A remote meter reading method, terminal, cloud server and storage medium

Abstract

The application relates to the technical field of computer vision and deep learning, and discloses a remote meter reading method, a terminal, a cloud server and a storage medium. By providing a "server-terminal" collaborative architecture with hierarchical deployment and low algorithm power dependence, the self-learning feature extraction capability of a convolutional neural network is utilized to improve the adaptability of meter dial data recognition to complex shooting environments and the recognition accuracy, and meanwhile, high algorithm power consumption operations such as model training and meter dial template preprocessing are placed on the server side, and the terminal only performs lightweight inference and extremely simple data processing, thereby avoiding the terminal from bearing a large amount of algorithm power and storage overhead of model training, and solving the problems of insufficient algorithm power of a low-resource terminal, difficult model deployment and limited storage resources. Furthermore, through lightweight technologies such as model pruning, quantization and knowledge distillation, and in combination with terminal inference optimization strategies, the algorithm power and hardware dependence of the terminal are further reduced, and accurate positioning of a meter dial data change area is realized.

Description

Technical Field

[0001] This application relates to the fields of computer vision and deep learning technology, and in particular to a remote meter reading method, terminal, cloud server and storage medium. Background Technology

[0002] Meter reading refers to the periodic or irregular recording of readings of various devices, systems, or resources, typically used in fields such as electricity and water. Compared to traditional manual meter reading methods, which heavily rely on manpower and suffer from low efficiency, high human error, high labor costs, and numerous blind spots, making them unsuitable for the development needs of smart cities and intelligent energy management, remote meter reading methods are gaining increasing application due to their ability to automatically collect data remotely.

[0003] In the past, the common approach was to replace the existing mechanical meter dials with smart remote meters to support remote meter reading. However, this would result in high equipment procurement and construction costs, as well as a long renovation period. Therefore, the current remote meter reading solutions typically involve adding an IoT terminal with image acquisition and recognition functions to the existing meter dial equipment. This terminal acquires images of the meter dial and obtains the meter readings through image processing.

[0004] However, existing remote meter reading solutions still suffer from high costs and insufficient real-time accuracy, which hinders the application of remote meter reading technology. Summary of the Invention

[0005] This application provides a remote meter reading method, terminal, cloud server, and storage medium. Through edge-cloud layered collaboration and lightweight model optimization, the computing power, storage, power consumption, and communication requirements of the terminal are significantly reduced while ensuring recognition accuracy, thereby achieving low-cost and highly reliable remote meter reading.

[0006] This application provides a remote meter reading method, comprising: acquiring at least two first dial image datasets, at least two first convolutional neural network models, and at least two second convolutional neural network models, wherein the first dial image datasets correspond one-to-one with dial types, each dial type corresponds to one first convolutional neural network model and one second convolutional neural network model, the first dial image datasets contain several dial images with labels, the labels indicating dial region detection boxes, dial reading region detection boxes, and dial readings in the dial images, and the dial images in the first dial image datasets are obtained by acquiring images of dials of corresponding types; pruning redundant convolutional layers and fully connected layers of the second convolutional neural network models and performing weight quantization to obtain a third convolutional neural network model corresponding to each second convolutional neural network model; training the first convolutional neural network model and the third convolutional neural network model with the same dial type as the first dial image datasets, based on the first dial image datasets, to obtain the first model and the second model, wherein... Each dial type corresponds to a first model and a second model. The first model is used to locate the dial area detection box and the dial reading area detection box in the input dial image. The second model is used to identify and output the dial readings in the image of the area where the input dial area detection box is located and the image of the area where the dial reading area detection box is located. The second model is trained by knowledge distillation. Redundant convolutional layers and fully connected layers of the first model are pruned and weighted to obtain a third model corresponding to each first model. All generated second models and third models are sent to the terminal so that the terminal can store the second model and the third model with the corresponding dial type as the target type. The terminal uses the locally stored second model and the third model to identify the dial readings in the collected dial images. The target type is the type of dial in the dial images collected by the terminal. The terminal sends the dial reading results and generates meter reading results according to the received dial reading structure and a preset data structure, which are then pushed to the meter reading management platform.

[0007] A second aspect of this application also provides a remote meter reading method, comprising: receiving a third model and a fourth model sent by a cloud server, wherein the cloud server sends the third model and the fourth model through the remote meter reading method described in the second aspect; determining and storing the second model and the third model with corresponding dial types as target types from the received second model and the third model, wherein the target type is the type of dial in the dial image collected by the terminal; acquiring a dial image to be processed; calling the third model to process the dial image to be processed, obtaining the positioning results of the dial area detection box and the dial reading area detection box in the dial image to be processed; acquiring the image of the area where the dial area detection box is located and the image of the area where the dial reading area detection box is located from the dial image to be processed based on the obtained positioning results; calling the second model to process the acquired image of the area where the dial area detection box is located and the image of the area where the dial reading area detection box is located, obtaining the dial reading result in the dial image to be processed; and sending the dial reading result to the cloud server.

[0008] A third aspect of this application also provides a terminal, comprising: a single-core MCU, a power management module, and an image sensor, an IoT communication module, external storage, and a clock module respectively connected to the single-core MCU; wherein, the single-core MCU is configured to control the image sensor to acquire a dial image to be processed, and, based on the dial image to be processed acquired by the image sensor, cooperate with the IoT communication module and the external storage to implement the method described in the second aspect; the power management module is configured to manage the power supply to the single-core MCU, the image sensor, the IoT communication module, the external storage, and the clock module.

[0009] A fourth aspect of this application also provides a cloud server, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method as described in the first aspect.

[0010] The fifth aspect of this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in the first aspect, or implements the method described in the second aspect.

