Electric power inspection evidence identification method and system based on data enhancement retraining

Abstract

The invention belongs to the technical field of image recognition, provides an electric power inspection evidence recognition method and system based on data enhancement and retraining, and aims to solve the problems of low accuracy of electric power marketing inspection image recognition, poor service scene applicability and the like, and adopts data enhancement to expand samples, YOLOv8 retraining optimization and network compression to improve the application range. Therefore, electric power inspection evidence identification requirements of multiple service scenes can be met. Firstly, in order to avoid low model identification accuracy caused by insufficient training sample quantity, a CutMix-based data enhancement algorithm is adopted to generate a simulation training sample; secondly, based on the expanded training sample set, carrying out applicability retraining on a YOLOv8 network identification model, and improving the accuracy of electric power inspection image identification; and finally, the retrained YOLOv8 network identification model is compressed by adopting a mode of combining structural design and knowledge distillation, so that the field identification requirement of the mobile terminal application is met, the network scale is effectively reduced, and the identification timeliness is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of image recognition, and in particular relates to a method and system for identifying power audit evidence based on data enhancement and retraining. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] Power marketing audit mainly investigates and handles responsible accidents, errors, illegal electricity use and over-capacity electricity use in power marketing work. It is based on marketing business data and aims to improve the management level and work quality of power marketing, which can largely avoid the risks that may exist in the power marketing process. At present, with the development of digital transformation of enterprises, power marketing audit work has carried out information construction. In the application scenarios of marketing audit business, it is necessary to compare and review unstructured data such as explanatory materials, on-site photos, system screenshots, etc. Traditionally, it relies on manual review, which has problems such as low efficiency, high work intensity, inconsistent standards, and easy negligence and omissions.

[0004] At present, commonly used text detection and recognition methods include DBNet text detection algorithm, CRNN text recognition algorithm, YOLOv8 network, etc. For example, the recognition of multi-angle and irregular texts can be realized by improving the Convolutional Recurrent Neural Network (CRNN); based on the structural characteristics of Uyghur, the DBNet network is improved to achieve accurate detection of Uyghur texts such as small texts, tilted texts, and curved texts; the above methods are based on specific text recognition requirements, and the text recognition methods are improved to improve the recognition performance;

[0005] However, on-site inspections require that the recognition module be embedded in mobile devices. The current YOLOv8 network recognition model has problems such as large scale and low real-time performance, so some institutions and scholars have conducted research on network model compression. For example, in order to solve the problem that deep neural networks are large in size, consume high resources, and are difficult to port to mobile devices, the model is compressed by integrating knowledge distillation, structural design, network pruning, and parameter quantization; another example is a bearing fault diagnosis method for compressed deep neural networks, which combines structured pruning, unstructured pruning, parameter quantization, and matrix compression multi-level processing to achieve multi-level network compression.

[0006] However, the above methods use network pruning, knowledge distillation, parameter quantization, structural design and other methods to compress the deep network individually or in combination. However, in actual marketing audit application scenarios, the actual amount of collected image data in different business scenarios is relatively small, which often cannot support the model's retraining and compression parameter adjustment needs. Therefore, the insufficient number of basic training samples leads to poor model recognition effect. Summary of the invention

[0007] In order to solve at least one technical problem existing in the above-mentioned background technology, the present invention provides a method and system for identifying power audit evidence based on data enhancement retraining. After the training sample set is expanded, the accuracy of recognizing power marketing audit images by the retrained deep YOLOv8 network recognition model is significantly improved, especially for mobile on-site recognition, the shallow recognition model after compression training effectively reduces the network scale and improves the recognition timeliness.

[0008] In order to achieve the above object, the present invention adopts the following technical solution:

[0009] A first aspect of the present invention provides a method for identifying power audit evidence based on data enhancement retraining, comprising the following steps:

[0010] Obtain power inspection evidence pictures, classify them according to power text recognition types, and obtain multiple types of power inspection evidence real training sample sets;

[0011] Expand the real training sample set for each type of power audit evidence to generate multiple types of power audit evidence simulation training sample sets;

[0012] Combine multiple types of power audit evidence real training sample sets and multiple types of power audit evidence simulation training sample sets to obtain a power audit evidence material identification sample library;

[0013] The YOLOv8 network recognition model is trained based on the electric power audit supporting material recognition sample library, wherein the YOLOv8 network recognition model includes a shallow model and a deep model, and the recognition information of the deep model is learned by the shallow model in layers to obtain the trained YOLOv8 network recognition model;

[0014] The trained YOLOv8 network recognition model is compressed, and the recognition result of mobile power inspection evidence image recognition is obtained based on the compressed YOLOv8 network recognition model.

[0015] Furthermore, the real training sample set of each type of power audit evidence is expanded to generate multiple types of power audit evidence simulation training sample sets, including:

[0016] While keeping the structure of the real training sample set unchanged, a simulated training sample set is generated based on the CutMix data enhancement algorithm, which includes:

[0017] Randomly select two power inspection evidence pictures, recorded as the first image and the second image;

[0018] Determine a mixed area ratio parameter of the first image and the second image in the mixed image;

[0019] Calculating the area range of the cropping frame according to the size of the first image and the mixed area ratio parameter, wherein the area of ​​the cropping region is inversely proportional to the mixed area ratio parameter;

[0020] Randomly cut a rectangular area in the first image and the second image according to the determined cropping frame, place the cropped area of ​​the second image in the cropped area of ​​the first image, merge them into a new image, and form a new sample;

[0021] According to the mixed area ratio parameter, the labels of the two original images are weighted averaged according to the area of ​​the cut area to calculate the labels of the new training samples, and the sample set supported by the new power audit is used as the simulation training sample set.

