A method and system for identifying a meter reading

Abstract

The application provides a kind of method and system for identifying electric meter reading, wherein the method comprises the following steps: obtaining the initial image set of electric meter;Feature extraction is carried out on the initial image set, and the first MSFHOG feature image set is obtained;The first MSFHOG feature image set is divided into training set image and verification set image according to the preset proportion;RDELM initial model is constructed;The initial model is trained based on the training set image, and the performance of the trained model is evaluated based on the verification set image until the preset convergence condition is reached, and the RDELM model is obtained;Get the image of the electric meter to be identified, and identify it based on the final model to obtain the electric meter reading result.The method and system for identifying electric meter reading provided by the application use the RDELM model trained and verified to identify the image of the electric meter to be identified, obtain the electric meter reading result, realize the accurate identification of the electric meter reading in low light environment, and effectively reduce the false detection rate.

Description

Technical Field

[0001] This invention relates to the fields of smart grids and computer vision, and in particular to a method and system for identifying electricity meter readings. Background Technology

[0002] With the rapid development of smart grids, automated meter reading recognition technology has become crucial for improving operation and maintenance efficiency. Currently, this technology mainly falls into two categories: traditional image processing and deep learning. Traditional methods are based on the structural features of digital regions (such as seven-segment code layout and character arrangement rules). First, preprocessing such as grayscale conversion and local adaptive binarization suppresses interference from uneven lighting. Then, projection methods or connected component analysis are used to locate the digital region, and methods such as projection histogram valleys are used to segment individual characters. Finally, based on preset seven-segment code stroke rules (such as the "seven-block scanning method"), the stroke state is judged to complete the digital recognition. Traditional methods simplify system complexity through multi-level projection, do not rely on large amounts of training data, and are suitable for low-power devices. Deep learning methods, on the other hand, use artificial neural networks to automatically extract deep image features, combining object detection and sequence recognition to achieve end-to-end reading extraction. The process typically includes three stages: first, using models such as YOLO and Faster R-CNN to locate the reading region; then, using models such as CRNN and LSTM to recognize character sequences; and finally, combining confidence thresholds to filter false detection results and embedding meter reading logic rules (such as numerical range verification) to optimize output reliability. Deep learning methods have strong robustness and generalization ability, can effectively cope with interference such as changes in lighting and shooting tilt, and reduce the manual feature design process, and have the ability to generalize across dial types.

[0003] While current technologies achieve high accuracy under normal lighting conditions, the false detection rate increases in low-light environments (such as at night or in enclosed electrical cabinets) due to issues like low brightness, significant noise, and dynamic shadow interference in meter images. Specifically, traditional single-frame low-light image enhancement methods, while improving image brightness, tend to amplify noise, damaging the topological structure of character strokes and affecting recognition accuracy. Traditional template matching methods are extremely sensitive to uneven lighting and are easily affected by changes in lighting and meter stains. The false matching rate increases under dynamic shadows or local lighting changes, making it difficult to meet stability requirements in complex environments. Deep learning methods rely on large amounts of labeled data, but data acquisition and labeling are costly and computationally resource-intensive, requiring GPU acceleration to meet real-time requirements, making them difficult to deploy and apply in edge embedded devices. Summary of the Invention

[0004] The present invention aims to provide a method and system for identifying electricity meter readings to solve the above-mentioned technical problems, achieve accurate identification of electricity meter readings in low light conditions, and effectively reduce the false detection rate.

[0005] To solve the above-mentioned technical problems, the present invention provides a method for identifying electricity meter readings, comprising the following steps:

[0006] Obtain the initial image set of the electricity meter;

[0007] Feature extraction is performed on the initial image set to obtain the first MSFHOG feature image set;

[0008] The first MSFHOG feature image set is divided into training set images and validation set images according to a preset ratio;

[0009] Construct the initial RDELM model;

[0010] The initial RDELM model is trained based on the training set images, and the performance of the trained model is evaluated based on the validation set images until the preset convergence condition is met, thus obtaining the RDELM model.

[0011] The system acquires an image of the electricity meter to be identified and performs identification based on the RDELM model to obtain the meter reading.

[0012] In the above scheme, an initial image set of the electricity meter is obtained as the basic data for the entire recognition process. Next, image features are extracted from the initial image set, and the resulting first MSFHOG feature image set provides input data for subsequent model training, validation, and testing. Then, the first MSFHOG feature image set is rationally divided according to a preset ratio to ensure the scientific nature of model training and the reliability of evaluation. Further, the initial RDELM model is trained using the training set images, and the trained model is evaluated using the validation set to determine if the model has reached the preset convergence condition. When the model converges, training is stopped to avoid over-optimization. Finally, the final RDELM model is used to recognize the electricity meter image to obtain the meter reading, achieving accurate recognition of meter readings in low-light environments and effectively reducing the false detection rate.

[0013] Further, acquiring the initial image set of the electricity meter includes:

[0014] Obtain initial images of several electricity meters;

[0015] The initial images of the aforementioned meters together constitute the initial image set.

[0016] In the above scheme, an initial image set is formed by acquiring initial images of several electricity meters to meet the sample size requirements for subsequent training and validation of the model, ensuring that each dataset contains different types of samples and avoiding model overfitting due to insufficient data.

