A weighted decision fusion recognition method of physical ID tags

Abstract

The application discloses a kind of weighted decision fusion identification methods of real ID label, comprising the following steps: the image of real ID label is obtained with visual sensor and carries out feature extraction and text recognition to obtain visual identification result, while RFID reading result of real ID label is obtained with wireless radio frequency sensor;Respectively calculate the weight corresponding to visual sensor and the weight corresponding to wireless radio frequency sensor, according to the weight, weighted fusion calculation is carried out to visual identification result and RFID reading result to obtain fusion identification result;According to the contrast of the image of real ID label and the change of RFID signal strength, the weight corresponding to visual sensor and the weight corresponding to wireless radio frequency sensor are dynamically updated.The application combines Paddle OCR intelligent vision and RFID electronic tag technology, realizes scene dynamic weight adjustment based on, improves the identification precision and robustness of real ID label.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a weighted decision fusion identification method for physical ID tags. Background Technology

[0002] With the accelerated advancement of smart grid construction, the State Grid Corporation of China has widely applied physical ID tag technology to the entire process of managing various equipment, materials, power transmission and transformation, power distribution, and metering assets. In equipment management, the State Grid Corporation uses physical ID tag technology to bind tags to dedicated equipment. Readers retrieve relevant equipment information, enabling real-time recording of data such as equipment status, usage time, location, and responsible personnel. This achieves systematic, automated, and registered management, establishing a comprehensive, full-cycle equipment asset management system. In materials management, the State Grid uses physical ID tag technology to establish a "State Grid Core Digital Identity" for materials, enabling rapid collection, real-time recording, accurate analysis, and judgment of materials management information. This helps the company achieve digitalization, standardization, automation, and intelligence in materials management. In on-site inspections, construction, and commissioning scenarios, the State Grid Corporation uses physical ID tags and its business systems to achieve a series of functions, including on-site safety supervision, on-site material and equipment allocation, and automated recording of construction manuals and commissioning records. This improves work efficiency and construction quality while reducing the risk of safety accidents.

[0003] Currently, the two mainstream recognition technologies are intelligent vision technology based on optical character recognition (OCR) and radio frequency identification (RFID) technology.

[0004] Traditional identification methods often rely on a single technology. For example, while RFID technology can provide efficient non-contact identification, it is easily limited by factors such as damaged tags or electromagnetic interference. While Optical Character Recognition (OCR) technology can provide visual recognition, its recognition rate decreases when the image quality is poor or the line of sight is obstructed. Summary of the Invention

[0005] The technical problem to be solved by this invention is: addressing the technical problems existing in the prior art, this invention provides a weighted decision fusion recognition method for physical ID tags, which combines Paddle OCR intelligent vision and RFID electronic tag technology to achieve dynamic weight adjustment based on the scene, thereby improving the recognition accuracy and robustness of physical ID tags.

[0006] To solve the above-mentioned technical problems, the technical solution proposed by this invention is as follows:

[0007] A weighted decision fusion identification method for physical ID tags includes the following steps:

[0008] S101) Use a visual sensor to acquire images of physical ID tags and perform feature extraction and text recognition to obtain visual recognition results, while using a radio frequency sensor to acquire RFID reading results of physical ID tags;

[0009] S102) Calculate the weights corresponding to the visual sensor and the RFID sensor respectively, and perform weighted fusion calculation on the visual recognition result and the RFID reading result according to the weights to obtain the fused recognition result;

[0010] S103) Based on the changes in image contrast and RFID signal strength of the physical ID tag, dynamically update the weights corresponding to the visual sensor and the RFID sensor.

[0011] Furthermore, in step S101, feature extraction and text recognition include: after preprocessing the image, inputting it into a pre-configured PaddleOCR model for feature extraction and text recognition. The preprocessing steps include:

[0012] S201) The first image is obtained by improving the overall brightness of the image by combining adaptive gamma correction with weighted distribution;

[0013] S202) Decompose the first image into the product of illumination component and reflection component, and establish a reflection model;

[0014] S203) The reflection model is transformed into a frequency domain representation through logarithmic and Fourier transforms;

[0015] S204) Set the specified parameters of the Gaussian filter and use the Gaussian high-pass filter to perform homomorphic filtering in the frequency domain to obtain the second image;

[0016] S205) Perform inverse transform and exponential operation on the second image to obtain the improved image.