[0011] The technical solution provided in this application has at least the following advantages: By training models separately for different dial types on a cloud server, the trained models become more adapted to the dial types, improving the accuracy and reliability of the model's reading recognition. Furthermore, compared to general-purpose models, the model used in this embodiment does not need to learn to recognize readings from various dial types. Specifically, the task of dial reading recognition is divided into two sub-tasks: dial area detection box and dial reading area detection box localization, and dial reading recognition within the image of the area containing the dial area detection box and the image of the area containing the dial reading area detection box. Each sub-task has a more refined learning objective, resulting in lower model complexity and computational complexity. Moreover, the terminal only needs to deploy and run the model corresponding to the dial type for which it requires recognition locally, facilitating lightweight model deployment. Furthermore, considering the impact of the two tasks on the final result, a lightweight third convolutional neural network model is first constructed and then trained to obtain the second model. Conversely, the first convolutional neural network model is trained first and then lightweighted. This process fully preserves the learning capabilities of the second model, which directly affects the reading recognition result, ensuring accuracy. Simultaneously, the second model is lightweighted, while the first model learns more complete content and avoids interference with reading recognition. Therefore, further lightweight optimization of the model is possible while maintaining accuracy. This eliminates the need for complex hardware support from the terminal, reducing hardware requirements and enabling real-time, efficient, accurate, and reliable remote meter reading with a low-hardware-requirement terminal. This improves the real-time accuracy of remote meter reading while reducing the cost of the remote meter reading solution, facilitating its widespread application. Attached Figure Description

[0012] One or more embodiments are illustrated by way of example with reference numerals in the accompanying drawings. These illustrations do not constitute a limitation on the embodiments. Elements with the same reference numerals in the drawings are denoted as similar elements. Unless otherwise stated, the figures in the drawings are not to be limited by scale.

[0013] Figure 1 This is a flowchart of a remote meter reading method applied to a cloud server provided in one embodiment of this application; Figure 2 This is a partial step diagram of another process of a remote meter reading method applied to a cloud server provided in another embodiment of this application; Figure 3 This is a partial step diagram of another process of a remote meter reading method applied to a cloud server provided in another embodiment of this application; Figure 4This is a partial step diagram of a remote meter reading method applied to a terminal provided in another embodiment of this application; Figure 5 This is a schematic diagram of the terminal structure provided in another embodiment of this application; Figure 6 This is a schematic diagram of the structure of a cloud server provided in another embodiment of this application; Figure 7 This is a schematic diagram of the system framework corresponding to the remote meter reading method provided in another embodiment of this application. Detailed Implementation

[0014] As described in the background section, existing remote meter reading solutions suffer from high costs and insufficient real-time accuracy.

[0015] Analysis revealed that the aforementioned problems stemmed from at least the following reasons: The terminal needs to be deployed at the location of the meter, which is often situated in complex environments such as outdoors or underground pipe networks. Limited by the installation environment and power supply methods, it faces core constraints including weak computing power, small storage capacity, power sensitivity, and limited communication bandwidth. Furthermore, the meter image acquisition process is susceptible to factors such as angle shifts, uneven lighting, meter stains, and background interference, resulting in complex and variable image features. While deep learning convolutional neural network (CNN) models possess the advantage of self-learning, their large size and high computing power requirements prevent direct deployment on low-resource terminals. Therefore, existing remote meter reading solutions implemented locally on the terminal typically deploy traditional machine vision solutions locally, using rule-driven recognition to identify readings from the meter images. However, this approach is also limited by the constraints of machine vision and struggles to adapt to complex scenarios. Additionally, while existing remote meter reading solutions based on edge-cloud collaboration exist, they rely excessively on cloud processing, leading to high communication and power consumption costs.

[0016] Specifically, when using traditional machine vision methods locally on the terminal to automate dial image recognition and obtain dial readings, the terminal needs to support a combination of algorithms including template matching, edge detection, and Hough transform. This allows for the use of manually designed image feature operators to segment and extract features from the dial image before numerical recognition. In this case, the entire remote meter reading process—image acquisition, processing, and recognition—is deployed on the terminal. This leads to the following problems: 1. High computing power and cost barriers: Traditional machine vision algorithms have high requirements for terminal hardware computing power, requiring high-performance processors and large-capacity storage, resulting in high production costs for terminal devices and making large-scale deployment difficult. 2. Poor environmental robustness; it is not adaptable to interference from real-world scenarios such as shooting angle deviation, uneven lighting, dial stains, and slight occlusion of data areas. It is prone to feature extraction failure, resulting in low numerical recognition accuracy, usually below 85%, which cannot meet the accuracy requirements of batch meter reading. 3. Poor scene adaptability: Since the data collection is based on a rule-driven recognition method, when the dial model, specifications, or layout changes, the feature operators and matching rules need to be redesigned manually, resulting in a long debugging cycle and high maintenance costs. 4. Insufficient positioning accuracy: It cannot accurately locate the area of ​​data change on the dial and is easily affected by the dial background pattern and scale, causing invalid areas to participate in the recognition calculation, further reducing recognition efficiency and accuracy.

[0017] When adopting a "device-side data acquisition - cloud-based centralized identification" architecture, the terminal device only integrates an image sensor to acquire dial images. The raw or simply compressed image data is then uploaded to a cloud server via cellular networks or IoT communication modules. The server runs a visual recognition algorithm to perform numerical analysis and then sends the identification results back to the terminal or meter reading management platform. This leads to the following problems: 1. High communication costs: Even with image compression, the data size of a single dial image still reaches hundreds of KB to several MB. Massive communication traffic is generated when large-scale terminals upload data in batches, resulting in high IoT communication costs. 2. The device consumes a lot of power. The radio frequency power consumption of wireless image transmission is much higher than that of local data processing. Moreover, data transmission is limited by bandwidth and takes a long time. The terminal device's battery is consumed quickly, making it difficult to meet the lifespan requirement of more than 3 years for IoT terminals powered by dry batteries or small lithium batteries. 3. Strong network dependence: In scenarios with weak or interrupted network signals (such as underground pipe networks and remote areas), image data cannot be uploaded in a timely manner, resulting in interruption of meter reading tasks and low system availability; 4. High data processing latency: Due to the influence of network transmission rate and cloud server scheduling, there is a significant delay from image acquisition to obtaining recognition results, making real-time meter reading impossible and unsuitable for scenarios requiring fast response.

[0018] Based on this, this application provides a remote meter reading method, terminal, cloud server and storage medium. Through edge-cloud layered collaboration and lightweight model optimization, the computing power, storage, power consumption and communication requirements of the terminal are greatly reduced while ensuring recognition accuracy, so as to achieve low-cost and highly reliable remote meter reading.

[0019] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the various embodiments of this application will be described in detail below with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details have been presented in the various embodiments of this application to enable readers to better understand this application. However, the technical solutions claimed in this application can be implemented even without these technical details and various changes and modifications based on the following embodiments.

[0020] The division of the following embodiments is for ease of description and should not constitute any limitation on the specific implementation of this application. The various embodiments can be combined with and referenced by each other without contradiction.

