Digital dial plate recognition device and edge coordination monitoring system
By using a lightweight digital recognition neural network model and an edge collaborative monitoring system, the problem of limited resources at the edge terminal is solved, enabling efficient and accurate dial recognition, suitable for non-intrusive dial reading in smart grids.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANGHAI INST OF MICROSYSTEM & INFORMATION TECH CHINESE ACAD OF SCI
- Filing Date
- 2026-01-27
- Publication Date
- 2026-06-12
AI Technical Summary
Existing technologies struggle to efficiently recognize digital dials on resource-constrained edge devices, and lightweight network architectures lack the flexibility to adapt to different hardware platforms, resulting in insufficient recognition accuracy.
A lightweight digital recognition neural network model is adopted, combined with a hybrid convolution strategy and depthwise separable convolution technology, to design a digital dial recognition device suitable for edge scenarios, and the model parameters are adaptively adjusted and optimized through an edge collaborative monitoring system.
It achieves efficient and accurate dial recognition on resource-constrained edge terminals, reduces computational complexity and parameter quantity, adapts to different hardware platforms, and is suitable for non-intrusive dial reading in smart grids.
Smart Images
Figure CN122200680A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of smart grid technology, and in particular to a digital dial recognition device and an edge collaborative monitoring system. Background Technology
[0002] With the rapid development of IoT, edge computing, and deep learning technologies, smart grids are evolving from traditional one-way power supply systems into digital infrastructures with bidirectional information interaction, real-time status awareness, and autonomous control capabilities. In this transformation process, low-cost, non-intrusive intelligent upgrades to a large number of existing monitoring devices in the power grid (such as mechanical energy meters, pointer meters, and LCD digital meters without communication interfaces) have become a key link in enhancing the overall sensing capabilities of the power grid.
[0003] Early dial readings relied on manual inspection, which was inefficient and prone to errors. Subsequently, methods based on traditional image processing and machine learning (such as MSER region detection and SVM classification) were used for automatic reading, which reduced labor costs but lacked robustness in complex field environments. In recent years, deep learning-driven digit recognition technologies have significantly improved accuracy, with representative models such as HRC-mCNNs and FCSRN performing excellently under laboratory conditions. However, these models generally have a large number of parameters and are computationally intensive, making them difficult to deploy on resource-constrained edge terminals (such as low-power MCUs). To adapt to edge scenarios, researchers have begun to explore lightweight network architectures (such as MobileNet and ShuffleNet) and dedicated compression techniques, but most solutions still lack the flexibility to adapt to different hardware platforms and have not been structurally optimized for the characteristics of digit recognition tasks (such as simple character structure, few categories, and significant local features). Summary of the Invention
[0004] The technical problem to be solved by the present invention is to provide a digital dial recognition device and an edge collaborative monitoring system that can meet the actual needs of non-intrusive dial reading in smart grids.
[0005] The technical solution adopted by the present invention to solve its technical problem is: to provide a digital dial recognition device, comprising:
[0006] The acquisition module is used to acquire images of the electricity meter's dial.
[0007] A recognition module is used to input the dial image into a lightweight digital recognition neural network model to obtain the dial reading of the electricity meter; wherein, the lightweight digital recognition neural network model includes:
[0008] An initial convolutional layer is used to extract low-level visual features from the dial image;
[0009] The main feature extraction part includes sequentially connected features. An RDSblock module employing a hybrid convolution strategy is used to perform feature enhancement and dimensionality compression on the low-level visual features.
[0010] The classification head is used to classify based on low-level visual features after feature enhancement and dimensionality compression to obtain the meter reading.
[0011] The hybrid convolution strategy refers to the fact that some convolutional layers in the RDSblock module use standard convolutional layers, while others use depthwise separable convolutional layers, and hyperparameters are introduced. Specifies the proportion of depthwise separable convolutional layers used in all RDSblock modules.
[0012] The RDSblock module includes a first convolutional layer and a second convolutional layer connected in sequence. A shortcut connection is provided between the input of the first convolutional layer and the output of the second convolutional layer. When the input size and output size of the RDSblock module are inconsistent, the shortcut connection is a projection shortcut connection. When the input size and output size of the RDSblock module are consistent, the shortcut connection is an identity shortcut connection.
[0013] In the backbone feature extraction section, a max pooling layer is set after both the first and last RDSblock modules.