[0017] Furthermore, before step S101, there is also a step of configuring the PaddleOCR model, which specifically includes:

[0018] S301) Uses cross-entropy loss and geometric loss as classification losses for text detection loss functions;

[0019] S302) Uses the intersection-union ratio (IoU) loss as the geometric loss of the text detection loss function;

[0020] S303) Sequence recognition loss using CTC loss as the text recognition loss function;

[0021] S304) Calculate the weighted sum of the text detection loss function and the text recognition loss function to obtain the loss function of the PaddleOCR model.

[0022] Furthermore, in step S102, when calculating the weights corresponding to the visual sensor and the wireless radio frequency sensor respectively, the following steps are included:

[0023] S401) The weight of each sensor is used as a variable to establish a multivariate objective function with constraints;

[0024] S402) Solve for the optimal solution of the objective function to obtain the weight of each sensor.

[0025] Furthermore, the objective function expression is as follows:

[0026] F(α i )=f(α i )+λθ(α i )

[0027] in,

[0028]

[0029]

[0030] θ(α i ) represents the constraint condition, α i Let σ be the weight of the i-th sensor. i Let λ be the variance of the i-th sensor, and λ be the Lagrange multiplier.

[0031] Furthermore, solving for the optimal solution of the objective function includes:

[0032] Find α for the objective function i The partial derivative is expressed as follows:

[0033]

[0034] make We obtain the first relation, which is:

[0035]

[0036] Simultaneous equation of first relation and constraint θ(α) i From the formula, we obtain the second relation, which is:

[0037]

[0038] Substituting the second relation into the first relation, we obtain the weight of the i-th sensor, as shown below:

[0039]

[0040] Where, α i Let σ be the weight of the i-th sensor. i Let be the variance of the i-th sensor.

[0041] Furthermore, step S103 specifically includes:

[0042] S501) Calculate the contrast of the image and the coefficient of variation of the RFID signal strength;

[0043] S502) The visual recognition result value is calculated based on the contrast of the image and the variance corresponding to the visual sensor, and the RFID reading result value is calculated based on the coefficient of variation of the RFID signal strength and the variance corresponding to the radio frequency sensor.

[0044] S503) The gradient corresponding to the visual sensor is calculated based on the value of the visual recognition result and the difference between the current fusion recognition result and the expected result; and the gradient corresponding to the radio frequency sensor is calculated based on the value of the RFID reading result and the difference between the current fusion recognition result and the expected result.

[0045] S504) Subtract the product of the gradient and the preset learning rate of the visual sensor from the weights corresponding to the current visual sensor, and subtract the product of the gradient and the preset learning rate of the wireless radio frequency sensor from the weights corresponding to the current wireless radio frequency sensor to obtain the new weights of the visual sensor and the wireless radio frequency sensor.

[0046] Furthermore, the gradient expressions corresponding to the visual sensor and the gradient expressions corresponding to the radio frequency sensor are as follows:

[0047]

[0048]

[0049] in, The gradient corresponding to the visual sensor. Let p be the gradient corresponding to the wireless radio frequency sensor, y be the probability of the current fusion recognition result being correct, and x be the label indicating whether the current fusion recognition result is correct. OCR It is a visual recognition result, x RFID The RFID reading result is defined as:

[0050]

[0051]

[0052] in, The variance corresponding to the visual sensor. is the variance corresponding to the wireless radio frequency sensor, function f is the contrast of the visual recognition image, and function g is the coefficient of variation of the RFID signal strength.

[0053] The present invention also proposes a weighted decision fusion identification system for physical ID tags, comprising a microprocessor and a computer-readable storage medium interconnected thereto, wherein the microprocessor is programmed or configured to execute the weighted decision fusion identification method for physical ID tags as described in any one of the present invention.

[0054] The present invention also proposes a computer-readable storage medium for being programmed or configured by a microprocessor to perform a weighted decision fusion identification method for physical ID tags as described in any one of the present invention.