[0017] Further, the step of extracting features from the initial image set to obtain a first MSFHOG feature image set includes:

[0018] The initial image set is fused to obtain a fused image set;

[0019] The fused image set is preprocessed to obtain a first binarized image set;

[0020] Feature extraction is performed on the first binarized image set to obtain the first MSFHOG feature image set.

[0021] In the above scheme, by fusing the initial image set, noise in the images can be reduced and image details enhanced, improving the overall image quality and reducing image quality fluctuations caused by shooting shake or changes in lighting. Then, the fused image set is preprocessed to remove residual noise and convert it into a first binarized image set, facilitating subsequent feature extraction. Next, image features are extracted from the first binarized image set, and the resulting first MSFHOG feature image set can provide input data for subsequent model training and validation.

[0022] Further, the step of fusing the initial image set to obtain a fused image set includes:

[0023] Align the initial images in the initial image set to obtain spatially aligned initial images;

[0024] Obtain the local entropy of the spatially aligned initial image and the gradient magnitude of the spatially aligned initial image;

[0025] The weight matrix sequence is obtained based on the local information entropy and gradient magnitude of the image, specifically as follows:

[0026]

[0027] In the formula, w i (x,y) represents the weight matrix sequence, E i (x,y) represents the local information entropy of the image, G i (x,y) represents the image gradient magnitude, and N represents the number of spatially aligned initial images;

[0028] The spatially aligned initial images are fused based on the weight matrix sequence to obtain a fused image set, specifically:

[0029]

[0030] In the formula, S(x,y) represents the fused image set, and I i (x,y) represents the spatially aligned initial image, w i (x,y) represents the weighting coefficients.

[0031] In the above scheme, the initial images in the initial image set are aligned to eliminate spatial misalignment that may exist between multiple initial images due to shooting shake, slight movement of the meter, etc. Then, the fusion weights are calculated based on the local information entropy and gradient magnitude of each spatially aligned initial image to obtain a weight matrix sequence. The spatially aligned initial images are then fused based on the weight matrix sequence, resulting in a fused image set with higher contrast, less noise, and clearer edges compared to the original, providing high-quality input for subsequent processing.

[0032] Further, the preprocessing of the fused image set to obtain a first binarized image set includes:

[0033] Obtain a grayscale image set based on the fused image set;

[0034] Enhance the grayscale image set to obtain the first enhanced grayscale image set;

[0035] Local contrast enhancement is performed on the first enhanced grayscale image set to obtain the second enhanced grayscale image set;

[0036] The second enhanced grayscale image set is subjected to target feature highlighting processing to obtain the third enhanced grayscale image set;

[0037] Perform a closing operation on the third enhanced grayscale image set to obtain the first binarized image set.

[0038] In the above scheme, the fused image set is first converted into a grayscale image set. Then, local contrast enhancement is performed on the first enhanced grayscale image set to alleviate local over-darkness or over-brightness caused by backlighting, shadows, etc., to obtain the first enhanced grayscale image set. Next, local contrast enhancement is performed on the first enhanced grayscale image set to preserve local details and prevent noise areas from interfering with recognition due to over-enhancement. This solves the problem that global enhancement cannot simultaneously preserve details in bright and dark areas, resulting in the second enhanced grayscale image set. Furthermore, target feature enhancement processing is performed on the second enhanced grayscale image set to further highlight the digit regions and suppress background interference, making the digit regions the main subject, resulting in the third enhanced grayscale image set. Finally, a closing operation is performed on the third enhanced grayscale image set to form the final first binarized image set that can be used for feature extraction, providing clear edge information for subsequent feature extraction.

[0039] Further, the step of performing local contrast enhancement on the first enhanced grayscale image set to obtain the second enhanced grayscale image set includes:

[0040] Divide the first enhanced grayscale image in the first enhanced grayscale image set into sub-regions of the first enhanced grayscale image;

[0041] The first enhanced grayscale image sub-region is constrained based on a pixel frequency threshold, and a constraint histogram is obtained, specifically as follows:

[0042] clip=clip_limit×(h×w / block_size 2 ×256)

[0043] In the formula, clip represents the pixel frequency threshold of the first enhanced grayscale image sub-region, h represents the height of the first enhanced grayscale image, w represents the width of the first enhanced grayscale image, block_size represents the division size of the first enhanced grayscale image sub-region, and clip_limit represents the preset contrast limit threshold.

[0044] Adaptive histogram equalization is performed based on the constrained histogram and the first enhanced grayscale image sub-region to obtain the second enhanced grayscale image sub-region;

[0045] Interpolation and fusion are performed on the sub-regions of the second enhanced grayscale image to obtain the second enhanced grayscale image set.

[0046] In the above scheme, the first enhanced grayscale image in the first enhanced grayscale image set is divided into non-overlapping sub-regions, allowing subsequent processing to be performed independently in each local region. This avoids the loss of details caused by uniform equalization of the entire region, preserving target features through local processing. Next, a preset contrast limit threshold is used to prevent excessive noise enhancement due to local pixel concentration during subsequent histogram equalization. Furthermore, histogram equalization is performed independently on each sub-region to avoid overall brightness shift, maintain the original brightness distribution of the meter image, and prevent the numbers from being too bright or too dark, leading to blurring and information loss. Finally, interpolation fusion is performed on the sub-regions to facilitate subsequent feature extraction.