[0014] The depth-separable convolutional layer includes:
[0015] The depthwise convolution part is used to process each channel of the input using an independent filter;
[0016] The pointwise convolution part is used to fuse the output feature maps of the depthwise convolution part to extract cross-channel information.
[0017] The technical solution adopted by this invention to solve its technical problem is: to provide an edge collaborative monitoring system, comprising:
[0018] The data perception layer includes multiple terminals deployed on-site, each equipped with the aforementioned digital dial recognition device. When the confidence level of the digital dial recognition device is less than a threshold, the dial image is uploaded.
[0019] The functional layer is equipped with the aforementioned digital dial recognition device, used to recognize the received dial image and return the recognition result to the terminal of the data perception layer. The functional layer is also used to train a lightweight digital recognition neural network model adapted to the terminal hardware resources based on data from a local experience database, and to send weight parameters to the terminal of the data perception layer after the accuracy requirements are met.
[0020] The transport layer is used to implement the communication connection between the data sensing layer and the functional layer using narrowband Internet of Things;
[0021] The number of parameters in the lightweight digital recognition neural network model of the digital dial recognition device mounted on the terminal is smaller than the number of parameters in the lightweight digital recognition neural network model of the digital dial recognition device configured in the functional layer.
[0022] The lightweight digital recognition neural network model of the digital dial recognition device mounted on the terminal adjusts its parameters based on the recognition results fed back from the functional layer.
[0023] The technical solution adopted by this invention to solve its technical problem is: to provide a digital dial recognition method, including the following steps:
[0024] Acquire an image of the electricity meter's dial;
[0025] The dial image is input into a lightweight digital recognition neural network model to obtain the dial reading of the electricity meter; wherein, the lightweight digital recognition neural network model includes:
[0026] An initial convolutional layer is used to extract low-level visual features from the dial image;
[0027] The main feature extraction part includes sequentially connected features. An RDSblock module employing a hybrid convolution strategy is used to perform feature enhancement and dimensionality compression on the low-level visual features.
[0028] The classification head is used to classify based on low-level visual features after feature enhancement and dimensionality compression to obtain the meter reading.
[0029] The technical solution adopted by the present invention to solve its technical problem is: to provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-mentioned digital dial recognition method.
[0030] The technical solution adopted by the present invention to solve its technical problem is: to provide a computer-readable storage medium on which a computer program is stored, wherein the computer program, when executed by a processor, implements the steps of the above-mentioned digital dial recognition method.
[0031] Beneficial effects
[0032] Due to the adoption of the above technical solutions, the present invention has the following advantages and positive effects compared with the prior art: The lightweight digital recognition neural network model in the present invention is designed specifically for edge-side digital recognition tasks, and integrates residual connection structure and depthwise separable convolution technology to significantly reduce the number of parameters and computational complexity; by introducing two adjustable hyperparameters to control the number of network layers and the proportion of depthwise separable convolution modules respectively, adaptive adaptation to MCUs with different storage capacities and computing power levels can be achieved. Attached Figure Description
[0033] Figure 1 This is a schematic diagram of a lightweight digit recognition neural network model in the first embodiment of the present invention;
[0034] Figure 2 This is a schematic diagram of the RDSblock module in the first embodiment of the present invention;
[0035] Figure 3 This is a schematic diagram of a depth-separable convolutional layer in the first embodiment of the present invention. Detailed Implementation
[0036] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.
[0037] The first embodiment of the present invention relates to a digital dial recognition device, comprising:
[0038] The acquisition module is used to acquire images of the electricity meter's dial.
[0039] The recognition module is used to input the dial image into a lightweight digital recognition neural network model (TRDS-Net) to obtain the dial reading of the electricity meter.
[0040] The lightweight digit recognition neural network model in this embodiment features a simple structure, high computational efficiency, and a small number of parameters. It supports flexible adjustment of hyperparameters to adapt to embedded microcontroller units (MCUs) with different resource levels.
[0041] like Figure 1 As shown, this lightweight digit recognition neural network model includes: an initial convolutional layer, a backbone feature extraction part, and a classification head. Detailed layer parameters of this lightweight digit recognition neural network model are shown in Table 1.
[0042] Table 1. Layer Parameters of Lightweight Digit Recognition Neural Network Model
[0043]
[0044] The initial convolutional layer is a standard one. The convolutional layer is used to initially extract low-level visual features from the dial image. This initial convolutional layer has 8 output channels and halves the size of the input dial image through a convolution operation with a stride of 2, which helps to quickly reduce computational complexity and expand the receptive field.