[0022] Furthermore, the YOLOv8 network recognition model includes a shallow model and a deep model, and the recognition information of the deep model is learned hierarchically by the shallow model, including:

[0023] After the deep model training is completed, the shallow model layer uses the output data of the deep model layer as the learning target output data, and feeds back to adjust the network parameters of the shallow model layer until the deep model layer and the shallow model layer have the same data input, and the output data error loss of the two is within the error threshold range, which means that the shallow model has completed the hierarchical learning of the deep model.

[0024] Furthermore, the YOLOv8 network recognition model is trained based on the power audit supporting material recognition sample library, and the YOLOv8 network recognition model includes a shallow model and a deep model. The shallow model hierarchically learns the recognition information of the deep model to obtain the trained YOLOv8 network recognition model, including:

[0025] The input and output definition of the target recognition model takes the image in the recognition sample library as input, divides the image into S×S grid cells, and outputs B prediction boxes for each grid cell. Each prediction box contains the position, size, confidence, and conditional probability of different categories of the prediction box.

[0026] The confidence of the target positioning is calculated based on the predicted position and predicted size of each prediction box of each grid unit, and the confidence probability value of the i-th type of character is calculated based on the obtained confidence of the target positioning, and the character recognition result is further output;

[0027] According to the coordinate size error between the predicted box and the true box, the confidence error between the predicted box and the true box, and the character recognition classification error, the loss function error of the YOLOv8 network recognition model is calculated. According to the loss function error feedback gradient, each weight value in the convolution kernel is adjusted and the weight is updated until the error is less than the set threshold. Then the model parameters are fixed and the YOLOv8 network recognition model is output.

[0028] Furthermore, the calculation formula of the confidence of the target positioning is:

[0029]

[0030] The calculation formula of the confidence probability value of the i-th type of character is:

[0031] s i =p i ×c

[0032] Wherein, p(Object) indicates whether there is a center point of the target (Object) in the grid unit. If so, p(Object) = 1; if not, p(Object) = 0. represents the intersection-over-union ratio of the predicted box and the true box area; area(t) represents the area of ​​the true box; area(p) represents the area of ​​the predicted box, p i It indicates the conditional probability that the feature value in the current prediction box meets the conditional probability of the i-th type of character.

[0033] Furthermore, the loss function error of the YOLOv8 network recognition model is:

[0034] loss = loss1 + loss2 + loss3,

[0035]

[0036] Among them, loss1 is the coordinate size error between the predicted box and the real box, λ coord Represents the set position coefficient of the prediction box, Used to determine whether the jth predicted box in the i-th grid unit is responsible for the target obj, and the real box of the target obj The largest prediction box is responsible for the target obj. If it is responsible, then Otherwise, 0; Indicates the position and size of the prediction box; (x i ,y i ), (w i ,h i ) represents the position and size of the real frame;

[0037] loss2 is the confidence error between the predicted box and the real box, λ noord Indicates the coefficient when there is no target obj in the grid cell. Used to determine whether the jth prediction box in the i-th grid unit is not responsible for the target obj. If not, Otherwise 0; c i Represents the confidence of the real box; Represents the confidence of the prediction box;

[0038] loss3 is the character recognition classification error, p i (C) represents the true probability that the i-th grid cell belongs to class C; The prediction for the i-th grid cell is the conditional probability of class C.

[0039] Furthermore, compressing the trained YOLOv8 network recognition model includes:

[0040] The network structure and weight parameters of the first layer of the YOLOv8 network recognition model are retained to extract the feature information of the image. The feature information of the image is output by the next layer of the first layer, and a four-layer shallow network is constructed to learn the implicit features of the outputs of the second to fifth layers.

[0041] In the process of extracting the implicit features of the outputs from the second to the fifth layers as shallow network learning parameters, the output implicit features are decomposed into multiple CBL shallow network representations to compress the network structure from the second to the fifth layers.

[0042] A second aspect of the present invention provides a power audit evidence recognition system based on data enhancement retraining, comprising:

[0043] The sample classification module is used to obtain power inspection evidence pictures, classify them according to the power text recognition type, and obtain multiple types of power inspection evidence real training sample sets;

[0044] A sample expansion module is used to expand each type of power audit evidence real training sample set to generate multiple types of power audit evidence simulation training sample sets; combine multiple types of power audit evidence real training sample sets and multiple types of power audit evidence simulation training sample sets to obtain a power audit evidence material identification sample library;

[0045] A recognition model retraining module, which is used to train a YOLOv8 network recognition model based on a sample library of power audit supporting materials recognition, wherein the YOLOv8 network recognition model includes a shallow model and a deep model, and the recognition information of the deep model is learned by the shallow model in layers to obtain a trained YOLOv8 network recognition model;

[0046] The model compression module is used to compress the trained YOLOv8 network recognition model, and obtain the recognition result of the mobile power inspection evidence image recognition based on the compressed YOLOv8 network recognition model.

[0047] A third aspect of the present invention provides a computer-readable storage medium.

[0048] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the above-mentioned method for identifying power audit evidence based on data enhancement and retraining.

[0049] A fourth aspect of the present invention provides a computer device.