[0055] Compared with the prior art, the advantages of the present invention are as follows:

[0056] This invention utilizes Paddle OCR technology to recognize textual information and RFID to read electronic information. It then dynamically updates the weighted decision fusion strategy using an online stochastic gradient descent (SGD) algorithm, adjusting the weights of the recognition results in real time according to different usage scenarios. This adaptive adjustment mechanism optimizes recognition accuracy based on environmental changes and sensor performance, thereby improving the efficiency and accuracy of physical ID tag management in the power grid, achieving the goal of increasing work efficiency and reducing labor intensity. Attached Figure Description

[0057] Figure 1 This is a flowchart of a method according to an embodiment of the present invention. Detailed Implementation

[0058] The present invention will be further described below with reference to the accompanying drawings and specific preferred embodiments, but this does not limit the scope of protection of the present invention.

[0059] Before introducing this embodiment, the relevant concepts will be explained.

[0060] Optical Character Recognition (OCR) is a technology that analyzes and processes image files containing text to obtain document and layout information. It primarily includes image input, pre-processing (binarization, image noise reduction, and tilt correction), mid-processing (layout analysis, character segmentation, character recognition, and layout restoration), and post-processing (recognition correction). Ultimately, it converts the text in the image into an editable text format. It works by scanning printed text in paper documents or images and using computer vision and pattern recognition algorithms to convert the characters in the image into machine-encoded text.

[0061] RFID is a wireless communication technology that uses radio waves to identify specific targets and read related data without requiring mechanical or optical contact. A typical RFID system includes tags, readers, and a data processing system.

[0062] Example 1

[0063] To improve the accuracy and robustness of physical ID tags during rapid and accurate reading, writing, and automatic entry in the power grid, this embodiment proposes a weighted decision fusion identification method for physical ID tags, such as... Figure 1 As shown, it includes the following steps:

[0064] S100 system construction. Configure visual sensors to process and identify visual information on physical ID tags; configure radio frequency sensors to read RFID information on physical ID tags.

[0065] S101) Data Acquisition and Preprocessing. Images of physical ID tags are acquired using a visual sensor and feature extraction and text recognition are performed using intelligent visual recognition to obtain visual recognition results. Simultaneously, RFID reading results of the physical ID tags are acquired using a radio frequency sensor.

[0066] S102) Weighted decision fusion. The weights corresponding to the visual sensor and the RFID sensor are calculated separately. Based on these weights, the visual recognition result and the RFID reading result are weighted and fused to obtain the fused recognition result.

[0067] S103) Online weight update. Based on the changes in image contrast and RFID signal strength of the physical ID tags, the weights corresponding to the visual sensor and the RFID sensor are dynamically updated using the online stochastic gradient descent (SGD) algorithm.

[0068] S104) Output and Application. The fusion recognition results will be output for subsequent data processing, storage, or business logic.

[0069] The above steps utilize OCR technology to recognize textual information and RFID to read electronic information. Then, an online stochastic gradient descent (SGD) algorithm is used to dynamically update the weighted decision fusion strategy, adjusting the weights of the recognition results in real time according to different usage scenarios. This adaptive adjustment mechanism can optimize recognition accuracy based on environmental changes and sensor performance, thereby improving the efficiency and accuracy of power grid physical ID tag management, achieving the goals of increasing work efficiency and reducing labor intensity.

[0070] The following is a detailed explanation of each step.

[0071] In step S100 of this embodiment, the visual sensor is a high-resolution camera that can capture ID tag images and use the PaddleOCR model for feature extraction and text recognition to obtain visual recognition results. The wireless radio frequency sensor uses an RFID reader / writer.