[0021] The first aspect of this application provides a remote meter reading method applied to a cloud server. In some embodiments, such as Figure 1 As shown, the remote meter reading method includes the following steps: Step 101: Obtain at least two first dial image datasets, at least two first convolutional neural network models, and at least two second convolutional neural network models. The first dial image dataset corresponds one-to-one with the dial type, and each dial type corresponds to one first convolutional neural network model and one second convolutional neural network model. The first dial image dataset contains several dial images with labels. The labels indicate the dial area detection box, dial reading area detection box, and dial reading in the dial image. The dial images in the first dial image dataset are obtained by acquiring images of the corresponding type of dial.

[0022] Step 102: Prune the redundant convolutional layers and fully connected layers of the second convolutional neural network model and quantize the weights to obtain the third convolutional neural network model corresponding to each second convolutional neural network model.

[0023] Step 103: Based on the first dial image dataset, train a first convolutional neural network model and a third convolutional neural network model with the same dial type as the first dial image dataset to obtain a first model and a second model. Each dial type corresponds to a first model and a second model. The first model is used to locate the dial region detection box and the dial reading region detection box in the input dial image. The second model is used to identify and output the dial reading in the image of the region where the input dial region detection box is located and the image of the region where the dial reading region detection box is located. The second model is trained by knowledge distillation.

[0024] Step 104: Prune the redundant convolutional layers and fully connected layers of the first model and quantize their weights to obtain the third model corresponding to each first model.

[0025] Step 105: Send all generated second and third models to the terminal so that the terminal can store the second and third models corresponding to the target type of the dial, and use the locally stored second and third models to identify the dial readings in the acquired dial images, wherein the target type is the type of the dial in the dial images acquired by the terminal.

[0026] Step 106: Receive the meter reading results sent by the terminal, and generate meter reading results according to the received meter reading structure and the preset data structure, and push them to the meter reading management platform.

[0027] Therefore, by training the model separately for different dial types on a cloud server, the trained model will be more adapted to the dial type, improving the accuracy and reliability of the model for reading recognition. At the same time, compared with a general model, the model used in this embodiment does not need to learn various types of dials for reading recognition. In particular, the task of dial reading recognition is divided into two sub-tasks: dial area detection box and dial reading area detection box localization, and dial reading recognition in the image of the area where the dial area detection box is located and dial reading recognition in the image of the area where the dial reading area detection box is located. The learning objectives of each sub-task are more refined. Therefore, the model complexity and model computation complexity are lower. Moreover, the terminal only needs to deploy and run the model corresponding to the dial type of the dial that it has recognition requirements locally, which is conducive to the lightweight deployment of the model. Furthermore, considering the impact of the two tasks on the final result, a lightweight third convolutional neural network model is first constructed and then trained to obtain the second model. Conversely, the first convolutional neural network model is trained first and then lightweighted. This process fully preserves the learning capabilities of the second model, which directly affects the reading recognition result, ensuring accuracy. Simultaneously, the second model is lightweighted, while the first model learns more complete content and avoids interference with reading recognition. Therefore, further lightweight optimization of the model is possible while maintaining accuracy. This eliminates the need for complex hardware support from the terminal, reducing hardware requirements and enabling real-time, efficient, accurate, and reliable remote meter reading with a low-hardware-requirement terminal. This improves the real-time accuracy of remote meter reading while reducing the cost of the remote meter reading solution, facilitating its widespread application.

[0028] For ease of understanding Figure 1 The steps included in the illustrated embodiment will be described below.

[0029] In step 101, this application does not limit the first convolutional neural network model and the second convolutional neural network model. In some embodiments, at least one of the first and second convolutional neural network models can be any convolutional neural network with ResNet or MobileNet as the backbone network. In some embodiments, the first convolutional neural network model can be a multi-task convolutional neural network with a feature extraction module, a dial positioning branch, and a data region regression branch. In some embodiments, the second convolutional neural network can be constructed according to the corresponding reading method. For example, when using pointer-type reading, a regression-type convolutional neural network can be constructed; and when using digital reading, a classification-type convolutional neural network can be constructed. This is more in line with the reading characteristics under the corresponding reading method and can perform reading recognition more efficiently and accurately. These are just a few examples; they will not be listed here.

[0030] Furthermore, this application does not limit the labeling of the dial image. It can be formed by drawing a dial area detection box and a dial reading area detection box in the dial image and marking the dial reading. Alternatively, it can be achieved by constructing the image and organizing the coordinates of the upper left and lower right corners of the dial area detection box and the dial reading area detection box in the image coordinate system according to a preset format, as well as the label data of the dial reading, etc. The layers will not be described in detail.

[0031] It should be noted that this application does not limit the method of obtaining the labels. They can be obtained through manual annotation, or by using other models that have been verified to be accurate and reliable to process the dial image generation, etc., which will not be listed here.

[0032] It should also be noted that this application does not limit the meter type, and it can be flexibly configured according to meter reading requirements, terminal hardware capabilities, etc. In some embodiments, the meter type may include: electricity meter type, water meter type, and gas meter type. In some embodiments, the meter type may include: pointer type and digital type. Of course, other meter types can also be used. For example, the pointer type can be further divided according to the reading values ​​and their distribution on the dial, etc., which will not be listed here.

[0033] In some embodiments, taking meter dial types including electricity meters, water meters, and gas meters as examples, meter dial images of different types of dials (water meters, electricity meters, gas meters, etc.) and different shooting scenarios (multi-angle, uneven lighting, dial stains, slight occlusion) are collected as samples. These samples are labeled to obtain a dataset. Then, the samples in this dataset are divided according to the meter dial type to obtain at least two first meter dial image datasets. Alternatively, the dataset can be left undivided, and label information indicating the meter dial type can be added. During training, the meter dial type indicated by the sample's label information determines which model to train on. Further details will not be elaborated here.

[0034] In step 102, this application does not limit the pruning of redundant convolutional layers and fully connected layers, nor the implementation method of weight quantization. The pruning of redundant convolutional layers and fully connected layers can be achieved using any channel pruning scheme described in related technologies; weight quantization can be achieved by converting 32-bit floating-point weights to 8-bit integer weights, etc., and will not be listed here.

[0035] By pruning redundant convolutional and fully connected layers and quantizing weights, the second convolutional neural network model can be compressed to less than 1M of parameters, which is more conducive to lightweight model deployment on the terminal.