[0050] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps in the method for identifying power audit evidence based on data enhancement and retraining as described above are implemented.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] 1. In view of the small scale of power corpus, the lack of public standard data sets, and the presence of numbers, Chinese and English, special symbols, and some text stamp representations, the present invention directly uses the current text detection and recognition method to perform character recognition on power audit supporting materials, and its recognition rate often cannot reach the industrial application standard. Based on different categories of power audit supporting image samples, the YOLOv8 network is retrained so that its internal network can learn information that is more in line with the characteristics of the power corpus, thereby improving the recognition accuracy;

[0053] 2. The number of power audit evidence samples obtained for different business scenarios is small, making it difficult to fully train the deep network model. The data enhancement algorithm based on CutMix is ​​used in this invention to randomly select two sample images and mix them together to create a new training example, thereby increasing the diversity of the data set and expanding the training sample set. This diversity helps the model learn more power text feature representations, thereby improving its generalization ability and reducing the risk of overfitting.

[0054] 3. Aiming at the needs of mobile applications for on-site audits and to ensure high text detection accuracy and detection rate, the present invention adopts a method that combines structural design with knowledge distillation to compress the deep YOLOv8 network recognition model, and ensures the recognition accuracy of the shallow model by learning the hidden "dark knowledge" of the middle layer in a layered manner.

[0055] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0057] Figure 1 The invention provides a method for identifying power audit evidence based on data enhancement and retraining.

[0058] Figure 2 It is a YOLOv8 network structure diagram provided by an embodiment of the present invention;

[0059] Figure 3 is a comparison chart of similarity indexes of generated samples provided by an embodiment of the present invention;

[0060] Figure 4 It is a comparison chart of the diversity evaluation index of generated samples provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0061] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0062] It should be noted that the following detailed descriptions are all illustrative and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0063] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0064] In view of the technical problems mentioned in the background technology, the present invention proposes a method for identifying power audit evidence based on data enhancement retraining, which mainly includes the following contents:

[0065] (1) In view of the small scale of the power corpus, the lack of a public standard data set, and the presence of numbers, Chinese and English, special symbols, and some text stamps, the current text detection and recognition method is directly used to perform character recognition on the power audit supporting materials, and its recognition rate often cannot reach the industrial application standard. Therefore, the present invention retrains the YOLOv8 network based on different categories of power audit supporting image samples, so that its internal network can learn information that is more in line with the characteristics of the power corpus, thereby improving the recognition accuracy;

[0066] (2) The number of power audit evidence samples obtained for different business scenarios is small, making it difficult to fully train the deep network model. The present invention is based on the CutMix data enhancement algorithm, which randomly selects two sample images and mixes them together to create a new training example, thereby increasing the power text diversity of the data set and expanding the training sample set. Power text diversity helps the model learn more power text feature representations, thereby improving its generalization ability and reducing the risk of overfitting.

[0067] (3) Aiming at the needs of mobile applications for on-site audits and to ensure high text detection accuracy and detection rate, the present invention adopts a method that combines structural design with knowledge distillation to compress the deep YOLOv8 network recognition model, and ensures the recognition accuracy of the shallow model by learning the hidden "dark knowledge" of the middle layer in a hierarchical manner.

[0068] Embodiment 1

[0069] like Figure 1 As shown, this embodiment provides a method for identifying power audit evidence based on data enhancement retraining, comprising the following steps:

[0070] Step 1: Obtain power inspection evidence pictures and classify them according to power text recognition types to obtain multiple types of power inspection evidence pictures;

[0071] In this embodiment, the power inspection evidence pictures are collected and summarized through the marketing business application system, the power consumption information collection system, etc.;

[0072] In this embodiment, according to different power audit evidence text recognition scenario requirements, power text recognition types are classified, including system screenshot recognition samples, electric energy representation number recognition samples, and on-site transformer identification plate text recognition samples.

[0073] Step 2: Build a sample library for identifying supporting materials for power audits;

[0074] In order to solve the problem of data enhancement diversity, this embodiment adopts the CutMix algorithm to expand samples, that is, randomly generate a bounding box (bbox) to crop the corresponding position of image A, and then put the corresponding area of ​​image B into the cropped area of ​​image A to form a new sample, which specifically includes the following steps:

[0075] Step 201: Randomly select two power inspection evidence images, denoted as image A and image B, and their corresponding labels, determine the hyperparameter α used to control the size of the cut area in CutMix, and generate the mixing ratio parameter λ.

[0076] A value λ is randomly sampled from the β(α, α) distribution. This value will be used to determine the area ratio of image A and image B in the mixed image. The value range of λ is between [0, 1], but λ = max(λ, 1-λ) is usually used to ensure that λ is not too small so that both images have enough area to be displayed.

[0077] Step 202: Calculate the area range of the cropping frame according to the size (width W and height H) of the image A and the blending ratio λ, and ensure that the area of ​​the cropping region is inversely proportional to λ.

[0078] A starting point (x, y) is randomly selected in image A. This point will be used as the coordinate of the upper left corner of the cropping area, and a rectangular cropping frame is generated around the starting point to ensure that the cropping area does not exceed the boundary of image A. According to the starting point, the size of the cropping area and the parameter λ, the size of the cropping frame is calculated to determine the bounding box B of the cropping area = (r x ,r y ,r w ,r h ).