[0072] In step S101 of this embodiment, feature extraction and text recognition include: after preprocessing the image, inputting it into a pre-configured PaddleOCR model for feature extraction and text recognition. In the image preprocessing process, adaptive gamma correction improves the overall brightness of the image by adjusting the cumulative distribution function of the compensation. The image is decomposed into the product of illumination and reflection components, and a reflection model is established. This model is then transformed into a frequency domain representation through logarithmic and Fourier transforms. Homomorphic filtering is performed using a Gaussian high-pass filter. The filtering effect is adjusted by setting specific parameters, and finally, an improved image is obtained through inverse transform and exponential operation, thereby enhancing the visual effect of the image and highlighting details. Therefore, the preprocessing steps include:

[0073] S201) The overall brightness of the image is improved by combining weighted distribution with adaptive gamma correction. In this embodiment, the image with improved brightness is referred to as the first image. The adaptive gamma correction formula is as follows:

[0074]

[0075] Where T(l) represents the intensity of the output image, l max Here, cdf(l) represents the maximum grayscale value of the input image, l represents the intensity of the input image, and cdf(l) represents the cumulative distribution of the brightness l. w (l) is the cumulative distribution function for compensation, which can approximate cdf(l); pdf w (l) refers to the probability density function of the histogram after l is adjusted by a weighted distribution, pdf max and PDF min These are the maximum and minimum values ​​of the probability density statistical histogram, ∑pdf. w It refers to the sum of probability densities of the entire image;

[0076] S202) Decompose the first image into the product of illumination component and reflection component, and establish a reflection model;

[0077] Specifically, the image to be processed, o(x,y), is considered as the product of the illumination component i1(x,y) and the reflection component r1(x,y), and the reflection model is established as follows:

[0078] o(x,y)=i1(x,y)·r1(x,y) (2)

[0079] S203) The reflection model is transformed into a frequency domain representation through logarithmic and Fourier transforms;

[0080] Specifically, performing a logarithmic transformation on the reflection model yields the following two additive components:

[0081] lno(x,y)=lni1(x,y)+lnr1(x,y) (3)

[0082] Next, performing a Fourier transform on the above equation yields its corresponding frequency domain representation:

[0083] O(u,v)=I(u,v)+R(u,v) (4)

[0084] S204) Set the specified parameters of the Gaussian filter and use the Gaussian high-pass filter to perform homomorphic filtering in the frequency domain to obtain the second image;

[0085] Specifically, the transfer function H(u,v) is used to process the results after the Fourier transform:

[0086] H(u,v)O(u,v)=H(u,v)I(u,v)+H(u,v)R(u,v) (5)

[0087] The homomorphic filter transfer function H(u,v) for filtering in the frequency domain using a Gaussian high-pass filter is:

[0088]

[0089] By setting the cutoff frequency D0 of the Gaussian filter, the constant e1 that controls the sharpness of the filter function's ramp, and the high-frequency gain γ... H and low-frequency gain γ L The homomorphically filtered frequency domain image is obtained, which is referred to as the second image in this embodiment;

[0090] (S205) Perform inverse transform and exponential operation on the second image to obtain the improved image. Performing inverse transform and exponential operation on the image is a conventional method used by those skilled in the art, and the specific implementation process will not be described in detail here.

[0091] The PaddleOCR model in this embodiment is an improved neural network model based on the PaddleOCR framework, used to process and recognize visual information on physical ID tags. How to construct an improved neural network model based on the PaddleOCR framework to process and recognize visual information on physical ID tags is well known to those skilled in the art and is not an innovation of this solution; therefore, the specific implementation process will not be elaborated here. In this embodiment, the overall loss function of the PaddleOCR model is optimized. The loss function is defined as the weighted sum of the text detection loss function and the text recognition loss function, used to simultaneously optimize the text detection and recognition tasks. Therefore, before step S101 in this embodiment, a step of configuring the PaddleOCR model is included, specifically including:

[0092] S301) Uses cross-entropy loss and geometric loss as classification losses for text detection loss functions;

[0093] Specifically, the loss function for text detection includes two parts: classification loss and geometric loss, which are used to ensure that the model can accurately locate the text regions in the image.

[0094] Classification loss (L cls This is used to distinguish between text and non-text regions. In this embodiment, cross-entropy loss and geometric loss are used, expressed as follows:

[0095]

[0096] Where N is the total number of pixels in the batch, y i p is the actual label (text or non-text) of the i-th pixel. i It is the probability that the model predicts that the point is a text region.