[0036] In step 103, this application does not limit the training method. In some embodiments, the MSE loss function can be used for optimization during the training of the first convolutional neural network model. In some embodiments, the data in the dataset can be divided into a training set and a validation set. In some embodiments, the dataset can be directly used as the training set. In some embodiments, the data in the dataset is divided into a training set and a validation set, and the IoU ≥ 0.9 of the validation set is set as the stopping condition for training iteration.

[0037] In some embodiments, the second model is trained as follows: a high-precision convolutional neural network model pre-trained on a cloud server; the cloud server uses a third convolutional neural network model as the student model and the pre-trained high-precision convolutional neural network model as the teacher model, guiding the student model training through the soft labels of the teacher model. This allows the inference speed of the third convolutional neural network model to be increased by 3 times while ensuring that the loss in recognition accuracy does not exceed 1%, thus improving efficiency.

[0038] In step 104, this application does not limit the pruning of redundant convolutional layers and fully connected layers, nor the implementation method of weight quantization. The pruning of redundant convolutional layers and fully connected layers can be achieved using any channel pruning scheme described in related technologies; weight quantization can be achieved by converting 32-bit floating-point weights to 8-bit integer weights, etc., and will not be listed here.

[0039] By removing redundant convolutional and fully connected layers and performing weight quantization, the first model is compressed. This reduces the model size to one-quarter of its original size while removing layers with low contribution and ensuring that the positioning accuracy loss does not exceed 0.5%. This further optimizes the model's lightweight design while still meeting the requirements for meter reading recognition, making it easier to deploy a lightweight model on the terminal.

[0040] In step 105, this embodiment does not limit the transmission method of the second and third models. In some embodiments, the second and third models corresponding to the same dial type can be encapsulated together, and all the second and third models can be broadcast to all terminals. Thus, the terminal only needs to match the target type to simultaneously obtain the second and third models, without needing to perform separate matching, improving terminal processing efficiency and enabling the terminal to apply more accurate and reliable models for reading recognition in real time. Of course, the second and third models can also be encapsulated separately, and / or the encapsulated second and third models can be multicast (i.e., sent to terminals whose target type is the dial type corresponding to the model), etc., which will not be listed here.

[0041] At this point, the terminal only stores the model it is adapted to, reducing terminal storage usage and facilitating the deployment of lightweight models on the terminal.

[0042] It should be noted that this application does not limit the terminal receiving the model. In some cases, it can be a terminal in the factory program flashing stage, thus covering terminals before they leave the factory, which is beneficial for product verification before leaving the factory and helps ensure terminal quality. In other cases, it can be a terminal after leaving the factory, in which case the transmission can be achieved through remote OTA upgrade push. Of course, it is also possible for both types of terminals to receive the model, etc., which will not be listed here.

[0043] In step 106, this application does not limit the preset data structure, which can be configured according to the meter reading data format requirements of the meter reading management platform. In some embodiments, the preset data structure includes the following information: device number, dial type, identification value, collection time, data status, etc., which will not be listed here.

[0044] In some embodiments, the data uploaded by the terminal can also be verified to further ensure the reliability of the meter reading data.

[0045] Based on this, in some embodiments, before generating the meter reading results, a correlation check is performed on the received meter reading results according to the association relationship between different terminals; and the mean of the received meter reading results is used to detect whether there are outliers in the received meter reading results. Thus, if the correlation check passes and there are no outliers, the meter reading results are generated according to the received meter reading structure and a preset data structure.

[0046] Of course, the above is merely an example of how to verify data sent by the terminal. In some embodiments, verification can also be implemented as follows: receiving normal watch face images sent by the terminal; randomly selecting at least one normal watch face image and sending it to the terminal, so that when the watch face type in the received normal watch face image is the same as the target type, the terminal can use locally stored second and third models to identify the watch face reading in the normal watch face image and send the identification result to the cloud server; receiving the identification result sent by the terminal and performing cross-machine verification based on the received identification result. This terminal-side verification avoids reading recognition discrepancies caused by hardware, allowing for incremental training of the model using the corresponding normal watch face images that have issues, further improving the model's reliability, etc. These will not be listed individually here.

[0047] In some embodiments, such as Figure 2 As shown, the remote meter reading method may also include the following steps: Step 107: Obtain at least two second dial image datasets, wherein each second dial image dataset corresponds one-to-one with a dial type. The second dial image dataset includes several standard dial images and several rotating dial images. The dials in the rotating dial images have an installation angle relative to the terminal. The dial images in the second dial image dataset are obtained by acquiring images of the corresponding type of dial.

[0048] Step 108: Extract target feature points from the standard dial images in the second dial image dataset, and generate a feature vector for each standard dial image based on the extracted target feature points. The target feature points indicate the dial outline, dial scale distribution, and reading area.

[0049] Step 109: Extract target feature points from the rotating dial images in the second dial image dataset, and generate a feature vector for each rotating dial image based on the extracted target feature points.

[0050] Step 110: Based on the mounting angle of the dial in the rotating dial image in the second dial image dataset relative to the terminal, generate an image correction matrix corresponding to each rotating dial image. The image correction matrix is ​​used to compensate for the mounting angle of the dial in the rotating dial image relative to the terminal to correct the rotating dial image.

[0051] Step 111: Generate a template library based on the feature vectors of the standard dial images in the same second dial image dataset, as well as the feature vectors and image correction matrices of the rotated dial images. Each template library corresponds to a dial type, and the feature vectors and image correction matrices of the same rotated dial image in the template library are associated.

[0052] Step 112: Send all generated template libraries to the terminal so that the terminal can store template libraries with the corresponding dial type as the target type, and use the locally stored third model, fourth model and template library to identify the dial readings in the acquired dial images.

[0053] Therefore, in Figure 1 Based on the illustrated embodiment, a template library is further constructed to complete template preprocessing through a cloud server. This enables the terminal to reduce the need for repeated complex calculations through template matching, thereby improving the real-time performance of terminal reading recognition.

[0054] For ease of understanding Figure 2 The steps included in the illustrated embodiment will be described below.

[0055] In step 107, the standard dial image is either the original dial design or a clear dial image taken from a positive position, thus enabling the construction of standardized feature templates. The rotated dial image is largely the same as the standard dial image, except that the dial in the rotated dial image has an installation angle relative to the terminal. This installation angle can be flexibly configured according to requirements, for example, it can be ±5°, ±10°, ±15°, ±20°, etc., which are typical rotation angles that the dial may have. Thus, the rotated dial image can cover features in some special cases outside of the standardized feature templates, improving the comprehensiveness, reliability, and accuracy of the template library, which is beneficial to improving the accuracy and reliability of reading recognition.