[0079] To sample the binary mask M, we first need to sample the bounding box B of the cropped region = (r x ,r y ,r w ,r h ) is used to sample the sample x A and x B Do the indication calibration of the cropping area. Sample the rectangular mask M, and the bounding box sampling formula of the cropping area is as follows:

[0080] r x ~Unif(0,W),

[0081] r y ~Unif(0,H),

[0082]

[0083]

[0084] The size of the matrix M is the same as the image, the value in the bbox is 0, and the other values ​​are 1. After determining the cropping area B, set the cropping area B in the binary mask to 0 and the other areas to 1, and the mask sampling is completed.

[0085] Step 203: Based on the cropping frame, randomly cut a rectangular area in the two images, place the cropped part of image B in the cropped area of ​​image A, and merge them into a new image to form a new sample.

[0086] Assume x A and xB Represents two different power inspection evidence pictures A and B, y A and B are their corresponding label values. CutMix needs to generate new training samples. and corresponding labels Its operation formula is as follows:

[0087]

[0088] Where M∈{0,1} w×h , is a mask matrix marked by 0 and 1 that is consistent with the image size, used to mark the areas that need to be cropped and the areas that are retained. A , the image part with M=1 will be retained; for x B , the part where M=0 will be retained. ⊙ represents element-wise multiplication, λ follows the β(α,α) distribution, and α is set to 1 here.

[0089] Step 204: According to the mixing ratio λ, the labels of the two original images are weighted averaged according to the area of ​​the cut region to calculate the label of the new training sample.

[0090] After the two images are fused, the label value is determined by the proportion of their respective areas. Assuming that 30% and 70% of the two images are fused together, the original labels are [1,0] and [0,1] respectively, then the fused label is [0.3,0.7]. The generated bbox size can be calculated as:

[0091]

[0092] The area ratio of bbox and original image is:

[0093]

[0094] It can be seen that image A retains the part outside the bbox, so The label weight is , The label weight is .

[0095] Step 205: Use the new sample set supported by the power audit to train the subsequent network recognition model.

[0096] The mixed image is passed as a new training sample data input to the recognition model. The loss is calculated based on the mixed labels (weighted average labels) and the weights of the model parameters are updated.

[0097] Step 3: Train the YOLOv8 network recognition model based on the power audit supporting material recognition sample library;

[0098] Among them, the YOLOv8 network recognition model is as follows Figure 2 As shown, it includes a shallow model and a deep model. The recognition information of the deep model is learned through the shallow model layering. That is, after the deep model training is completed, the shallow model layering uses the output data of the deep model layering as the learning target output data, and feedback is used to adjust the network parameters of the shallow model layering until the deep model layer and the shallow model layer have the same data input, and the output data error loss of the two is within the error threshold range, which means that the shallow model has completed the hierarchical learning of the deep model.

[0099] Among them, the shallow model layer uses the output data of the deep model layer as the learning target output data, that is, the shallow model layer 2 uses the output data of the deep model layer 2 as the learning target output data, the shallow model layer 3 uses the output data of the deep model layer 3 as the learning target output data, the shallow model layer 4 uses the output data of the deep model layer 4 as the learning target output data, and the shallow model layer 5 uses the output data of the deep model layer 5 as the learning target output data. This embodiment uses the detection method based on the YOLOv8 network to realize the target positioning and character recognition of the collected images in the power marketing inspection and audit scene. Taking the transformer identification plate recognition as an example, the specific process is as follows:

[0100] like Figure 2 As shown, first, the model input and output are defined.

[0101] The size of the original image of the transformer sign is initialized to 1024×1024×3, where 3 represents the three color channels of RGB, as the input of the model; the model output is defined, the image is divided into S×S grid units, each grid unit outputs B prediction boxes, and each prediction box outputs 4 types of data, including the position (x, y) of the prediction box, the size (w, h), the confidence c, and the conditional probabilities p1,…,p of different categories N , where the number of categories is N.

[0102] Subsequently, this embodiment uses the YOLOv8 network to build a recognition model. The network in the feature extraction stage contains 52 CBL layers, 23 residual layer structures, and three feature map outputs of different scales are designed.

[0103] like Figure 2 The basic structural unit structures of each network are shown as follows:

[0104] The CBL layer is a network layer that combines convolution operations (Convolution), batch normalization operations (BN, Batch Normal) and activation operations represented by the activation function Leaky ReLU. The image sample data is first convolved using a convolution kernel constructed with a weight of size b×b, as shown below:

[0105]

[0106] The convolution kernel moves on the image with a defined step size, the grayscale value of the image pixel is multiplied by the weight in the convolution kernel, and the multiplied values ​​of the same pixel obtained by the RGB three channels of the same convolution kernel are added as the output feature value of the convolution operation.

[0107] The BN operation regularizes the eigenvalues ​​obtained after the convolution operation into a normal distribution with a mean of 0 and a variance of 1, so as to speed up the convergence of the training model and alleviate overfitting of the training sample set.

[0108] The eigenvalues ​​obtained after the batch normalization operation use the activation function LeakyReLU to obtain the eigenvalue output of the CBL layer. The specific formula is:

[0109]

[0110] Among them, a i represents a fixed parameter, a i ∈(1,+∞).

[0111] The CBLset layer refers to a collection of multiple CBL layers. The residual layer (Res) adds a skip connection to the original network layer to solve the gradient vanishing problem in the deep network training process. The upsampling layer (Upsample) refers to the use of interpolation to enlarge the original image so that it can be displayed on a higher resolution display device. The Concat layer refers to a multi-scale feature fusion operation, which usually superimposes shallow network features on deep network features to avoid contour recognition degradation caused by the deep network training process.