[0047] Further, the step of extracting features from the first binarized image set to obtain the first MSFHOG feature image set includes:

[0048] Based on the first binarized image set, obtain the bottom layer image set, the intermediate layer image set, and the top layer image set;

[0049] HOG feature extraction is performed on the images in the bottom layer image set, the middle layer image set, and the top layer image set respectively to obtain the bottom layer feature image, the middle layer feature image, and the top layer feature image;

[0050] The bottom layer feature image, the intermediate layer feature image and the top layer feature image are fused to obtain the first MSFHOG feature image set.

[0051] In the above scheme, the first binarized image set is processed hierarchically, decomposing the images to express image information at different levels. This provides input at different levels for subsequent feature extraction, enabling the capture of image features from multiple dimensions. Next, HOG feature extraction and feature fusion are performed on each layer of the image set, combining these feature images from different levels. The fused feature image can more comprehensively and completely describe the image content. Obtaining the fused first MSFHOG feature image integrates the advantages of each layer, enhancing the expressive power of the image and providing high-quality input for subsequent model training and validation.

[0052] Further, the step of training the initial RDELM model based on the training set images and evaluating the performance of the trained model based on the validation set images, until a preset convergence condition is met, thereby obtaining the RDELM model, includes:

[0053] The initial RDELM model is trained based on the training set images to obtain the first hidden layer output matrix;

[0054] The cosine similarity matrix is ​​calculated based on the output matrix of the first hidden layer, specifically as follows:

[0055] S=(H T H) / (||H||·||H||)

[0056] In the formula, S represents the cosine similarity matrix, H represents the first hidden layer output matrix, and T is the matrix transpose;

[0057] The second hidden layer output matrix is ​​obtained based on the cosine similarity matrix and a preset similarity threshold.

[0058] Obtain the model output matrix based on training set images;

[0059] The model weights are calculated based on the second hidden layer output matrix and the model output matrix, specifically as follows:

[0060] β=(H ′T H′+λI) -1 H′ T Y;

[0061] In the formula, β represents the model weights, H' represents the output matrix of the second hidden layer, λ represents the regularization parameter, and Y represents the model output matrix;

[0062] The parameters of the initial RDELM model are updated based on the model weights, and the performance of the updated model is evaluated based on the validation set images until the preset convergence condition is met, thus obtaining the RDELM model.

[0063] In the above scheme, the first hidden layer output matrix is ​​obtained by inputting the training set images into the initial RDELM model, realizing the mapping from pixel space to feature space. This preserves the model's initial feature extraction results from the training data, providing data support for subsequent cosine similarity calculation and hidden layer structure optimization. Then, the cosine similarity matrix is ​​calculated to reflect the degree of correlation between different training samples in the hidden layer feature space, measuring the model's ability to represent similar samples. Next, the cosine similarity matrix is ​​used to identify highly correlated feature vectors in the hidden layer, and a preset similarity threshold is used to filter out excessively correlated feature vectors, retaining discriminative feature vectors to generate the second hidden layer output matrix. This reduces feature redundancy, improves model efficiency, and ensures the diversity and discriminativity of hidden layer features. Finally, the performance of the updated model is evaluated using validation set images, and the model is dynamically adjusted until the preset convergence condition is met, obtaining the final RDELM model.

[0064] This invention provides a system for identifying electricity meter readings, comprising an image acquisition module, a feature extraction module, a dataset partitioning module, a model building module, a training and optimization module, and a recognition application module, specifically:

[0065] The image acquisition module is used to acquire the initial image set of the electricity meter;

[0066] The feature extraction module is used to extract features from the initial image set to obtain a first MSFHOG feature image set;

[0067] The dataset partitioning module is used to divide the first MSFHOG feature image set into training set images and validation set images according to a preset ratio.

[0068] The model building module is used to build the initial RDELM model;

[0069] The training and optimization module is used to train the initial RDELM model based on the training set images and evaluate the performance of the trained model based on the validation set images until the preset convergence condition is met, and then obtain the RDELM model.

[0070] The identification application module is used to acquire an image of the electricity meter to be identified, and to identify the image of the electricity meter based on the RDELM model to obtain the electricity meter reading result.

[0071] This invention provides a system for identifying electricity meter readings. In practical applications, only an initial image set of the electricity meter needs to be acquired through an image acquisition module as the basic data for the entire identification process. Next, a feature extraction module extracts image features from the initial image set, and the resulting first MSFHOG feature image set provides input data for subsequent model training, validation, and testing. Then, a dataset partitioning module rationally divides the first MSFHOG feature image set according to a preset ratio, ensuring the scientific nature of model training and the reliability of evaluation. Further, a training and optimization module trains the initial RDELM model constructed by the model building module based on the training set images, and uses a validation set to evaluate the trained model and determine whether it has reached the preset convergence condition. When the model converges, training stops to avoid over-optimization. Finally, the identification application module uses the final RDELM model to identify the electricity meter image to obtain the meter reading result, achieving accurate identification of electricity meter readings in low-light environments and effectively reducing the false detection rate.

[0072] Furthermore, the image acquisition module is used to acquire an initial image set of the electricity meter; including:

[0073] Obtain initial images of several electricity meters;

[0074] The initial images of the aforementioned meters together constitute the initial image set.