[0056] It should be noted that the acquisition of the second dial image dataset is roughly the same as that of the first dial image dataset. The main difference is that the second dial image dataset does not introduce labels, and dial images with specified characteristics are added to the second dial image dataset.

[0057] In step 108, this application does not limit the target feature points, which can be any feature points indicating the dial outline, dial scale distribution, and reading area. In some embodiments, the target feature points include the dial center, key points of the dial outer contour, scale start point, scale end point, pointer axis, four corner points of the digital area, and reading window boundary points.

[0058] Furthermore, this application does not limit the generation of feature vectors; they can be generated by preset rules or by model processing, etc., which will not be listed here.

[0059] Step 109 is largely the same as step 108, so it will not be described in detail here.

[0060] In step 110, this application does not limit the image correction matrix, which can be any matrix capable of compensating for the mounting angle of the dial relative to the terminal in the rotated dial image. In some embodiments, in order to ensure that the dial remains within the image area after rotation and to avoid offsetting out of bounds, the image correction matrix can be constructed as follows: ; Where θ is the installation offset angle (θ is fixed for each dial; it can be obtained by the cloud server accurately identifying and sending the image to the terminal for storage after the first image acquisition of the dial is uploaded, or by the terminal calculating the effective data area based on the pixel position of that area after extracting the effective data area according to the third model), tx=x0×(1 cosθ)+y0×sinθ,ty=y0×(1 cosθ) x0×sinθ, where x0 and y0 are the x-coordinate and y-coordinate of the center pixel of the dial area in the image, respectively.

[0061] In step 111, this application does not limit the size of the template library. In some embodiments, in order to reduce the processing pressure on the terminal, the size of each feature template library can be controlled to be less than 300KB. For example, clustering can be performed on the obtained feature vectors, and a preset number of templates can be extracted from each class.

[0062] In step 112, in this application, the template library can be sent using the same method as the second and third models, or even sent together with the second and third models. Of course, in some embodiments, the terminal may experience significant load. Therefore, the sending of the template library can also be passive, meaning the terminal determines whether it needs to request the template library from the cloud server based on actual needs. If it needs to obtain the template library, it pulls the template library corresponding to the target type from the cloud server using low-bandwidth communication, thereby flexibly adapting to terminal needs and reducing terminal processing pressure, etc., which will not be listed here.

[0063] In some embodiments, such as Figure 3 As shown, the remote meter reading method may also include the following steps: Step 113: Receive the abnormal dial image sent by the terminal.

[0064] Step 114: Based on the abnormal dial image, determine the dial type corresponding to the abnormal dial image, and generate a label for the abnormal dial image.

[0065] Step 115: Determine the first and second models that need to be updated based on the dial type corresponding to the abnormal dial image. Step 116: Based on the labels of the abnormal dial images, perform incremental training on the first and second models that need to be updated, so as to update the first and second models that need to be updated.

[0066] Step 117: Prune redundant convolutional layers and fully connected layers in the updated first model to update the third model corresponding to the first model that needs to be updated.

[0067] Step 118: Send the updated second model and the updated third model to the terminal.

[0068] Therefore, in Figure 1 Based on the illustrated embodiment, model updates are further introduced, especially the ability to perform incremental training based on abnormal dial images sent by the terminal. This allows the model to be improved in a targeted manner, efficiently identify model defects, and make up for them, thereby further improving the accuracy and reliability of reading recognition.

[0069] For ease of understanding Figure 3 The steps included in the illustrated embodiment will be described below.

[0070] In step 113, this application does not limit the abnormal dial image. It can be a dial image in which the terminal side fails to recognize the reading. This includes dial images in which the positioning fails due to occlusion, poor image quality, etc., and dial images in which the reading result output is problematic due to poor image quality, large changes in dial features, etc.

[0071] In step 114, this application does not limit the determination of the dial type and the generation of the label. It can be obtained by information parsing when the abnormal dial image carries the dial type and label, or it can be achieved by image detection or other methods when the abnormal dial image does not carry relevant information. These methods will not be elaborated here.

[0072] In step 115, since the cloud server contains models deployed on various terminals, it is necessary to find the model that needs to be updated based on the dial type. Simultaneously, since the first model is trained first and then optimized using lightweight methods, incremental training requires finding the first model for incremental training, followed by lightweight optimization to obtain the third model sent to the terminal.

[0073] In step 116, incremental training can be implemented using any incremental training scheme described in the relevant art, which will not be elaborated here.

[0074] Step 117 is largely the same as step 104, so it will not be described in detail here.

[0075] Step 118 is largely the same as step 105, so it will not be described in detail here.

[0076] Correspondingly, the second aspect of this application also provides a remote meter reading method applied to a cloud server. In some embodiments, such as Figure 4 As shown, the remote meter reading method includes the following steps: Step 201: Receive the third and fourth models sent by the cloud server.

[0077] Step 202: From the received second model and third model, determine the second model and third model with the corresponding dial type as the target type and store them, where the target type is the type of dial in the dial image collected by the terminal.

[0078] Step 202: Obtain the dial image to be processed.

[0079] Step 203: Call the third model to process the dial image to be processed, and obtain the localization results of the dial area detection box and the dial reading area detection box in the dial image to be processed.

[0080] Step 204: Based on the obtained positioning results, obtain the image of the area where the dial area detection box is located and the image of the area where the dial reading area detection box is located from the dial image to be processed.

[0081] Step 205: Call the second model to process the image of the area where the dial area detection box is located and the image of the area where the dial reading area detection box is located, and obtain the dial reading result in the dial image to be processed.

[0082] Step 206: Send the dial reading results to the cloud server.

[0083] It is not hard to see that Figure 4 The embodiments shown correspond to the embodiments shown in the first aspect. Figure 4The embodiments shown can be implemented in conjunction with the embodiments shown in the first aspect. The relevant technical details mentioned in the embodiments shown in the first aspect are as follows: Figure 4 The illustrated embodiments remain valid, and to avoid repetition, they will not be described again here. Accordingly, Figure 4 The relevant technical details mentioned in the illustrated embodiments can also be applied to the embodiments shown in the first aspect.

[0084] In step 201, the cloud server sends the third and fourth models via a remote meter reading method as described in any of the first aspects.