[0112] Next, the YOLOv8 network structure diagram outputs the prediction box of each grid unit. The target is located and the character is recognized based on the output value. The target is located in the form of a prediction box, which is the predicted position of the prediction box. Predicted size Calculate the target location confidence c:

[0113]

[0114] Wherein, p(Object) indicates whether there is a center point of the target (Object) in the grid unit. If so, p(Object) = 1, if not, p(Object)t = 0; represents the intersection-over-union ratio of the predicted box and the true box area; area(t) represents the area of ​​the true box; area(p) represents the area of ​​the predicted box, which is determined by the predicted position and predicted size Find the area.

[0115] Set the target location confidence threshold M. If the target location confidence c>M, add it to the prediction box set D for character recognition and predict the category. Complete character recognition, specifically calculate the confidence probability value s of the i-th character based on the obtained prediction box i :

[0116] s i =p i ×c (6),

[0117] Among them, p i It represents the conditional probability that the feature value in the current prediction box meets the i-th type of character, and c represents the confidence of the current prediction box.

[0118] Take the maximum confidence probability value s i The corresponding character class i is the character recognition output C:

[0119] C = argmax(s i ) (7),

[0120] Finally, according to the YOLOv8 network recognition model output Calculate the model loss function and provide feedback for parameter adjustment.

[0121] The loss function consists of three parts: the coordinate size error between the predicted box and the true box, the confidence error between the predicted box and the true box, and the character recognition classification error.

[0122] The coordinate size error calculation formula between the predicted box and the real box is:

[0123]

[0124] Among them, λ coord Represents the set position coefficient of the prediction box, and its value is set to a constant 5; Used to determine whether the jth predicted box in the i-th grid unit is responsible for the target obj, and the real box of the target obj The largest prediction box is responsible for the target obj. If it is responsible, then Otherwise, 0; Indicates the position and size of the prediction box; (xi ,y i ), (w i ,h i ) represents the position and size of the real box.

[0125] The confidence error calculation formula between the predicted box and the true box is:

[0126]

[0127] Among them, λ noord It represents the coefficient when there is no target obj in the grid cell, and its value is set to a constant of 0.5; Used to determine whether the jth prediction box in the i-th grid unit is not responsible for the target obj. If not, Otherwise 0; c i Represents the confidence of the real box; Represents the confidence of the prediction box.

[0128] The category conditional probability error calculation formula is:

[0129]

[0130] Among them, p i (C) represents the true probability that the i-th grid cell belongs to class C; The prediction for the i-th grid cell is the conditional probability of class C.

[0131] loss=loss1+loss2+loss3 (11),

[0132] According to the loss function error loss feedback gradient, each weight value in the convolution kernel is adjusted and the weight is updated until the error is less than the set threshold. Then the model parameters are fixed and the YOLOv8 network recognition model is output.

[0133] Step 4: Compress the trained YOLOv8 network recognition model, and obtain the recognition result of the mobile power inspection evidence image recognition based on the compressed YOLOv8 network recognition model.

[0134] Since the marketing audit business has different requirements for image recognition, such as power system screenshots and images uploaded by high-speed scanners, they are completed by the system-side image and text recognition service module, and the image resolution is high, the background scene is simple, and the background resources are relatively abundant. In particular, the document images uploaded by high-speed scanners often contain manual handwriting content, so the accuracy of character recognition can be improved through a large-scale model. The overall model is constructed as follows: Figure 2As shown in Output1, in order to express the discrimination, this embodiment describes the obtained model as a deep YOLOv8 network recognition model; and when taking images on-site through mobile applications, there are requirements such as relatively low image resolution, complex background scenes, tight mobile device resources and high timeliness. Therefore, it is necessary to compress the deep YOLOv8 network recognition model under the premise of ensuring recognition accuracy.

[0135] This embodiment uses structural design combined with knowledge distillation to compress the deep YOLOv8 network recognition model. Figure 2 As shown in the figure, the network structure and weight parameters of layer 1 are retained to realize the extraction of basic image data and feature information. The image recognition features are derived from the output of the next layer of layer 1, and 4 shallow network layers are constructed to learn the output implicit features of layers 2-5. Finally, the mobile image recognition result is obtained by output Output2.

[0136] In the process of extracting the output implicit features of layers 2-5 as shallow network learning parameters, the network structure of layers 2-5 is compressed by structural design decomposition. Taking layer 2 as an example, its output data is 64×16×256, which can be expressed as a three-dimensional matrix κ 64,16,256 , the structural design decomposes 3 CBL shallow network representations:

[0137] κ 64,16,256 ≈A 64,3×3 B 16,3×3 C 258,1×1 (12)

[0138] Among them, A 64,3×3 , B 16,3×3 , C 258,1×1 Represents the compressed CBL shallow network;

[0139] The implicit data based on the output of layer 1 is used as the input data of the shallow network of layer 2, the output implicit data of layer 2 64×16×256 is used as the learning target, and the goal is to minimize the loss function L. The shallow network parameters of layer 2 are adjusted by gradient feedback. It is a three-connected CBL shallow network with 64 3×3 convolution kernels, 16 3×3 convolution kernels and 256 1×1 convolution kernels. The output values ​​of the three shallow networks are merged as the input implicit data of the deep layer 3, that is, the shallow network layer of layer 2 replaces the original deep layer 2 network layer, and the soft parameter adjustment loss function and the hard parameter adjustment loss function are constructed to jointly feedback and adjust the parameters of the three shallow networks.