[0075] In the above scheme, an initial image set is formed by acquiring initial images of several electricity meters to meet the sample size requirements for subsequent training and validation of the model, ensuring that each dataset contains different types of samples and avoiding model overfitting due to insufficient data. Attached Figure Description

[0076] Figure 1 This is a schematic flowchart of a method for identifying electricity meter readings according to an embodiment of the present invention;

[0077] Figure 2 This is a system architecture diagram for identifying electricity meter readings according to an embodiment of the present invention. Detailed Implementation

[0078] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0079] This embodiment primarily addresses the shortcomings of existing electricity meter reading recognition technologies, such as insufficient noise suppression, incomplete feature preservation, and difficulties in deploying edge devices under low-light conditions. To overcome the limitations of traditional image processing methods, which are susceptible to illumination interference and have low segmentation accuracy, and deep learning methods, which rely on high computing power and have high data annotation costs, an innovative scheme integrating multi-exposure frame fusion and morphological feature enhancement is proposed. By dynamically fusing multiple frames to compensate for brightness imbalances in low-light images, and combining this with directional morphological filtering to enhance the topological structure of digital strokes, a lightweight classification model is designed. This achieves high-precision character recognition under extreme lighting conditions without requiring complex hardware modifications or massive amounts of labeled data. This embodiment balances algorithm robustness with edge computing efficiency, effectively solving the trade-off between noise suppression and detail preservation in existing technologies, adapting to the real-time requirements of embedded devices, and promoting the practical application of automated electricity meter reading.

[0080] The implementation of this embodiment is completed in two stages. The first stage is the training stage, which aims to build a lightweight classification model (MSFHOG+RDELM, where MSFHOG is a multi-scale fused directional gradient histogram and RDELM is a regularized dynamic extreme learning machine) and optimize the parameters. The second stage is the recognition stage, which automatically recognizes and outputs structured readings from the input low-light image.

[0081] This embodiment provides a method for identifying electricity meter readings; for detailed steps, please refer to [link to relevant documentation]. Figure 1 ,include:

[0082] Step S1: Obtain the initial image set of the electricity meter;

[0083] Step S2: Extract features from the initial image set to obtain the first MSFHOG feature image set;

[0084] Step S3: Divide the first MSFHOG feature image set into training set images and validation set images according to a preset ratio;

[0085] Step S4: Construct the initial RDELM model;

[0086] Step S5: Train the initial RDELM model based on the training set images, and evaluate the performance of the trained model based on the validation set images until the preset convergence condition is met, and obtain the RDELM model.

[0087] Step S6: Obtain the image of the meter to be identified, and identify the meter image based on the RDELM model to obtain the meter reading result.

[0088] In this embodiment, an initial image set of the electricity meter is acquired as the foundational data for the entire recognition process. Next, image features are extracted from the initial image set, and the resulting first MSFHOG feature image set provides input data for subsequent model training, validation, and testing. Then, the first MSFHOG feature image set is rationally divided according to a preset ratio, which can be 7:2, to ensure the scientific nature of model training and the reliability of evaluation. Furthermore, the initial RDELM model is trained using the training set images, and the trained model is evaluated using the validation set to determine if the model has reached the preset convergence condition. When the model converges, training is stopped to avoid over-optimization. This embodiment also incorporates a logical rule verification and confidence threshold mechanism, which can reduce the false recognition rate and ensure output stability in complex scenarios. Finally, the final RDELM model is used to recognize the image of the electricity meter to obtain the meter reading result, which includes the meter value, confidence level, and verification status, such as "12345.6kWh, confidence level 92%, verification passed". This significantly improves the character recognition accuracy under low light conditions, more than 40% higher than the traditional method, and realizes accurate recognition of electricity meter readings in low light environment, effectively reducing the false detection rate.

[0089] Further, acquiring the initial image set of the electricity meter includes:

[0090] Obtain initial images of several electricity meters;

[0091] The initial images of the aforementioned meters together constitute the initial image set.

[0092] In this embodiment, initial images of several electricity meters in low-light environments are obtained by using a camera or mobile phone's burst shooting function, enabling the "multiple exposure mode" in the camera settings, and selecting the number of overlays (greater than 3 times), or by continuously capturing multiple images with different exposure durations (by fusing multiple images with different exposures, the dynamic range is expanded and details in both bright and dark areas are preserved, directly utilizing physical multi-frame data rather than algorithmic simulation, thus overcoming the limitations of brightness information in a single exposure). These initial images constitute an initial image set, which covers images of electricity meters acquired under different lighting conditions, scales, and dial types. This increases data diversity, meets the sample size requirements for subsequent model training and validation, ensures that each dataset contains different types of samples, avoids overfitting due to insufficient data, and improves the robustness and generalization ability of the subsequent segmentation and recognition model in complex real-world scenarios, thereby achieving industrial-grade robust recognition. Simultaneously, the initial images are uniformly scaled to a fixed size (512×512 pixels) to eliminate size differences.

[0093] Further, the step of extracting features from the initial image set to obtain a first MSFHOG feature image set includes:

[0094] The initial image set is fused to obtain a fused image set;

[0095] The fused image set is preprocessed to obtain a first binarized image set;

[0096] Feature extraction is performed on the first binarized image set to obtain the first MSFHOG feature image set.