[0045] The main feature extraction section includes sequentially connected... An RDSblock module employing a hybrid convolution strategy, wherein... These are hyperparameters used to control the network depth. The backbone feature extraction section is used to perform feature enhancement and dimensionality compression on the low-level visual features. A hyperparameter is inserted after both the first and last RDSblock modules in the backbone feature extraction section. Max pooling layers are used to further compress the spatial resolution of feature maps, thereby effectively reducing the computational burden of subsequent layers and increasing the receptive field.
[0046] Each RDSblock module in this embodiment includes a first convolutional layer and a second convolutional layer connected in sequence, and a shortcut connection is provided between the input of the first convolutional layer and the output of the second convolutional layer. This shortcut connection can be a projection shortcut connection or an identity shortcut connection. Figure 2 As shown, when the input and output dimensions of an RDSblock module are inconsistent, a projective quick join is used to achieve dimension alignment; when the input and output dimensions of an RDSblock module are the same, an identity quick join is used. Specifically, a single RDSblock module can be represented as:
[0047] ;
[0048] Given Indicates the first One RDSblock module. Collection Representative and the The weights (parameter set) associated with each RDSblock module, among which... This represents the number of layers in this RDSblock module. (Function) Refers to the activation function, while Defined as a quick link. Furthermore, This represents the residual mapping to be learned.
[0049] The lightweight digit recognition neural network model in this embodiment employs a hybrid convolution strategy in the RDSblock module. Specifically, some convolutional layers in the RDSblock module use standard convolutional layers, while others use depthwise separable convolution (DSC) layers. This design not only reduces computational costs but also maintains necessary model accuracy. To achieve fine-tuning of computational overhead, hyperparameters are introduced. This specifies the proportion of DSC used across all RDSblock modules. It allows for joint adjustment of hyperparameters. (Controlling network depth) and hyperparameters (By controlling the degree of lightweight design), the model accuracy and computational efficiency can be flexibly balanced on embedded platforms with different resource constraints, as shown in Table 2. This design enables the lightweight digital recognition neural network model to have both high recognition accuracy, low parameter count, and strong hardware adaptability. It is particularly suitable for resource-constrained edge digital recognition scenarios in smart grids, and can efficiently and automatically read the numerical information of meters or other instruments from monitoring images.
[0050] Table 2 Resource Usage Table for Cross-Parameter Configuration
[0051]
[0052] The strategy of employing depthwise separable convolution in specific RDSblock modules aims to reduce computational overhead and model size while maintaining necessary model accuracy. For example... Figure 3 As shown, a depthwise separable convolutional layer consists of a depthwise convolutional part and a pointwise convolutional part. Specifically, the depthwise convolutional part processes each channel of the input using an independent filter. Then, the pointwise convolutional part fuses the feature maps output by the depthwise convolutional part by applying a 1×1 convolution to extract cross-channel information. This depthwise separable convolution greatly optimizes the structure of lightweight digit recognition neural network models, reducing the model's size and computational burden. For a standard convolutional layer, its feature map output (considering a stride of 1 and padding of 1) can be represented as follows:
[0053] ;
[0054] in, Indicates the first Layer, First line, number The output of the column convolution. Indicates application to the first The feature map locations are The convolution kernel. This indicates that after the kernel position is shifted, the position is located at... The input feature map at the th position The values on each channel.
[0055] The computational cost of a standard convolutional layer It can be represented as:
[0056] ;
[0057] in, Indicates the size of the convolution kernel. Indicates the number of input channels. Indicates the number of output channels. Indicates the size of the feature map.
[0058] The computational cost of a depthwise separable convolutional layer is the sum of the computational costs of depthwise convolution and pointwise convolution, which can be expressed as:
[0059] ;
[0060] Therefore, the reduction in computational cost achieved by replacing standard convolution with depthwise separable convolution can be expressed as:
[0061] ;
[0062] Compared to standard convolutional layers, the computational cost of depthwise separable convolutional layers is reduced by approximately [percentage missing]. Most convolutional kernels used in lightweight digit recognition neural network models are 3 × 3 in size. If all of them are replaced with depthwise separable convolutional layers, the computational cost is reduced to 1 / 8 to 1 / 9 of the original.