[0140] The loss function is:

[0141]

[0142] in, Represents the character recognition probability output after the shallow layer 2 network layer replaces the original deep layer 2 network layer, L hard is the loss value obtained based on the actual character recognition result, C i is the actual category of the i-th character, with values ​​of 0 and 1, L soft The loss value is calculated based on the output probability of the deep YOLOv8 network recognition model and the output probability of the model after the shallow layer 2 is replaced. The similar cross entropy is used and the recognition training loss is L cls , μ is the weight adjustment parameter.

[0143] The loss function constructed based on the intermediate hidden layer learning objective is:

[0144]

[0145] Among them, the target loss of the intermediate hidden layer learning is L reg , represents the loss function generated during the learning process of the shallow network on the output of the intermediate layer, h s and h s ′ represents the sth hidden layer output value of the deep layer and the shallow layer 2 output respectively, so the loss function of layer 2 compression parameter adjustment is:

[0146] L=λL cls +(1-λ)L reg (17)

[0147] Among them, λ is the adjustment ratio;

[0148] Feedback adjustment is performed based on the target loss function L until the loss function is lower than the set threshold, completing the compression of the shallow network recognition model of layer 2. The network compression of layers 3 to 5 is consistent with the principle of layer 2.

[0149] After compressing the deep YOLOv8 network recognition model based on knowledge distillation, the lightweight recognition model with Output2 representing the output result compresses the network capacity from 95.48MB to 31.53MB. The "knowledge" of the deep YOLOv8 network recognition model is shallowly learned, including the hidden "dark knowledge" of the middle layer, thereby ensuring the recognition accuracy of the recognition model.

[0150] Experiment and analysis

[0151] The present invention is based on 2642 pictures to be identified for different businesses obtained from the marketing on-site audit in a certain area of ​​a certain city from August to December 2022 as a sample set, including system screenshot recognition samples, electric energy indication number recognition samples, and on-site transformer identification plate text recognition samples. Table 1 shows the information of the supporting infrastructure project of the 10kV charging pile power grid in a certain district of a certain city. The specific businesses involved include user information review of unallocated meter reading sections, total abnormality review of reverse active power, over-capacity electricity consumption audit, and abnormal verification of dismantled electric energy indication numbers, as shown in Table 1:

[0152] Table 1 Detailed information on overhead line project support

[0153]

[0154]

[0155] In order to clearly illustrate the effectiveness of the method described in the present invention, the present invention uses transformer identification as a specific case for identification example description, and for different specific business processing processes, uses the above four types of business as cases to illustrate the effectiveness.

[0156] Before training the deep YOLOv8 network recognition model, the captured images are first normalized to a size of 1024×512, 450 images are randomly extracted as the training sample set, and the remaining 192 images are the test sample set. Due to the small number of training samples actually obtained, it is difficult to meet the needs of sufficient training of deep neural networks. Therefore, the present invention uses data enhancement to expand the sample set based on 450 training sample set images and historical archived texts, and uses similarity and diversity as two indicators to evaluate the quality of generated samples. The specific indicators are defined as follows:

[0157] (1) Similarity SSIM index

[0158]

[0159] Among them, the similarity index defines structural information from the perspective of image composition as independent of brightness and contrast, reflecting the properties of the structure of objects in the scene, and models distortion as a combination of three different factors: brightness, contrast, and structure. The mean μ is used as an estimate of brightness, the standard deviation σ is used as an estimate of contrast, and the covariance cov is used as a measure of structural similarity. The closer the SSIM index value is to 1, the more similar the generated training sample image is to the real sample image.

[0160] (2) Diversity indicator Distinct (abbreviated as Dist):

[0161]

[0162] Among them, there are a total of n types of characters in the power transformer identification plate sample library, C n Indicates the actual number of characters. It indicates the number of character categories in the generated training samples. The more balanced the distribution of characters in the generated training samples, the higher the entropy value, and the closer the Dist value is to 0, indicating that the diversity of the generated samples is higher.

[0163] Based on the current research status of sample generation, three methods are selected for comparison, as follows:

[0164] Method 1: Generate expanded training sample data based on randomly shuffled characters;

[0165] Method 2: Generate extended training sample data based on Encoder+GAN, that is, extract training sample text data through Encoder as GAN data input, generate extended training samples through the generator, and use the discriminator to distinguish the generated extended training samples from the real sample data;

[0166] Method 3: Generate expanded training sample data based on CutMix without the enhanced feature module;

[0167] Method 4: The method of the present invention generates expanded training sample data.

[0168] Figure 3 and Figure 4 The similarity index SSIM and diversity index Dist evaluation results under different scene text recognition requirements are shown. Method 1 generates expanded training samples based on randomly shuffled characters. The similarity SSIM is about 55%, and the diversity index Dist is 13%. As the number of covered prediction boxes and the number of character categories increase, their values ​​are relatively stable. This is because synthetic data is difficult to fit the state of the captured image, and random extraction also manifests itself as a certain proportion of fat characters, so its effect is stable but the simulation effect is poor. Method 2 generates expanded training samples based on GAN. Since the generated data is not generated based on actual real sample data, the generated data has Instability, especially with the increase of the number of covered prediction boxes and characters, the decline ratio increases significantly; Method 3 generates expanded training sample data based on CutMix without including the data enhancement module, and the similarity index SSIM of the generated samples is similar to that of the method of the present invention, but the diversity index Dist is worse than that of the method of the present invention, indicating that the data enhancement extraction module can effectively extract the power word segmentation features; Method 4 is the method of the present invention, which maximizes the available information of the power word segmentation of the relevant scenes through the data enhancement extraction module, and retains the image feature influence of the real training samples, thereby showing its superiority in the similarity index SSIM and the diversity index Dist. Therefore, the training samples expanded by the method of the present invention have better diversity and rationality.

