L1 regularization is added to the parameters of the SE module during training to make the parameters sparse. In subsequent deployment, some channels are removed using the SE module. To ensure model accuracy, the modified model is fine-tuned.

[0079] 311: Output model results; For LED data meters, the model simultaneously identifies the original image and the fused image, and selects the optimal result to output based on the two recognition confidence levels. For other types of digital meters, the results are output directly.

[0080] The improved YOLOv5s model structure is as follows: The module in the Backbone is replaced with the lighter IR module from MobileNetV3, as shown in the dashed box in the figure. The activation function is replaced with the high-performance Hard-Swish activation function. A landmark output is added, which can simultaneously supervise detection and correct display regions. This is mainly achieved by adding a landmark branch to the original head, outputting a feature map with the same width and height as the head and 8 channels.

[0081] Figure 4 The diagram shown is a device composition diagram 500 provided in the embodiment, which is a specific implementation of the method and apparatus, and its system composition is as follows.

[0082] 501: Processor, used to execute related programs to implement the above-described text detection, recognition, and verification methods. A processor refers to any computing device capable of running the methods described in this application, including but not limited to microprocessors (MCUs), programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), system-on-a-chip (SoCs), and other digital computing devices and their various combinations.

[0083] 502: Memory, used to store related instruction code, including 5021 the operating system (e.g., Windows, Mac OS, Unix, Linux), 5022 computer programs, and 5023 images and video data required or saved by the running programs. Memory is divided into internal memory and external memory. Internal memory includes, but is not limited to, read-only memory (ROM), programmable memory (PROM), random access memory (RAM), and various combinations thereof. External memory includes, but is not limited to, hard drives, USB flash drives, optical discs, and various combinations thereof. Internal and external memory are used in conjunction.

[0084] 503: Input devices, including 5031 image input devices and 5032 command input devices. 5031 image input devices include various image or video capture hardware devices, including those that transmit data to a cloud server via an image input device. 5032 command input devices include, but are not limited to, external devices capable of inputting commands, such as mice and keyboards.

[0085] 504: Output Device: Its main function is to display the output results of the model, including content verification information and specific content output information. The device mainly includes various displays and their various combinations.

[0086] Figure 5 This is a flowchart illustrating the data synthesis method provided in this application. This application specifically generates a dataset for digital meter readings, including LED data and data of common digit types. For LED data with shadow backgrounds, this application has specifically developed a method using CSS (Cascading Style Sheets) to design display styles for different types of LED digit displays, designing combinations of different background and font colors to generate digit types that meet the requirements. The OCR recognition model is then trained using actual data.

[0087] Figure 6 This is a phased result in the digital meter reading process of the embodiments of this application, and is a summary of the verification of the method.

[0088] First, after the camera acquires the image, the location and type of the meter are obtained using the target detection model. Then, the improved YOLOv5s model is used to obtain the digital region and key point information of the digital region. The data region is corrected using the key points. The dynamic threshold binarization method proposed in this application is used to obtain the mask template. After fusion, a new digital region image is obtained. The two images are simultaneously fed into the OCR recognition model, and the optimal result is selected for output based on the confidence level.

[0089] As described above, although the invention has been shown and described with reference to specific preferred embodiments, it should not be construed as limiting the invention itself. Various changes in form and detail may be made without departing from the spirit and scope of the invention as defined in the appended claims.