[0097] In this embodiment, by fusing the initial image set, noise in the images can be reduced and image details enhanced, improving the overall image quality and reducing image quality fluctuations caused by shooting shake or changes in lighting. The fused image set is then preprocessed to remove residual noise and convert it into a first binarized image set, facilitating subsequent feature extraction. Next, image features are extracted from the first binarized image set, and the resulting first MSFHOG feature image set can provide input data for subsequent model training and validation.

[0098] Further, the step of fusing the initial image set to obtain a fused image set includes:

[0099] Align the initial images in the initial image set to obtain spatially aligned initial images;

[0100] Obtain the local entropy of the spatially aligned initial image and the gradient magnitude of the spatially aligned initial image;

[0101] The weight matrix sequence is obtained based on the local information entropy and gradient magnitude of the image, specifically as follows:

[0102]

[0103] In the formula, w i (x,y) represents the weight matrix sequence, E i (x,y) represents the local information entropy of the image, G i (x,y) represents the image gradient magnitude, and N represents the number of spatially aligned initial images;

[0104] The spatially aligned initial images are fused based on the weight matrix sequence to obtain a fused image set, specifically:

[0105]

[0106] In the formula, S(x,y) represents the fused image set, and I i (x,y) represents the spatially aligned initial image, w i (x,y) represents the weighting coefficients.

[0107] In this embodiment, motion compensation is calculated using optical flow to align the initial images in the initial image set, eliminating spatial misalignment that may exist between multiple initial images due to shooting shake, slight movement of the meter, etc., thus aligning multiple frames spatially to preserve the spatiotemporal consistency of continuous scenes. Next, fusion weights are calculated based on the local information entropy (reflecting texture complexity) and gradient magnitude (reflecting edge strength) of each spatially aligned initial image, thereby suppressing noise regions and preserving details. A weight matrix sequence is obtained, and a weighted average method is used to fuse the spatially aligned initial images based on the weight matrix sequence. This results in a fused image set with higher contrast, less noise, clearer edges, and preserved details in dark areas compared to the original set, providing high-quality input for subsequent processing.

[0108] Further, the preprocessing of the fused image set to obtain a first binarized image set includes:

[0109] Obtain a grayscale image set based on the fused image set;

[0110] Enhance the grayscale image set to obtain the first enhanced grayscale image set;

[0111] Local contrast enhancement is performed on the first enhanced grayscale image set to obtain the second enhanced grayscale image set;

[0112] The second enhanced grayscale image set is subjected to target feature highlighting processing to obtain the third enhanced grayscale image set;

[0113] Perform a closing operation on the third enhanced grayscale image set to obtain the first binarized image set.

[0114] In this embodiment, the fused image set is first converted into a grayscale image set. By enhancing structural and texture features, color interference is suppressed, while reducing data volume and improving processing speed. Then, local contrast enhancement is performed on the digital stroke topology structure of the first enhanced grayscale image set based on multi-scale morphological filtering. Opening operations are performed on the 1×3 vertical line structural elements in the digital stroke topology structure to eliminate horizontal noise and enhance vertical strokes; closing operations are performed on the 3×1 horizontal line structural elements to repair broken strokes and connect horizontal strokes; 5×5 circular structural elements are used to fill the digital intersection area (such as the center of "8") to optimize the intersection point, thereby mitigating the problem of local over-darkness or over-brightness caused by backlighting, shadows, etc., to obtain the first enhanced grayscale image set. Next, local contrast enhancement is performed on the first enhanced grayscale image set to preserve the local details of the first enhanced grayscale image set and prevent noise areas from interfering with recognition due to over-enhancement. This solves the problem that global enhancement cannot take into account the details of bright and dark areas simultaneously, to obtain the second enhanced grayscale image set. Furthermore, the second enhanced grayscale image set is processed for target feature prominence based on the adaptive binarization Sauvola algorithm, specifically: Where mean represents the average gray level of neighboring pixels, std represents the standard deviation of gray level of neighboring pixels, and k represents the correction parameter (usually 0.1-0.5, which needs to be adjusted according to the image). The image is simplified to black and white to further highlight the digit region, suppress background interference, and provide clear structural information with high contrast and low noise for subsequent character segmentation and recognition, making the digit region the main body, and obtaining the third enhanced gray level image set. Then, a closing operation is performed on the third enhanced gray level image set to form the first binarized image set that can be used for feature extraction, providing clear edge information for subsequent feature extraction. Finally, this embodiment also performs accurate character segmentation on the first binarized image in the output first binarized image set, and standardizes the size and gray level range to eliminate illumination differences. Combined with vertical / horizontal projection dual constraints, the segmentation accuracy is improved. Specifically, the digit row region and row boundary are located by vertical projection histogram to eliminate dial border interference; individual characters are segmented by horizontal projection combined with connected component analysis; at the same time, they are uniformly normalized to 28×28 pixels and the [0,1] gray level range to adapt to the requirements of subsequent feature extraction and classification models.

[0115] Further, the step of performing local contrast enhancement on the first enhanced grayscale image set to obtain the second enhanced grayscale image set includes:

[0116] Divide the first enhanced grayscale image in the first enhanced grayscale image set into sub-regions of the first enhanced grayscale image;

[0117] The first enhanced grayscale image sub-region is constrained based on a pixel frequency threshold, and a constraint histogram is obtained, specifically as follows:

[0118] clip=clip_limit×(h×w / block_size 2 ×256)

[0119] In the formula, clip represents the pixel frequency threshold of the first enhanced grayscale image sub-region, h represents the height of the first enhanced grayscale image, w represents the width of the first enhanced grayscale image, block_size represents the division size of the first enhanced grayscale image sub-region, and clip_limit represents the preset contrast limit threshold.