[0085] In step 203, the table image to be processed is obtained by real-time image acquisition of the area where the dial is located.

[0086] In step 204, during the invocation of the third model, a frame-by-frame inference + feature reuse optimization strategy can be adopted, that is, the feature map of the dial positioning branch is directly reused to the data region regression branch, which helps to avoid the terminal repeatedly performing convolution calculations, so that the time for positioning processing of a single dial image to be processed can be controlled within 100ms, improving the real-time performance of reading recognition, and thus improving the real-time performance of meter reading.

[0087] In some embodiments, considering that image quality affects the reading recognition effect, the image to be processed can be preprocessed after the model is invoked. For example, in some cases, preprocessing can be performed as follows: color channel conversion and resolution conversion are performed on the dial image to be processed; the color channel converted and resolution-reduced dial image to be processed is filtered with a mean filter with a convolution kernel of 3×3; the pixel values ​​of the filtered dial image to be processed are normalized to update the dial image to be processed. This process involves color channel conversion (e.g., converting to grayscale) to remove redundant color information and scaling the image to a preset low resolution (e.g., 320×240 pixels) to match the terminal's processing power. Furthermore, it replaces computationally complex Gaussian and median filters with mean filtering, performing only a single round of 3×3 kernel mean filtering to eliminate slight salt-and-pepper noise and reduce computational load. Pixel value normalization (e.g., linearly scaling grayscale pixel values ​​from the 0-255 range to the 0-1 range) skips complex mean and variance normalization operations, reducing the terminal's floating-point computation time. After preprocessing, the data size of a single image is ≤30KB. Ultimately, this significantly reduces the terminal's processing load and improves the real-time performance of meter reading.

[0088] In step 206, as described above Figure 2In the illustrated embodiment, the cloud server can send templates to the terminal. Therefore, before calling the second model, it can be determined whether the terminal has cached a template library for the corresponding watch face type. If so, the template library for the target watch face type is determined from the received template library and stored. Target feature points are extracted from the watch face image to be processed, and a feature vector of the watch face image to be processed is generated based on the extracted target feature points. Matching is performed in the locally stored template library based on the feature vector of the watch face image to be processed. If the matched feature vector in the template library is associated with an image correction matrix, an affine transformation is performed on the image of the area where the watch face reading area detection box is located based on the image correction matrix associated with the matched feature vector to update the image of the area where the watch face reading area detection box is located.

[0089] Correspondingly, prior to this, the terminal will also receive a template library sent by the cloud server. Each template library includes the feature vector of a standard dial image, as well as the feature vector and image correction matrix of a rotating dial image. The feature vector is generated based on the target feature points of the dial image. The target feature points indicate the dial outline, dial scale distribution, and reading area. The image correction matrix is ​​used to compensate for the installation angle of the dial relative to the terminal in the rotating dial image to correct the rotating dial image. The feature vector and image correction matrix of the same rotating dial image in the template library are associated.

[0090] In other words, before recognizing the reading result, the image of the area where the detection box of the dial reading area is located can be quickly matched by the template to determine whether there is an angular offset. If there is an offset, the pre-calculated correction matrix in the template library is directly called to complete the image correction. There is no need for the terminal to perform complex affine transformation calculations, which further improves the reading recognition efficiency of the terminal and thus improves the real-time performance of meter reading.

[0091] Furthermore, to reduce the computational complexity of the second model, the image containing the dial scale and pointer can be cropped based on the pixel dimension within the image of the area where the dial reading detection box is located, and resolution conversion can be performed to update the image of the area where the dial reading detection box is located. Thus, through cropping and resolution conversion (e.g., converting to a resolution of 128×64 pixels), the computational power consumption for subsequent numerical recognition and inference is further reduced, improving the real-time performance of meter reading.

[0092] In step 207, as described above Figure 3In the illustrated embodiment, the cloud server can also perform incremental training on the model using abnormal watch face images uploaded by the terminal. Based on this, in some embodiments, if the third model fails to output or the fourth model's output does not fall within a preset range, the terminal will send the watch face data to be processed as an abnormal watch face image to the cloud server. This allows the cloud server to trigger incremental training of the first and second models based on the received abnormal watch face image, thereby updating the third and fourth models.

[0093] As can be seen from the above embodiments, the remote meter reading method provided in this application addresses the problems of high computing power cost and poor environmental robustness of local terminal identification in existing remote meter reading solutions, as well as high communication power consumption and strong network dependence of centralized cloud identification. It offers a layered deployment, low computing power dependency "server-terminal" collaborative architecture. Utilizing the self-learning feature extraction capability of convolutional neural networks, it improves the adaptability and recognition accuracy of dial data recognition in complex shooting environments. Furthermore, it places high computing power-consuming operations such as model training and dial template preprocessing on the server side, while the terminal only performs lightweight inference and simplified data processing. It avoids the large computing and storage overhead of the terminal for model training, and solves the problems of insufficient computing power, difficult model deployment, and limited storage resources of low-resource terminals. Furthermore, through lightweight technologies such as model pruning, quantization, and knowledge distillation, combined with terminal inference optimization strategies, it further reduces the computing power and hardware dependence of the terminal, and achieves accurate positioning of the dial data change area. Combined with a lightweight recognition model, it improves the accuracy and efficiency of numerical recognition, and takes into account the low cost, low power consumption, high robustness and real-time performance of the meter reading system. The dial data can be accurately, in real time and on a large scale, meeting the actual needs of large-scale IoT remote meter reading.

[0094] The steps of the various methods described above are only for clarity. In practice, they can be combined into one step or some steps can be split into multiple steps. As long as they include the same logical relationship, they are all within the scope of protection of this patent. Adding insignificant modifications or introducing insignificant designs to the algorithm or process, but without changing the core design of the algorithm and process, are also within the scope of protection of this patent.

[0095] A third aspect of the embodiments of this application also provides a terminal, such as... Figure 5 As shown, it includes: a single-core MCU, a power management module, and an image sensor, an IoT communication module, external storage, and a clock module, which are respectively connected to the single-core MCU.

[0096] The single-core MCU is configured to control the image sensor to acquire the dial image to be processed, and, based on the dial image acquired by the image sensor, cooperate with the IoT communication module and external storage to implement the method as described in any one of the second aspects; the power management module is configured to manage the power supply to the single-core MCU, image sensor, IoT communication module, external storage and clock module.

[0097] in, Figure 5 The connection methods and functional descriptions of the hardware shown are shown in Table 1 below.