Claims

1. A method for identifying digital electrical meters, characterized in that, Includes the following steps: Acquire the initial image of the digital electrical meter for which readings are required; The acquired initial image is preprocessed; the preprocessing operation includes at least image scaling and image normalization to obtain a preprocessed image. For the preprocessed image, the display area is identified and output as a rectangular prediction box surrounding the display area, which is the first display area image; For the first display area image, the dial is detected and key points are regressed using an improved YOLOv5 model; the improved YOLOv5 model is based on YOLOv5, the focus layer is replaced with a 3-layer 3×3 convolutional module, the backbone network adopts MobileNet-v3, and a key point regression branch is added. The key points are used to perform horizontal correction on the displayed font; Determine the type of meter; the marking type includes LED meters and non-LED meters. A. If the current meter is determined to be an LED meter, then perform dynamic threshold binarization processing on the first display area image of the LED meter to generate a binarized mask template of the first display area image. The binarized mask template is fused with the first display area image to form a second display area image; The SVTR_LCNet lightweight text recognition network is used to perform OCR recognition on the first display area image and the second display area simultaneously, and the best of the two recognition results is output. B. If the current meter is determined to be a non-LED meter, the display area is directly sent to the SVTR_LCNet lightweight text recognition network, and the output result with a confidence level higher than the predetermined value is completed, and the process ends. In the SVTR_LCNet lightweight text recognition network, Hswish is replaced with the ReLU activation function, and L1 regularization is added to the parameters of the SE module to achieve sparsity. In subsequent deployment, the SE module is used to delete some channels in the SVTR_LCNet lightweight text recognition network and make fine adjustments.

2. The method for identifying digital electrical meters according to claim 1, characterized in that, The acquired initial image should simultaneously possess at least the following characteristics: The image resolution is greater than the predetermined value; The meter angle deviation is within the predetermined range; The initial image includes pictures or videos in various formats.

3. The method for identifying digital electrical meters according to claim 1, characterized in that, The image scaling operation includes: Maintain the aspect ratio of the initial image, calculate the scaling ratio of the width and height based on the model input size, and adjust the length and width of the image to multiples of 32 respectively.

4. The method for identifying digital electrical meters according to claim 3, characterized in that, Also includes: Scale the long side L of the initial image to a multiple of 32, and record the scaled long side as L. S And record the scaling ratio as S; The width W of the initial image is scaled proportionally according to the scaling ratio S; If the width W is scaled according to the scaling ratio S, then W S If the value is also a multiple of 32, then the image scaling is complete, and the image size is now L. S ×W S ; If the width is scaled according to the scaling ratio S, then W S If the number is not a multiple of 32, then the search term is greater than W. S And the nearest multiple of 32 The image size at this time is .

5. The method for identifying digital electrical meters according to claim 1, characterized in that, The process of using the aforementioned key points to perform horizontal correction on the displayed font includes: The transmission matrix is ​​calculated using the information of the four corner points of the first display area, and the display area is corrected into a rectangular image on the same plane using transmission transformation. The formula for the transmission transformation is as follows: ； In the formula, u and v represent the left side of the original image, corresponding to the x and y coordinates of the transformed image; where ; Transformation matrix It is divided into 4 parts, among which Represents a linear transformation; Used for translation; Generates transmission transformation; get: 。 6. The method for identifying digital electrical meters according to claim 1, characterized in that, The dynamic threshold binarization process includes: Convert the first display area into a grayscale image; Apply Gaussian filtering to the grayscale image; Statistical analysis of grayscale images' grayscale histograms; Smooth the grayscale histogram; Calculate the histogram gradient change from back to front, and find the gray values ​​at the first and third changes; Calculate the dynamic threshold, which is the average of two grayscale values.

7. The method for identifying digital electrical meters according to claim 6, characterized in that, The process of fusing the binarized mask template with the first display area image includes: The mask binary image is obtained based on the calculated dynamic threshold, then a closing operation is performed on the mask, and finally the mask image is fused with the display area image to form a new display area image.

8. A digital electrical meter identification device, used to perform the identification method for a digital electrical meter as described in any one of claims 1 to 7, characterized in that, include: The image acquisition unit is used to acquire the image to be detected; The digital display area detection unit is used to perform display area detection operations on the image to be detected and output the results to the digital recognition unit. The digit recognition unit is used to receive the display area image from the digit display area detection unit and perform digit recognition operations.

9. A digital meter reading recognition device, characterized in that, include: Memory devices used to store related programs and data; Input device, used to acquire the image to be detected; Output device, used to output the recognition results of digital electrical meter readings; The processor is configured to execute the identification method for digital electrical meters as described in any one of claims 1 to 7; when the program is executed, it runs the algorithms of each module to realize the reading of the digital meter.

Images