[0120] Adaptive histogram equalization is performed based on the constrained histogram and the first enhanced grayscale image sub-region to obtain the second enhanced grayscale image sub-region;

[0121] Interpolation and fusion are performed on the sub-regions of the second enhanced grayscale image to obtain the second enhanced grayscale image set.

[0122] In this embodiment, the first enhanced grayscale image in the first enhanced grayscale image set is divided into non-overlapping sub-regions, allowing subsequent processing to be performed independently in each local region. This avoids the loss of detail caused by uniform equalization of the entire region, preserving target features through local processing. Next, a preset contrast limit threshold is used to prevent excessive noise enhancement due to local pixel concentration during subsequent histogram equalization. Furthermore, histogram equalization is performed independently on each sub-region to avoid overall brightness shift, maintain the original brightness distribution of the meter image, and prevent blurring of numbers due to excessive brightness or darkness, thus preventing information loss. Then, bilinear interpolation fusion is performed on the sub-regions to eliminate block artifacts, ensuring overall image smoothness and facilitating subsequent feature extraction.

[0123] Further, the step of extracting features from the first binarized image set to obtain the first MSFHOG feature image set includes:

[0124] Based on the first binarized image set, obtain the bottom layer image set, the intermediate layer image set, and the top layer image set;

[0125] HOG feature extraction is performed on the images in the bottom layer image set, the middle layer image set, and the top layer image set respectively to obtain the bottom layer feature image, the middle layer feature image, and the top layer feature image;

[0126] The bottom layer feature image, the intermediate layer feature image and the top layer feature image are fused to obtain the first MSFHOG feature image set.

[0127] In this embodiment, the first binarized image set is processed hierarchically to decompose the image into three Gaussian pyramids (original image, 0.5×, 0.25×), allowing image information to be expressed at different levels and providing different levels of input for subsequent feature extraction, enabling the capture of image features from multiple dimensions. Next, HOG (Histogram of Oriented Gradients) feature extraction is performed on each layer of the image set (with parameters set as follows: Cell size 8×8 pixels, Block size 2×2 Cell, Number of orientation bins 8 (0°, 45°, 90°, 135°)). HOG features are calculated independently for each layer, and then feature fusion is performed by weighting according to scale (weight: original image > 0.5× > 0.25×). This combines these feature images from different levels, preserving details while enhancing scale invariance, so that the fused feature image can more comprehensively and completely describe the image content. Obtaining the fused first MSFHOG feature image can integrate the advantages of each layer, enhance the expressive power of the image, and provide high-quality input for subsequent model training and validation.

[0128] Further, the step of training the initial RDELM model based on the training set images and evaluating the performance of the trained model based on the validation set images, until a preset convergence condition is met, thereby obtaining the RDELM model, includes:

[0129] The initial RDELM model is trained based on the training set images to obtain the first hidden layer output matrix;

[0130] The cosine similarity matrix is ​​calculated based on the output matrix of the first hidden layer, specifically as follows:

[0131] S=(H T H) / (||H||·||H||)

[0132] In the formula, S represents the cosine similarity matrix, H represents the first hidden layer output matrix, and T is the matrix transpose;

[0133] The second hidden layer output matrix is ​​obtained based on the cosine similarity matrix and a preset similarity threshold.

[0134] Obtain the model output matrix based on training set images;

[0135] The model weights are calculated based on the second hidden layer output matrix and the model output matrix, specifically as follows:

[0136] β=(H ′T H′+λI) -1 H′ T Y;

[0137] In the formula, β represents the model weights, H' represents the output matrix of the second hidden layer, λ represents the regularization parameter, and Y represents the model output matrix;

[0138] The parameters of the initial RDELM model are updated based on the model weights, and the performance of the updated model is evaluated based on the validation set images until the preset convergence condition is met, thus obtaining the RDELM model.

[0139] In this embodiment, by inputting the training set images into the RDELM initial model and randomly initializing the parameters, the first hidden layer output matrix is ​​obtained, realizing the mapping from pixel space to feature space. This preserves the model's initial feature extraction results from the training data, providing data support for subsequent cosine similarity calculation and hidden layer structure optimization. Then, the cosine similarity matrix is ​​calculated to reflect the degree of correlation between different training samples in the hidden layer feature space, used to measure the model's ability to represent similar samples. Next, the cosine similarity matrix is ​​used to identify highly correlated feature vectors in the hidden layer, and a preset similarity threshold is used to filter out excessively correlated feature vectors, retaining discriminative feature vectors to generate the second hidden layer output matrix. This reduces feature redundancy, improves model efficiency, and ensures the diversity and discriminativeness of hidden layer features. Then, the model weights are calculated using the second hidden layer output matrix and the model output matrix, where λ is used to control the regularization strength and suppress overfitting. Finally, the updated model performance is evaluated using validation set images, and the model is dynamically adjusted until the preset convergence condition is met. The final RDELM model parameters, CLAHE parameters, morphological filter structuring element size, and other operating parameters are stored.