[0098] Table 1. Hardware in the terminal, its connection methods, and functional descriptions.

[0099] It is not difficult to see that this embodiment is a terminal embodiment corresponding to the method embodiment, and this embodiment can be implemented in conjunction with the method embodiment. The relevant technical details mentioned in the method embodiment are still valid in this embodiment, and will not be repeated here to reduce repetition. Accordingly, the relevant technical details mentioned in this embodiment can also be applied to the method embodiment.

[0100] Furthermore, in order to highlight the innovative aspects of this application, no units that are not closely related to solving the technical problems proposed in this application are introduced in this embodiment, but this does not mean that there are no other units in this embodiment.

[0101] The fourth aspect of this application also provides an electronic device, such as... Figure 6 As shown, it includes: at least one processor 601; and a memory 602 communicatively connected to at least one processor 601; wherein the memory 602 stores instructions executable by at least one processor 601, the instructions being executed by at least one processor 601 to enable at least one processor 601 to perform the method described in any of the second aspects.

[0102] The memory 602 and processor 601 are connected via a bus, which may include any number of interconnecting buses and bridges. The bus connects various circuits of one or more processors 601 and memory 602 together. The bus can also connect various other circuits, such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. A bus interface provides an interface between the bus and the transceiver. The transceiver can be a single element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. Data processed by processor 601 is transmitted over a wireless medium via an antenna, which further receives data and transmits it to processor 601.

[0103] Processor 601 is responsible for managing the bus and general processing, and can also provide various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. Memory 602 can be used to store data used by processor 601 during operation.

[0104] It should be noted that, Figure 5 and Figure 6 This is a hardware description only for the terminal and cloud server. Based on the above hardware, the terminal and cloud server can be organized into various logical modules. For example, based on... Figure 5 With the hardware shown, the terminal can form the following logical functional modules to support the implementation of the corresponding methods.

[0105] Image acquisition module: used to acquire the image of the dial to be processed; Lightweight preprocessing module: used to perform the preprocessing of the dial image to be processed as described above; Model caching module: Used for local caching of models and template libraries generated on the cloud server; The localization inference module is used to load and call a third-party model to achieve accurate localization. Data region cropping module: used for image cropping and fast affine correction based on the positioning results; Numerical recognition module: used to load and call the second model to obtain the dial reading results; Reasonableness verification module: Used to verify the recognition results of dial readings and distinguish between valid values ​​and outliers; Data upload module: Used to upload dial readings and abnormal dial images to the cloud server in batches on a periodic basis (of course, it can also upload non-periodicly and in non-batch).

[0106] Accordingly, the cloud server includes the following logical modules: Sample library management module: used to build the first dial image dataset and the second dial image dataset; Dual convolutional neural network model training module: used for model training to obtain the first and second models; Model lightweighting module: used to perform lightweight optimization on the first trained model and to use a knowledge distillation strategy when training the second model; Template preprocessing module: used to build a template library based on the second dial image dataset; Model package push module: Used to package models according to dial type and push the packaged model to the terminal through factory flashing or remote OTA upgrade; Abnormal Sample Receiving Module: Used to receive abnormal dial images uploaded in batches by terminals; Incremental training module: used for incremental training based on abnormal dial images to achieve iterative optimization of the model; Meter reading data management module: used to uniformly verify and correct the meter reading results uploaded by the terminal, and generate meter reading results.

[0107] Therefore, based on the above structural description of the cloud server and terminal, the following can be formed: Figure 7 The remote meter reading system architecture is shown below.

[0108] Above the system architecture is the cloud server layer, which includes: S1: Sample library management module; S2: Dual CNN model training module; S3: Module lightweight processing module; S4: Template preprocessing module; S5: Model incremental training module; S6: Meter reading data management module.

[0109] In the middle layer of the system architecture, there is a communication layer that enables bidirectional data interaction between the upper and lower layers via an IoT communication network, including: T1: Communication technology, such as NB-IoT / LoRa / SigFox (or other low-power wide area networks); T2: Communication protocol, such as MQTT (or other lightweight IoT protocols); T3: Data characteristics, including model packages / template libraries (OTA push), meter reading data / abnormal dial images (≤1KB / piece, batch upload); T4: Data flow: server → terminal (model / template library); terminal → server (dial reading results / abnormal dial images).

[0110] At the lower layer of the system architecture is the terminal layer, which includes: D1: Image acquisition module; D2: Lightweight preprocessing module; D3: Model caching module; D4: Localization and inference module; D5: Numerical recognition module; D6: Reasonableness verification module; D7: Data transmission module.

[0111] Of course, the above is only an example of the division of a logical module and system framework. Other division methods can also be used, which will not be listed here.

[0112] A fifth aspect of this application also provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the above-described method embodiments.

[0113] That is, those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0114] Those skilled in the art will understand that the above embodiments are specific embodiments for implementing this application, and in practical applications, various changes can be made to them in form and detail without departing from the spirit and scope of this application.

Claims

1. A remote meter reading method, characterized in that, Applied to a cloud server, the method includes: Acquire at least two first dial image datasets, at least two first convolutional neural network models, and at least two second convolutional neural network models. The first dial image dataset corresponds one-to-one with the dial type, and each dial type corresponds to one first convolutional neural network model and one second convolutional neural network model. The first dial image dataset contains several dial images with labels, the labels indicating dial area detection boxes, dial reading area detection boxes, and dial readings in the dial images. The dial images in the first dial image dataset are obtained by acquiring images of dials of the corresponding types. Redundant convolutional and fully connected layers of the second convolutional neural network model are pruned and their weights are quantized to obtain the third convolutional neural network model corresponding to each second convolutional neural network model. Based on the first dial image dataset, a first convolutional neural network model and a third convolutional neural network model with the same dial type as the first dial image dataset are trained to obtain a first model and a second model. Each dial type corresponds to the first model and the second model. The first model is used to locate the dial region detection box and the dial reading region detection box in the input dial image. The second model is used to identify and output the dial reading in the image of the region where the input dial region detection box is located and the image of the region where the dial reading region detection box is located. The second model is trained by knowledge distillation. Redundant convolutional and fully connected layers of the first model are pruned and their weights are quantized to obtain the third model corresponding to each of the first models. All generated second and third models are sent to the terminal so that the terminal can store the second and third models with the corresponding dial type as the target type, and use the locally stored second and third models to identify the dial readings in the acquired dial images, wherein the target type is the type of dial in the dial images acquired by the terminal; The system receives the meter reading results sent by the receiving terminal and generates meter reading results according to the received meter reading structure and a preset data structure, which are then pushed to the meter reading management platform.