[0063] In this implementation, each convolution operation in the RDSblock module is followed by a batch normalization (BN) layer and a ReLU6 activation function to improve training stability and inference efficiency.
[0064] During the training of a lightweight digit recognition neural network model, the input distribution of each layer within the network shifts as parameters are updated; this phenomenon is known as internal covariate shift. This issue limits the choice of learning rate, leading to instability in the training process and slower convergence. To mitigate this impact, this implementation introduces a batch normalization (BN) layer after each convolutional layer in the RDSblock module. The BN layer normalizes the input distribution of each layer by normalizing each batch of data, effectively stabilizing the distribution of inputs to each layer. This not only allows for a larger learning rate to accelerate training but also provides a certain degree of regularization, thereby significantly improving the model's convergence speed and generalization ability.
[0065] In this implementation, each convolution operation in the RDSblock module is followed by an activation function, specifically the ReLU6 activation function. Similar to the classic ReLU activation function, the ReLU6 activation function is a piecewise linear activation function. During forward and backward propagation, it only requires simple thresholding and multiplication operations, resulting in extremely low computational overhead. This makes it particularly suitable for deployment on embedded devices with limited computing resources and memory.
[0066] Compared to the ReLU activation function, the ReLU6 activation function also possesses activation sparsity, which helps improve the model's generalization ability and further reduces storage and computational burden, while effectively mitigating the gradient vanishing problem. The key difference lies in the fact that the ReLU6 activation function restricts the output value to a bounded interval of [0, 6]. This bounded characteristic has a significant advantage in low-precision inference environments, not only helping to maintain model performance but also significantly improving the accuracy and stability after quantization, thus better supporting efficient deployment at edge environments.
[0067] The second embodiment of the present invention relates to an edge collaborative monitoring system, which adopts a three-layer architecture design to achieve efficient data processing, low-power transmission, and accurate digital recognition. The three-layer architecture design includes a data perception layer, a functional layer, and a transmission layer.
[0068] The data perception layer includes multiple terminals deployed on-site. Each terminal is equipped with a digital dial recognition device and a camera module according to the first embodiment. The camera module periodically captures images of the electricity meter's dial and records the timestamp and device ID. The digital dial recognition device performs preliminary recognition on the dial images captured by the camera module. If the recognition confidence level is lower than a set threshold (e.g., 95%), the original dial image is uploaded to the edge server (i.e., the functional layer). If the recognition confidence level is not lower than the set threshold, only the recognition result is reported to the edge server. In this embodiment, the terminal can use a low-power MCU (e.g., STM32L series) as the main control unit, the camera module is a camera that supports grayscale image output (e.g., OV2640), and the wireless communication module is a LoRa low-power communication module used to establish a connection with the aggregation node.
[0069] The functional layer (i.e., the edge server) is equipped with a digital dial recognition device according to the first embodiment. This device performs high-precision recognition on the received raw dial images uploaded from the terminal and returns the recognition results to the corresponding terminal. The lightweight digital recognition neural network model of the digital dial recognition device configured in the functional layer is trained based on data from a local experience database and adapted to the terminal's hardware resources. After meeting the accuracy requirements, the model sends weight parameters to the terminal.
[0070] The transport layer is used to realize the communication connection between the data sensing layer and the functional layer using narrowband IoT. It can use narrowband IoT technology (LoRa) to build a transmission network consisting of a single access node and multiple aggregation nodes, ensuring that the terminal can randomly access the network at any access point, reducing power consumption and improving coverage. The data of all terminals is finally aggregated at the access node and forwarded to the edge server to achieve reliable data transmission.
[0071] The hyperparameters of the lightweight digital recognition neural network model of the digital dial recognition device configured in the functional layer of this embodiment. With only 11,930 parameters, it achieves a recognition accuracy of 99.41%. The hyperparameters of the lightweight digital recognition neural network model in the digital dial recognition device mounted on the terminal are... The number of parameters was further compressed to 5314, while the recognition accuracy remained at 99.22%. During the supervised phase, the functional layer compares the recognition results uploaded from the terminal with those from the edge server to evaluate model performance, and retrains and updates the model if necessary. The lightweight digit recognition neural network model of the digital dial recognition device mounted on the terminal also adjusts its parameters based on the high-precision recognition results fed back from the edge server to improve subsequent recognition accuracy.
[0072] The following specific example illustrates this implementation method.