[0097] S302) Uses the intersection-union ratio (IoU) loss as the geometric loss of the text detection loss function;

[0098] Specifically, geometric loss (L geo This is used to optimize the shape of the text region predicted by the model, making it closer to the actual text box. In this embodiment, the IoU (Intersection over Union) loss is used, which is suitable for rotated rectangular or quadrilateral boxes, and its expression is as follows:

[0099]

[0100] Among them, Pred i It is the i-th text region predicted by the model, GT i It corresponds to the actual text area.

[0101] S303) Sequence recognition loss using CTC loss as the text recognition loss function;

[0102] Specifically, the loss function in the text recognition part is used to ensure that the model can accurately recognize the text content within the detected text region.

[0103] Sequence recognition loss (L seq This method is used to handle text sequence recognition problems. In this embodiment, the CTC (Connectionist Temporal Classification) loss is used, and its expression is as follows:

[0104] L seq =-logP(Y|X) (9)

[0105] Where X is a given input image feature sequence, Y is a target text sequence, and P(Y|X) is the probability that the model predicts Y given X.

[0106] S304) Calculate the weighted sum of the text detection loss function and the text recognition loss function to obtain the loss function of the PaddleOCR model, as shown in the following expression:

[0107] L total =λ cls L cls +λ geo L geo +λ seq L seq (10)

[0108] Where, λ cls , λ geo and λ seq These are weighting coefficients used to balance the contributions of different loss functions. The values ​​of these weighting coefficients can be adjusted according to the application scenario. For example, if accurate text recognition is of greater concern, then the weight of sequence recognition loss needs to be increased, therefore λ should be increased. seq The value of λ is reduced. cls , λ geo The value of .

[0109] Step S102 of this embodiment aims to design a weighted decision fusion algorithm based on the confidence and reliability of the visual recognition results and RFID reading results. By setting appropriate weight parameters and calculating these parameters through the weighted fusion algorithm, the best fusion effect can be achieved, thereby combining the advantages of the two technologies to obtain the final ID recognition information.

[0110] In multi-sensor data fusion, sensors with smaller variance are considered to have higher confidence and reliability because variance is an indicator of the dispersion (i.e., deviation from the true value) of measured values. Smaller variance means that the difference between the sensor's measurement and the true value is smaller, indicating that the information provided by the sensor is closer to the reality, and therefore it is considered more reliable.

[0111] The purpose of multi-sensor data fusion is to improve the accuracy of target estimation. However, since the variance of the sensors is constant, the impact of sensor variance on the fusion weights must be considered during fusion. Therefore, step S102, which calculates the weights corresponding to the visual sensors and the wireless radio frequency sensors respectively, includes the following steps:

[0112] S401) The weight of each sensor is used as a variable to establish a multivariate objective function with constraints;

[0113] Specifically, let the multi-sensor fusion weights be: α i (i = 1, 2, ..., n), where i represents the current sensor number. For example, the weight corresponding to the visual sensor can be set as α1, and the weight corresponding to the radio frequency sensor can be set as α2. Therefore, solving for α... i What conditions must be met for this to be true? To minimize the value of , we can give a multivariate objective function with constraints, as shown in the following expression:

[0114] F(α i )=f(α i )+λθ(α i (11)

[0115] The constraints are:

[0116]

[0117] In the above formula, α i Let σ be the weight of the i-th sensor. i Let λ be the variance of the i-th sensor, and λ be the Lagrange multiplier.

[0118] S402) Solve for the optimal solution of the objective function to obtain the weight of each sensor. The derivation and calculation process for finding the optimal solution is as follows:

[0119] Find α for the objective function i The partial derivative is expressed as follows:

[0120]

[0121] make Objective function F(α) i To obtain the minimum value, the following relationship exists, which is referred to as the first relationship in this embodiment:

[0122]

[0123] Simultaneous equation of first relation and constraint θ(α) iThe formula, that is, the following relation is obtained from formulas (12) and (14), which is called the second relation in this embodiment:

[0124]

[0125] Substituting the second relation into the first relation, we obtain the weight of the i-th sensor, as shown below:

[0126]

[0127] Where, α i Let σ be the weight of the i-th sensor. i Let be the variance of the i-th sensor. It can be seen that in the formula for calculating the fusion weight, the reliability of the sensor is reflected by assigning a larger weight to the sensor with a smaller variance, that is, the weight is inversely proportional to the sensor variance.