2. The method according to claim 1, characterized in that, The method further includes: Acquire at least two second dial image datasets, wherein each second dial image dataset corresponds one-to-one with a dial type, and the second dial image dataset includes several standard dial images and several rotating dial images. The dials in the rotating dial images have an installation angle relative to the terminal. The dial images in the second dial image dataset are obtained by acquiring images of the corresponding type of dial. Target feature points are extracted from the standard dial images in the second dial image dataset, and feature vectors are generated for each standard dial image based on the extracted target feature points, wherein the target feature points indicate the dial outline, dial scale distribution, and reading area; The target feature points are extracted from the rotating dial images in the second dial image dataset, and the feature vector of each rotating dial image is generated based on the extracted target feature points. Based on the mounting angle of the dial in the rotating dial image relative to the terminal in the second dial image dataset, an image correction matrix is ​​generated for each rotating dial image, wherein the image correction matrix is ​​used to compensate for the mounting angle of the dial in the rotating dial image relative to the terminal, so as to correct the rotating dial image. Based on the feature vector of the standard dial image in the same second dial image dataset, and the feature vector of the rotating dial image and the image correction matrix, a template library is generated, wherein each template library corresponds to a dial type, and the feature vector and the image correction matrix of the same rotating dial image in the template library are associated. All the generated template libraries are sent to the terminal so that the terminal can store the template library corresponding to the target type of the dial, and use the locally stored third model, fourth model and template library to identify the dial readings in the acquired dial images.

3. The method according to claim 1 or 2, characterized in that, After sending all the generated second and third models to the terminal, the method further includes: Receive abnormal dial images sent by the terminal; Based on the abnormal watch face image, determine the watch face type corresponding to the abnormal watch face image, and generate the label of the abnormal watch face image; Based on the dial type corresponding to the abnormal dial image, determine the first model and the second model that need to be updated; Based on the labels of the abnormal dial images, incremental training is performed on the first and second models that need to be updated, so as to update the first and second models that need to be updated. Redundant convolutional and fully connected layers of the updated first model are pruned and their weights quantized to update the third model corresponding to the first model that needs to be updated. The updated second model and the updated third model are sent to the terminal.

4. The method according to claim 1 or 2, characterized in that, Before generating the meter reading result according to the received dial reading structure and a preset data structure, the method further includes: Based on the association relationship between different terminals, the received dial reading results are verified for correlation. Based on the mean of the received dial readings, detect whether there are outliers in the received dial readings. The step of generating meter reading results based on the received meter reading structure and according to a preset data structure includes: If the correlation verification is passed and no outliers are found, the meter reading result is generated according to the received meter reading structure and the preset data structure.

5. A remote meter reading method, characterized in that, Applied to a terminal, the method includes: The system receives a third model and a fourth model sent by a cloud server, wherein the cloud server sends the third model and the fourth model using the remote meter reading method as described in any one of claims 1 to 4. From the received second model and third model, determine the second model and third model with the corresponding dial type as the target type and store them, wherein the target type is the type of dial in the dial image collected by the terminal; Obtain the watch face image to be processed; The third model is invoked to process the dial image to be processed, and the positioning results of the dial area detection box and the dial reading area detection box in the dial image to be processed are obtained. Based on the obtained positioning results, images of the area where the dial area detection box is located and images of the area where the dial reading area detection box is located are obtained from the dial image to be processed; The second model is invoked to process the image of the area where the dial area detection box is located and the image of the area where the dial reading area detection box is located, so as to obtain the dial reading result in the dial image to be processed. Send the dial reading results to the cloud server.

6. The method according to claim 5, characterized in that, Before invoking the second model, the method further includes: The system receives a template library sent by a cloud server. Each template library includes a feature vector of a standard dial image, as well as the feature vector and an image correction matrix of a rotating dial image. The feature vector is generated based on target feature points of the dial image, which indicate the dial outline, dial scale distribution, and reading area. The image correction matrix is ​​used to compensate for the installation angle of the dial relative to the terminal in the rotating dial image to correct the rotating dial image. The feature vector and the image correction matrix of the same rotating dial image in the template library are associated. From the received template library, determine the template library whose corresponding dial type is the target type and store it; The target feature points are extracted from the dial image to be processed, and the feature vector of the dial image to be processed is generated based on the extracted target feature points. Matching is performed in the locally stored template library based on the feature vector of the dial image to be processed; When the feature vector in the template library is associated with the image correction matrix, an affine transformation is performed on the image of the area where the dial reading area detection box is located, based on the image correction matrix associated with the matched feature vector, to update the image of the area where the dial reading area detection box is located.

7. The method according to claim 6, characterized in that, Before invoking the second model, the method further includes: In the image of the area where the dial reading area detection box is located, the image containing the dial scale and pointer is cropped based on the pixel dimension, and the resolution is converted to update the image of the area where the dial reading area detection box is located.

8. The method according to any one of claims 5 to 7, characterized in that, Before invoking the third model, the method further includes: Perform color channel conversion and resolution conversion on the dial image to be processed; The color channel converted and resolution downsized dial image is filtered using a mean filter with a 3×3 convolution kernel. The pixel values ​​of the filtered dial image to be processed are normalized to update the dial image to be processed.

9. The method according to any one of claims 5 to 7, characterized in that, The method further includes: If the third model fails to output or the fourth model's output does not fall within a preset range, the dial data to be processed is sent to the cloud server as an abnormal dial image. The cloud server then triggers incremental training of the first and second models based on the received abnormal dial image to update the third and fourth models.

10. A terminal, characterized in that, include: A single-core MCU, a power management module, and an image sensor, an IoT communication module, external storage, and a clock module respectively connected to the single-core MCU; The single-core MCU is configured to control the image sensor to acquire the dial image to be processed, and, based on the dial image to be processed acquired by the image sensor, cooperate with the IoT communication module and the external storage to implement the method as described in any one of claims 5 to 9. The power management module is configured to manage the power supply to the single-core MCU, the image sensor, the IoT communication module, the external storage, and the clock module.

11. A cloud server, characterized in that, include: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method as described in any one of claims 1 to 4.

12. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 4, or implements the method as described in any one of claims 5 to 9.

Images