[0073] During the initialization phase of the TRDS-Net-based remote monitoring system, the image acquisition parameters of the terminal equipment need to be configured according to the installation location of the electricity meter, lighting conditions, and meter type (such as LCD digital meter, mechanical roller meter, etc.). The system first triggers the camera to capture meter images at a default cycle (e.g., every 15 minutes), and simultaneously records the timestamp and device ID of the capture time. The original image and preliminary recognition results are stored in the local cache and simultaneously uploaded to the edge server.
[0074] In the initial stage, the terminal did not have an effective TRDS-Net model deployed, so all images were processed by a high-performance, stable deep learning model (such as a more complex benchmark model) on the edge server. The model first completed two key tasks: (1) dial area detection - locating the bounding box of the digit display area; (2) digit segmentation and cropping - dividing the multi-digit dial image into individual digit sub-images.
[0075] These cropped single-digit images, along with their real labels, were used to construct a local experience database specific to the terminal. Subsequently, the edge server trained a lightweight TRDS-Net model adapted to the current terminal hardware resources (such as memory and computing power) based on the data in this experience database. Its structure was dynamically determined by hyperparameters α and β (e.g., α=1, β=1, with approximately 5300 parameters).
[0076] After the TRDS-Net model is trained, if its accuracy on the validation set has not yet reached a satisfactory level (e.g., inference confidence <95% for three consecutive times), the system will not distribute the model immediately. Instead, a conservative strategy will continue: the terminal will continuously upload the original images, the edge server will perform high-precision recognition, and the dial area coordinates (ROI coordinates) will be fed back to the terminal. These coordinates only need to be transmitted once or updated periodically (e.g., weekly). The terminal uses these coordinates to fix the cropping area of subsequent images, improving input consistency and creating conditions for local model deployment.
[0077] Once the TRDS-Net trained on the edge meets the accuracy requirements (e.g., validation accuracy ≥ 98.5%), the edge server sends the model weight parameters to the terminal. After receiving the data, the terminal loads the new model and begins digitally cropping the acquired image based on the preset dial area coordinates, then calls the local TRDS-Net to perform edge recognition.
[0078] To ensure the reliability of the model after migration, the system includes a supervision phase: within a few cycles after the initial update of the model parameters (e.g., the first 10 recognitions), the terminal simultaneously uploads the cropped digital image and its local recognition result. The edge server re-performs high-precision recognition on the same image and compares the two results. If the terminal's recognition accuracy reaches a preset threshold (e.g., ≥99%), the supervision is passed, and the terminal enters autonomous operation mode, reporting only structured readings and no longer uploading images; if the threshold is not met, the edge server re-optimizes the model and issues new weights, and the system re-enters the supervision process.
[0079] Subsequently, the supervision phase can be triggered periodically as needed (e.g., monthly or when the network environment changes) to verify the long-term stability of the model. Through a closed-loop mechanism of "edge training - terminal deployment - supervision verification - autonomous operation," this system ensures high recognition accuracy while achieving efficient model deployment and reliable operation, significantly reducing communication overhead and system latency. It is suitable for non-intrusive smart meter reading in large-scale power Internet of Things scenarios.
[0080] The third embodiment of the present invention relates to a digital dial recognition method, comprising the following steps:
[0081] Acquire an image of the electricity meter's dial;
[0082] The dial image is input into a lightweight digital recognition neural network model to obtain the dial reading of the electricity meter; wherein, the lightweight digital recognition neural network model includes:
[0083] An initial convolutional layer is used to extract low-level visual features from the dial image;
[0084] The main feature extraction part includes sequentially connected features. An RDSblock module employing a hybrid convolution strategy is used to perform feature enhancement and dimensionality compression on the low-level visual features.
[0085] The classification head is used to classify based on low-level visual features after feature enhancement and dimensionality compression to obtain the meter reading.
[0086] The hybrid convolution strategy refers to the fact that some convolutional layers in the RDSblock module use standard convolutional layers, while others use depthwise separable convolutional layers, and hyperparameters are introduced. Specifies the proportion of depthwise separable convolutional layers used in all RDSblock modules.
[0087] The RDSblock module includes a first convolutional layer and a second convolutional layer connected in sequence. A shortcut connection is provided between the input of the first convolutional layer and the output of the second convolutional layer. When the input size and output size of the RDSblock module are inconsistent, the shortcut connection is a projection shortcut connection. When the input size and output size of the RDSblock module are consistent, the shortcut connection is an identity shortcut connection.