[0140] This embodiment provides a system for identifying electricity meter readings, including an image acquisition module, a feature extraction module, a dataset partitioning module, a model building module, a training and optimization module, and a recognition application module, specifically:

[0141] The image acquisition module is used to acquire the initial image set of the electricity meter;

[0142] The feature extraction module is used to extract features from the initial image set to obtain a first MSFHOG feature image set;

[0143] The dataset partitioning module is used to divide the first MSFHOG feature image set into training set images and validation set images according to a preset ratio.

[0144] The model building module is used to build the initial RDELM model;

[0145] The training and optimization module is used to train the initial RDELM model based on the training set images and evaluate the performance of the trained model based on the validation set images until the preset convergence condition is met, and then obtain the RDELM model.

[0146] The identification application module is used to acquire an image of the electricity meter to be identified, and to identify the image of the electricity meter based on the RDELM model to obtain the electricity meter reading result.

[0147] This embodiment provides a system for recognizing electricity meter readings. In practical applications, only an initial image set of the electricity meter needs to be acquired through an image acquisition module as the basic data for the entire recognition process. Next, a feature extraction module extracts image features from the initial image set, and the resulting first MSFHOG feature image set provides input data for subsequent model training, validation, and testing. Then, a dataset partitioning module rationally partitions the first MSFHOG feature image set according to a preset ratio, which can be 7:2, to ensure the scientific nature of model training and the reliability of evaluation. Furthermore, a training and optimization module trains the initial RDELM model built by the model construction module based on the training set images, and uses a validation set to evaluate the trained model and determine whether it has reached the preset convergence condition. When the model converges, training stops to avoid over-optimization. This embodiment also embeds a logical rule verification and confidence threshold mechanism, which can reduce the false recognition rate and ensure output stability in complex scenarios. Finally, the final RDELM model is used to recognize the image of the electricity meter to obtain the meter reading result, which includes the meter value, confidence level, and verification status, such as "12345.6kWh, confidence level 92%, verification passed". This significantly improves the character recognition accuracy under low light conditions, more than 40% higher than the traditional method, and realizes accurate recognition of electricity meter readings in low light environment, effectively reducing the false detection rate.

[0148] Furthermore, the image acquisition module is used to acquire an initial image set of the electricity meter; including:

[0149] Obtain initial images of several electricity meters;

[0150] The initial images of the aforementioned meters together constitute the initial image set.

[0151] In this embodiment, initial images of several electricity meters in low-light environments are obtained by using a camera or mobile phone's burst shooting function, enabling the "multiple exposure mode" in the camera settings, and selecting the number of overlays (greater than 3 times), or by continuously capturing multiple images with different exposure durations (by fusing multiple images with different exposures, the dynamic range is expanded and details in both bright and dark areas are preserved, directly utilizing physical multi-frame data rather than algorithmic simulation, thus overcoming the limitations of brightness information in a single exposure). These initial images constitute an initial image set, which covers images of electricity meters acquired under different lighting conditions, scales, and dial types. This increases data diversity, meets the sample size requirements for subsequent model training and validation, ensures that each dataset contains different types of samples, avoids overfitting due to insufficient data, and improves the robustness and generalization ability of the subsequent segmentation and recognition model in complex real-world scenarios, thereby achieving industrial-grade robust recognition. Simultaneously, the initial images are uniformly scaled to a fixed size (512×512 pixels) to eliminate size differences.

[0152] This embodiment employs the MSFHOG+RDELM lightweight classification model, which reduces the number of parameters by more than 90% compared to deep learning solutions. It achieves low power consumption (≤3W) and high-efficiency recognition without GPU acceleration, making it suitable for embedded devices. Furthermore, this embodiment significantly reduces its reliance on labeled data compared to deep learning solutions.

[0153] This embodiment can solve the problems of poor image quality under low light and amplified noise when single-frame enhancement increases brightness; this embodiment can solve the problem of existing technologies being sensitive to sudden changes in illumination, which leads to a significant increase in binarization and positioning errors; this embodiment can solve the problem of matching algorithm accuracy, speed, power consumption with the limited resources (computing power, memory, energy consumption) of edge devices.

[0154] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for identifying electricity meter readings, characterized in that, Includes the following steps: Obtain the initial image set of the electricity meter; Feature extraction is performed on the initial image set to obtain a first MSFHOG feature image set. Specifically, the initial images in the initial image set are aligned to obtain spatially aligned initial images; the local information entropy and gradient magnitude of the corresponding spatially aligned initial images are obtained; and a weight matrix sequence is obtained based on the local information entropy and gradient magnitude. In the formula, Represents the sequence of weight matrices. Represents the local information entropy of an image. The gradient magnitude is represented by N, which represents the number of spatially aligned initial images. The spatially aligned initial images are fused based on the weight matrix sequence to obtain a fused image set, specifically: In the formula, Indicates a fused image set. This indicates that the initial image is spatially aligned. Indicates the weighting coefficient; The fused image set is preprocessed to obtain a first binarized image set; Based on the first binarized image set, obtain the bottom layer image set, the intermediate layer image set, and the top layer image set; HOG features are extracted from the images in the bottom layer image set, the middle layer image set, and the top layer image set respectively to obtain the bottom layer feature image, the middle layer feature image, and the top layer feature image; the bottom layer feature image, the middle layer feature image, and the top layer feature image are then fused to obtain the first MSFHOG feature image set. The first MSFHOG feature image set is divided into training set images and validation set images according to a preset ratio; Construct the initial RDELM model; The initial RDELM model is trained based on the training set images, and the performance of the trained model is evaluated based on the validation set images until the preset convergence condition is met, thus obtaining the RDELM model. The system acquires an image of the electricity meter to be identified and performs identification based on the RDELM model to obtain the meter reading.