[0128] After obtaining the weights corresponding to the visual sensor and the RFID sensor, the visual recognition result and the RFID reading result are weighted and calculated to obtain the fused recognition result. Weighted calculation is a conventional technique well known to those skilled in the art, and the specific calculation process will not be described in detail here.

[0129] Step S103 of this embodiment aims to calculate the coefficient of variation of the contrast of the visual recognition image and the RFID signal strength, combine them with the variance of the corresponding sensor, and then dynamically update the calculated weights using the online stochastic gradient descent method.

[0130] Specifically, it includes:

[0131] S501) Calculate the coefficient of variation of the image contrast and RFID signal strength. How to calculate the coefficient of variation of the image contrast and RFID signal strength is well known to those skilled in the art, and the specific calculation process will not be described in detail here.

[0132] S502) The visual recognition result value is calculated based on the contrast of the image and the variance corresponding to the visual sensor, and the RFID reading result value is calculated based on the coefficient of variation of the RFID signal strength and the variance corresponding to the radio frequency sensor.

[0133] Specifically, x OCR It is a visual recognition result, x RFID The RFID reading result is defined as:

[0134]

[0135]

[0136] in, The variance corresponding to the visual sensor. This represents the variance of the wireless radio frequency sensor, where f is the contrast of the visual recognition image and g is the coefficient of variation of the RFID signal strength. By combining these with the corresponding sensor variance, the final read value is obtained.

[0137] S503) The gradient corresponding to the visual sensor is calculated based on the value of the visual recognition result and the difference between the current fusion recognition result and the expected result; and the gradient corresponding to the radio frequency sensor is calculated based on the value of the RFID reading result and the difference between the current fusion recognition result and the expected result.

[0138] Specifically, a loss function is defined to measure the difference between the fusion result and the expected result (actual ID) under the current weight parameters. In this embodiment, logistic regression loss is chosen:

[0139] L=-[ylog(p)+(1-y)log(1-p)] (19)

[0140] Where y is the actual label (1 for correct identification and 0 for incorrect identification), and p is the probability of the correctness of the fusion result under the current weight.

[0141] For each identified event, the gradient of the loss function with respect to each weight parameter is calculated. In this embodiment, there are two weight parameters: α... OCR The weights corresponding to the visual sensor, α RFID These are the weights corresponding to the wireless radio frequency sensor, and the gradient is calculated as follows:

[0142]

[0143] Where, x OCR It is a visual recognition result, x RFID This is the result of RFID reading;

[0144] S504) Subtract the product of the gradient and the preset learning rate of the visual sensor from the weights corresponding to the current visual sensor, and subtract the product of the gradient and the preset learning rate of the wireless radio frequency sensor from the weights corresponding to the current wireless radio frequency sensor to obtain the new weights of the visual sensor and the wireless radio frequency sensor.

[0145] Specifically, in this embodiment, the online SGD algorithm is used to update the weight parameters in real time based on the calculated gradient:

[0146]

[0147] Here, η is the learning rate, which determines the step size for weight updates.

[0148] like Figure 1As shown, in this embodiment, after obtaining the updated weights in step S103, data acquisition and preprocessing are performed again. The updated weights are then used to perform a weighted fusion calculation on the visual recognition results and RFID reading results to obtain a new fused recognition result. Simultaneously, the weight updates are continuously repeated, updating the weight parameters based on the results of each recognition event and user feedback. Specifically, the weight of the visual sensor is increased in well-lit areas, and the weight of the radio frequency sensor is increased in areas with good reader signal, until the required number of repetitions is reached or the weight parameters converge. Correspondingly, in step S104 of this embodiment, the output recognition result is the recognition result obtained after the aforementioned multiple weight updates and weighted fusion calculation.