[0088] In the backbone feature extraction section, a max pooling layer is set after both the first and last RDSblock modules.
[0089] The depth-separable convolutional layer includes:
[0090] The depthwise convolution part is used to process each channel of the input using an independent filter;
[0091] The pointwise convolution part is used to fuse the output feature maps of the depthwise convolution part to extract cross-channel information.
[0092] The fourth embodiment of the present invention relates to an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the digital dial recognition method of the third embodiment.
[0093] The fifth embodiment of the present invention relates to a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the digital dial recognition method of the fourth embodiment.
[0094] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.
[0095] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0096] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction methods implemented in a process. Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0097] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0098] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A digital dial recognition device, characterized in that, include: The acquisition module is used to acquire images of the electricity meter's dial. A recognition module is used to input the dial image into a lightweight digital recognition neural network model to obtain the dial reading of the electricity meter; wherein, the lightweight digital recognition neural network model includes: An initial convolutional layer is used to extract low-level visual features from the dial image; The main feature extraction part includes sequentially connected features. An RDSblock module employing a hybrid convolution strategy is used to perform feature enhancement and dimensionality compression on the low-level visual features. The classification head is used to classify based on low-level visual features after feature enhancement and dimensionality compression to obtain the meter reading.
2. The digital dial recognition device according to claim 1, characterized in that, The hybrid convolution strategy refers to the fact that some convolutional layers in the RDSblock module use standard convolutional layers, while others use depthwise separable convolutional layers, and hyperparameters are introduced. Specifies the proportion of depthwise separable convolutional layers used in all RDSblock modules.
3. The digital dial recognition device according to claim 1, characterized in that, The RDSblock module includes a first convolutional layer and a second convolutional layer connected in sequence. A shortcut connection is provided between the input of the first convolutional layer and the output of the second convolutional layer. When the input size and output size of the RDSblock module are inconsistent, the shortcut connection is a projection shortcut connection. When the input size and output size of the RDSblock module are consistent, the shortcut connection is an identity shortcut connection.
4. The digital dial recognition device according to claim 1, characterized in that, In the backbone feature extraction section, a max pooling layer is set after both the first and last RDSblock modules.
5. The digital dial recognition device according to claim 2, characterized in that, The depth-separable convolutional layer includes: The depthwise convolution part is used to process each channel of the input using an independent filter; The pointwise convolution part is used to fuse the output feature maps of the depthwise convolution part to extract cross-channel information.
6. An edge collaborative monitoring system, characterized in that, include: The data perception layer includes multiple terminals deployed on-site, each terminal being equipped with a digital dial recognition device as described in any one of claims 1-5, which uploads a dial image when the confidence level of the digital dial recognition device is less than a threshold. The functional layer is configured with a digital dial recognition device as described in any one of claims 1-5, used to recognize the received dial image and return the recognition result to the terminal of the data perception layer; the functional layer is also used to train the lightweight digital recognition neural network model adapted to the terminal hardware resources based on data in a local experience database, and to send weight parameters to the terminal of the data perception layer after the accuracy requirements are met. The transport layer is used to implement the communication connection between the data sensing layer and the functional layer using narrowband Internet of Things; The number of parameters in the lightweight digital recognition neural network model of the digital dial recognition device mounted on the terminal is smaller than the number of parameters in the lightweight digital recognition neural network model of the digital dial recognition device configured in the functional layer.
7. The digital dial recognition device according to claim 6, characterized in that, The lightweight digital recognition neural network model of the digital dial recognition device mounted on the terminal adjusts its parameters based on the recognition results fed back from the functional layer.
8. A method for recognizing a digital dial, characterized in that, Includes the following steps: Acquire an image of the electricity meter's dial; The dial image is input into a lightweight digital recognition neural network model to obtain the dial reading of the electricity meter; wherein, the lightweight digital recognition neural network model includes: An initial convolutional layer is used to extract low-level visual features from the dial image; The main feature extraction part includes sequentially connected features. An RDSblock module employing a hybrid convolution strategy is used to perform feature enhancement and dimensionality compression on the low-level visual features. The classification head is used to classify based on low-level visual features after feature enhancement and dimensionality compression to obtain the meter reading.
9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the digital dial recognition method as described in claim 8.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the digital dial recognition method as described in claim 8.