2. The method for identifying meter readings according to claim 1, characterized in that, The acquisition of the initial image set of the electricity meter includes: Obtain initial images of several electricity meters; The initial images of the aforementioned meters together constitute the initial image set.

3. The method for identifying meter readings according to claim 1, characterized in that, The step of preprocessing the fused image set to obtain a first binarized image set includes: Obtain a grayscale image set based on the fused image set; Enhance the grayscale image set to obtain the first enhanced grayscale image set; Local contrast enhancement is performed on the first enhanced grayscale image set to obtain the second enhanced grayscale image set; The second enhanced grayscale image set is subjected to target feature highlighting processing to obtain the third enhanced grayscale image set; Perform a closing operation on the third enhanced grayscale image set to obtain the first binarized image set.

4. The method for identifying meter readings according to claim 3, characterized in that, The step of performing local contrast enhancement on the first enhanced grayscale image set to obtain the second enhanced grayscale image set includes: Divide the first enhanced grayscale image in the first enhanced grayscale image set into sub-regions of the first enhanced grayscale image; The first enhanced grayscale image sub-region is constrained based on a pixel frequency threshold, and a constraint histogram is obtained, specifically as follows: In the formula, represents the pixel frequency threshold of the first enhanced grayscale image sub-region, h represents the height of the first enhanced grayscale image, and w represents the width of the first enhanced grayscale image. This indicates the partition size of the first enhanced grayscale image sub-region. Indicates the preset contrast limit threshold; Adaptive histogram equalization is performed based on the constrained histogram and the first enhanced grayscale image sub-region to obtain the second enhanced grayscale image sub-region; Interpolation and fusion are performed on the sub-regions of the second enhanced grayscale image to obtain the second enhanced grayscale image set.

5. The method for identifying meter readings according to claim 1, characterized in that, The initial RDELM model is trained based on the training set images, and the performance of the trained model is evaluated based on the validation set images until the preset convergence condition is met, thereby obtaining the RDELM model. include: The initial RDELM model is trained based on the training set images to obtain the first hidden layer output matrix; The cosine similarity matrix is ​​calculated based on the output matrix of the first hidden layer, specifically as follows: In the formula, Represents the cosine similarity matrix. T represents the output matrix of the first hidden layer, and T is the matrix transpose. The second hidden layer output matrix is ​​obtained based on the cosine similarity matrix and a preset similarity threshold. Obtain the model output matrix based on training set images; The model weights are calculated based on the second hidden layer output matrix and the model output matrix, specifically as follows: In the formula, β Represents the model weights. H’ This represents the output matrix of the second hidden layer. λ represents the regularization parameter, and Y represents the model output matrix; The parameters of the initial RDELM model are updated based on the model weights, and the performance of the updated model is evaluated based on the validation set images until the preset convergence condition is met, thus obtaining the RDELM model.

6. A system for identifying electricity meter readings, characterized in that, It includes an image acquisition module, a feature extraction module, a dataset partitioning module, a model building module, a training and optimization module, and a recognition application module, specifically: The image acquisition module is used to acquire the initial image set of the electricity meter; The feature extraction module is used to extract features from the initial image set to obtain a first MSFHOG feature image set. Specifically, it aligns the initial images in the initial image set to obtain spatially aligned initial images; it obtains the local entropy and gradient magnitude of the image corresponding to the spatially aligned initial images; and it obtains a weight matrix sequence based on the local entropy and gradient magnitude. In the formula, Represents the sequence of weight matrices. Represents the local information entropy of an image. The gradient magnitude is represented by N, which represents the number of spatially aligned initial images. The spatially aligned initial images are fused based on the weight matrix sequence to obtain a fused image set, specifically: In the formula, Indicates a fused image set. This indicates that the initial image is spatially aligned. Indicates the weighting coefficient; The fused image set is preprocessed to obtain a first binarized image set; Based on the first binarized image set, obtain the bottom layer image set, the intermediate layer image set, and the top layer image set; HOG features are extracted from the images in the bottom layer image set, the middle layer image set, and the top layer image set respectively to obtain the bottom layer feature image, the middle layer feature image, and the top layer feature image; the bottom layer feature image, the middle layer feature image, and the top layer feature image are then fused to obtain the first MSFHOG feature image set. The dataset partitioning module is used to divide the first MSFHOG feature image set into training set images and validation set images according to a preset ratio. The model building module is used to build the initial RDELM model; The training and optimization module is used to train the initial RDELM model based on the training set images and evaluate the performance of the trained model based on the validation set images until the preset convergence condition is met, and then obtain the RDELM model. The identification application module is used to acquire an image of the electricity meter to be identified, and to identify the image of the electricity meter based on the RDELM model to obtain the electricity meter reading result.

7. A system for identifying electricity meter readings according to claim 6, characterized in that, The image acquisition module is used to acquire an initial image set of the electricity meter; it includes: Obtain initial images of several electricity meters; The initial images of the aforementioned meters together constitute the initial image set.

Images