[0149] After verification, in the application of intelligent and rapid reading, writing and entry of power asset information, even when faced with adverse conditions such as damaged tags, obstructed vision or electromagnetic interference, the above steps can still maintain high recognition performance and achieve efficient and accurate physical ID tag recognition.

[0150] Example 2

[0151] This embodiment proposes a weighted decision fusion recognition system for physical ID tags, including a microprocessor and a computer-readable storage medium interconnected with each other. The microprocessor is programmed or configured to execute the weighted decision fusion recognition method for physical ID tags described in Embodiment 1.

[0152] This embodiment also proposes a computer-readable storage medium that can be programmed or configured by a microprocessor to execute the weighted decision fusion identification method for physical ID tags described in Embodiment 1.

[0153] The process by which the system in this embodiment identifies physical ID tags is as follows:

[0154] Step 1: System Construction. Build an improved neural network model based on the PaddleOCR framework to process and recognize visual information on physical ID tags; configure an RFID reader to read RFID information from the physical ID tags.

[0155] Step 2: Data Acquisition and Preprocessing. Images of the ID tags are acquired using a high-resolution camera and preprocessed using PaddleOCR, including image enhancement, noise reduction, and normalization. Simultaneously, the electronic information on the ID tags is read using an RFID reader.

[0156] Step 3: Intelligent Visual Recognition. In the image preprocessing workflow, adaptive gamma correction improves the overall brightness of the image by adjusting the cumulative distribution function of the compensation. The image is decomposed into the product of illumination and reflection components, and a reflection model is established. This model is then transformed into a frequency domain representation through logarithmic and Fourier transforms. Homomorphic filtering is performed using a Gaussian high-pass filter, and the filtering effect is adjusted by setting specific parameters. Finally, the improved image is obtained through inverse transform and exponential operations, thereby enhancing the visual effect and highlighting details. The preprocessed image is input into the improved PaddleOCR model for feature extraction and text recognition, yielding the visual recognition result.

[0157] Step 4: RFID Information Reading. The electronic information of the tag is acquired using an RFID reader / writer, serving as a second, independent identification result.

[0158] Step 5: Weighted Decision Fusion. Based on the confidence and reliability of the visual recognition results and RFID reading results, a weighted decision fusion algorithm is designed to combine the advantages of both technologies to obtain the final ID identification information; an appropriate weight parameter (α) is set. OCR α RFID ), where α OCR Regarding the visual recognition results, α RFID Based on the RFID reading results, these parameters are calculated using a weighted fusion algorithm to achieve the best fusion effect.

[0159] Step 6: Update weights online. Calculate the coefficients of variation of the contrast of the visual recognition image and the RFID signal strength, combine them with the variance of the corresponding sensors, and then dynamically update the weights calculated in Step 5 using the online stochastic gradient descent method.

[0160] Step 7, Output and Application. The fused recognition results are output for subsequent data processing, storage, or business logic; in the application of intelligent and rapid reading, writing, and entry of power asset information, efficient and accurate physical ID tag recognition is achieved.

[0161] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the invention. Therefore, any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention should fall within the protection scope of the present invention.

Claims

1. A weighted decision fusion recognition method of physical ID tags, characterized in that, The method comprises the following steps: S101) acquiring an image of a physical ID tag by using a visual sensor and performing feature extraction and text recognition to obtain a visual recognition result, and acquiring an RFID reading result of the physical ID tag by using a wireless radio frequency sensor; S102) calculating a weight corresponding to the visual sensor and a weight corresponding to the wireless radio frequency sensor respectively, and performing weighted fusion calculation on the visual recognition result and the RFID reading result according to the weights to obtain a fusion recognition result; When the weight corresponding to the visual sensor and the weight corresponding to the wireless radio frequency sensor are calculated respectively, the following steps are included: S401) taking the weight of each sensor as a variable, and establishing a multivariable objective function with a constraint condition, and the objective function expression is as follows: Wherein, is a constraint, is a weight of the i-th sensor, is a variance of the i-th sensor, is a Lagrange multiplier; S402) solving the optimal solution of the objective function to obtain the weight of each sensor, and when the optimal solution of the objective function is solved, the following steps are included: Find the objective function The partial derivative is expressed as follows: Let , we obtain the first relation, which is simultaneously with the first relationship and the constraint condition The second relationship is obtained by simultaneously solving the first relationship and the constraint condition Substitute the second relationship into the first relationship to obtain the weight of the i th sensor, and the expression is as follows: wherein, wi is the weight of the i-th sensor, σi2is the variance of the i-th sensor; S103) dynamically updating the weight corresponding to the visual sensor and the weight corresponding to the wireless radio frequency sensor according to the contrast of the image of the physical ID tag and the change of the RFID signal strength.