[0169] Model identification results and analysis

[0170] The training parameters for defining the YOLOv8 network recognition model are shown in Table 2.

[0171] Table 2 Training parameters of YOLOv8 network recognition model

[0172]

[0173] The model evaluation index consists of two parts, namely the prediction box positioning evaluation index and the character recognition evaluation index, which are defined as follows:

[0174] (1) Prediction box positioning evaluation index S:

[0175]

[0176] Where S represents the average value of the intersection-over-union ratio between the predicted box and the true box.

[0177] (2) Character recognition evaluation indicators

[0178] The accuracy of text character recognition F1:

[0179]

[0180] Among them, N1(True) represents the number of correctly recognized text characters, N1(Total) represents the total number of text characters, and the test set N1(Total)=13231.

[0181] Text character line recognition accuracy F2:

[0182]

[0183] Among them, N2(True) represents the number of text character lines that are correctly recognized, and N2(Total) represents the actual total number of text character lines, where N2(Total) of the test set = 2313.

[0184] The present invention takes transformer identification plate text recognition as an example. Based on a sample library without transformer identification plate text recognition, the current general YOLOv8 method is used to construct a recognition model, and it is used to recognize the transformer identification plate text. Since the transformer identification plate characters contain power text and the concave and convex characteristics of steel stamps, the current general recognition method is not ideal.

[0185] Based on the transformer identification plate image currently taken on site, the present invention adopts data enhancement to expand the sample set by training samples, and trains the YOLOv8 network recognition model based on the expanded training sample set. The recognition effects of the prediction frame, stamped text and overall character are better than those of the general recognition method. The specific recognition comparison effects are shown in Table 3:

[0186] Table 3 Comparison of recognition results of different recognition models

[0187]

[0188] Method 1: The current general YOLOv8 method is used to build a recognition model, and the model training optimization is not performed based on the transformer sign text recognition sample library;

[0189] Method 2: Use the YOLOv8 method to build a recognition model, and optimize the model training based on the transformer sign text recognition sample library;

[0190] Method 3: The YOLOv8 method is used to build a recognition model. After expanding the sample library of transformer identification plate text recognition, the model training optimization is performed based on real samples and simulation samples.

[0191] Method 4: Based on the optimization model of Method 3, the structural design described in the present invention is combined with knowledge distillation to achieve model network compression, and a shallow YOLOv8 network recognition model is constructed.

[0192] It can be seen from Table 2 that method 1 directly uses the general YOLOv8 method to detect the characters on the transformer identification plate. Its prediction box positioning evaluation index, text character recognition accuracy and text character line recognition accuracy cannot meet the application standards of specific power scenarios. The self-trained text character recognition models represented by methods 2-4 have significantly improved the above three indicators, indicating that the self-training network can learn to meet the specific power scenarios.

[0193] The sample of the application scenario is the information of the present invention, but due to the lack of real training sample data represented by method 2, the model optimization effect is limited; method 3 is based on real samples and simulation samples to jointly optimize the YOLOv8 network recognition model for secondary optimization, and its prediction box positioning evaluation index and character recognition evaluation index are the best, but its network parameter scale is large, which is not suitable for on-site photography of mobile terminal applications. Therefore, method 4 of the present invention uses structural design combined with knowledge distillation to compress the model trained by method 3. Compared with the deep YOLOv8 network recognition model, the prediction box positioning evaluation index, text character recognition accuracy and text character line recognition accuracy of the shallow YOLOv8 network recognition model are all reduced to within 2%, but the model parameter scale is only one-third of the original, and the average recognition time efficiency is improved, which is only half of the time consumed by the original model. Therefore, the model constructed by method 3 is suitable for the system side, and the model constructed by method 4 is suitable for the mobile terminal.

[0194] Comparison of the application effects of intelligent auditing in power inspection

[0195] In order to demonstrate the feasibility and applicability of the method described in the present invention, a regional pilot application was carried out in a certain power company, and four types of intelligent audit business scenarios in the current online audit were selected, namely, the user information audit of unassigned meter reading segments, the total abnormality audit of reverse active power, the audit of overcapacity power consumption, and the abnormality verification of the number of returned electric energy. The specific audit rules are described as follows:

[0196] (1) Review of user information for unassigned meter reading segments: Regularly identify and verify the user's meter reading segment number in the screenshot of the marketing archive system. If a meter reading segment has been assigned, it will be marked as automatically approved;

[0197] (2) Review of abnormal reverse active power. Identify the frozen data collected on the last day of the previous month in the screenshot of the collection system. If the total forward active power is 0 and the total reverse active power is not zero, it is considered abnormal.

[0198] (3) Inspection of over-capacity electricity consumption: The mobile terminal uploads the factory number and nameplate capacity identified by the on-site photo. The system side identifies and selects the only on-site transformer photo that can be used for judgment. The system side takes a screenshot to identify the user's electricity capacity information, determines that the electricity consumption conditions for breach of contract are met, and automatically approves the review;

[0199] (4) Check the abnormal readings of the electric energy meter and compare them with the system screenshots in the marketing system. Check whether the data of the electric energy meter image taken on site is consistent. If they are consistent, the meter will be automatically approved. If they are inconsistent, the system will identify and select the only abnormal photo of the electric energy meter image taken on site that can be used for judgment.