2. The weighted decision fusion recognition method of physical ID tags according to claim 1, characterized in that, When the feature extraction and text recognition are performed in step S101, the following steps are included: after the image is preprocessed, the PaddleOCR model configured in advance is inputted to perform feature extraction and text recognition, and the steps of the preprocessing include: S201) improving the overall brightness of the image by combining the adaptive gamma correction of the weighted distribution to obtain a first image; S202) decomposing the first image into the product of illumination component and reflection component to establish a reflection model; S203) converting the reflection model into frequency domain representation through logarithmic transformation and Fourier transformation; S204) setting the specified parameters of the Gaussian filter and using the Gaussian high-pass filter to perform homomorphic filtering in the frequency domain space to obtain a second image; S205) performing inverse transformation and exponential operation on the second image to obtain an improved image.

3. The weighted decision fusion recognition method of physical ID tags according to claim 1, characterized in that, The step S101 further includes the step of configuring the PaddleOCR model, specifically including: S301) using cross-entropy loss function and geometric loss as the classification loss of the text detection loss function; S302) using the intersection over union IoU loss as the geometric loss of the text detection loss function; S303) using the CTC loss as the sequence recognition loss of the text recognition loss function; S304) calculating the weighted sum of the text detection loss function and the text recognition loss function to obtain the loss function of the PaddleOCR model.

4. The weighted decision fusion recognition method of physical ID tags according to claim 1, characterized in that, Step S103 specifically includes: S501) calculating the coefficient of variation of the contrast of the image and the RFID signal strength; S502) calculating the value of the visual recognition result according to the contrast of the image and the variance corresponding to the visual sensor, and calculating the value of the RFID reading result according to the coefficient of variation of the RFID signal strength and the variance corresponding to the wireless radio frequency sensor; S503) calculating a gradient corresponding to the visual sensor according to the value of the visual recognition result and the difference between the current fusion recognition result and the expected result, and calculating a gradient corresponding to the wireless radio frequency sensor according to the value of the RFID reading result and the difference between the current fusion recognition result and the expected result; S504) subtracting the product of the gradient corresponding to the visual sensor and a preset learning rate from the weight corresponding to the current visual sensor, and subtracting the product of the gradient corresponding to the wireless radio frequency sensor and the preset learning rate from the weight corresponding to the wireless radio frequency sensor, to obtain a new weight of the visual sensor and a new weight of the wireless radio frequency sensor.

5. The weighted decision fusion recognition method of physical ID tags according to claim 4, characterized in that, The gradient corresponding to the visual sensor and the gradient corresponding to the wireless radio frequency sensor are expressed as follows: wherein, is the gradient corresponding to the visual sensor, is the gradient corresponding to the wireless radio frequency sensor, is the probability of the correctness of the current fusion recognition result, is the label of whether the current fusion recognition result is correct, is the visual recognition result, is the RFID reading result, defined as: wherein, is the variance corresponding to the visual sensor, is the variance corresponding to the wireless radio frequency sensor, function f is the contrast of the visual recognition image, function g is the coefficient of variation of the RFID signal strength.

6. A weighted decision fusion recognition system of physical ID tags, characterized by, The computer readable storage medium is used to be programmed or configured by the microprocessor to execute the weighted decision fusion recognition method of the physical ID tag according to any one of claims 1-5.

7. A computer readable storage medium characterized in that, The computer readable storage medium is used to be programmed or configured by the microprocessor to execute the weighted decision fusion recognition method of the physical ID tag according to any one of claims 1-5.

Images