[0200] Table 4 Comparison of the application effects of intelligent audit of power inspection

[0201]

[0202] When facing different business scenarios, according to the specific business processes, the deep or shallow YOLOv8 network recognition model for the specific identification sample library is selected. Taking the transformer nameplate as an example, the mobile terminal embeds a shallow YOLOv8 network recognition model to identify the factory number and nameplate capacity of the taken picture, and uploads multiple recognizable pictures and recognition results to the system. The system side further uses the deep YOLOv8 network recognition model on the uploaded picture to identify the transformer nameplate, and together with the electricity consumption data of the electricity customer identified by the system screenshot on the system side, enter the customized process of over-capacity electricity consumption inspection to automatically identify over-capacity electricity users.

[0203] As shown in Table 4, the present invention sets the business volume of the above four types of business to 50 cases each, and compares the method time efficiency by taking the average total manual time consumption of a single business case and the average total automatic review time consumption. The total automatic review time consumption of the four types of business is reduced to 0.87%, 0.47%, 0.52% and 0.45% of the total manual review time consumption respectively, and the review efficiency is improved by a hundred times. The review accuracy rate is not less than 80%, especially for the abnormal verification of the electric energy indication number, the image recognition is digital recognition, so the accuracy rate is the highest, and only one case of review error is found in the 50 business processing cases; the transformer identification plate recognition is mostly engraved, so the recognition accuracy rate is low, but for clear and standardized photos, it can be accurately recognized, and the review accuracy rate reaches 82%. 8 cases that cannot be automatically reviewed enter the manual review process, which still effectively reduces the labor cost and improves the review efficiency.

[0204] Embodiment 2

[0205] This embodiment provides a power audit evidence recognition system based on data enhancement retraining, including:

[0206] The sample classification module is used to obtain power inspection evidence pictures, classify them according to the power text recognition type, and obtain multiple types of power inspection evidence real training sample sets;

[0207] A sample expansion module is used to expand each type of power audit evidence real training sample set to generate multiple types of power audit evidence simulation training sample sets; combine multiple types of power audit evidence real training sample sets and multiple types of power audit evidence simulation training sample sets to obtain a power audit evidence material identification sample library;

[0208] A recognition model retraining module, which is used to train a YOLOv8 network recognition model based on a sample library of power audit supporting materials recognition, wherein the YOLOv8 network recognition model includes a shallow model and a deep model, and the recognition information of the deep model is learned by the shallow model in layers to obtain a trained YOLOv8 network recognition model;

[0209] The model compression module is used to compress the trained YOLOv8 network recognition model, and obtain the recognition result of the mobile power inspection evidence image recognition based on the compressed YOLOv8 network recognition model.

[0210] Embodiment 3

[0211] This embodiment provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the method for identifying power audit evidence based on data enhancement and retraining as described above.

[0212] Embodiment 4

[0213] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the method for identifying power audit evidence based on data enhancement and retraining as described above are implemented.

[0214] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage and optical storage, etc.) containing computer-usable program code.

[0215] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0216] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0217] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0218] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes of the embodiments of the above-mentioned methods. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0219] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for identifying power audit evidence based on data enhancement and retraining, characterized in that: The steps include: Obtain power inspection evidence pictures, classify them according to power text recognition types, and obtain multiple types of power inspection evidence real training sample sets; Expand the real training sample set for each type of power audit evidence to generate multiple types of power audit evidence simulation training sample sets; Combine multiple types of power audit evidence real training sample sets and multiple types of power audit evidence simulation training sample sets to obtain a power audit evidence material identification sample library; The YOLOv8 network recognition model is trained based on the electric power audit supporting material recognition sample library, wherein the YOLOv8 network recognition model includes a shallow model and a deep model, and the recognition information of the deep model is learned by the shallow model in layers to obtain the trained YOLOv8 network recognition model; The trained YOLOv8 network recognition model is compressed, and the recognition result of mobile power inspection evidence image recognition is obtained based on the compressed YOLOv8 network recognition model.

2. The method for identifying power audit evidence based on data enhancement retraining according to claim 1, characterized in that: The real training sample set for each type of power audit evidence is expanded to generate multiple types of power audit evidence simulation training sample sets, including: While keeping the structure of the real training sample set unchanged, a simulated training sample set is generated based on the CutMix data enhancement algorithm, which includes: Randomly select two power inspection evidence pictures, recorded as the first image and the second image; Determine a mixed area ratio parameter of the first image and the second image in the mixed image; Calculating the area range of the cropping frame according to the size of the first image and the mixed area ratio parameter, wherein the area of ​​the cropping region is inversely proportional to the mixed area ratio parameter; Randomly cut a rectangular area in the first image and the second image according to the determined cropping frame, place the cropped area of ​​the second image in the cropped area of ​​the first image, merge them into a new image, and form a new sample; According to the mixed area ratio parameter, the labels of the two original images are weighted averaged according to the area of ​​the cut area to calculate the labels of the new training samples, and the sample set supported by the new power audit is used as the simulation training sample set.

3. The method for identifying power audit evidence based on data enhancement retraining according to claim 1, characterized in that: The YOLOv8 network recognition model includes a shallow model and a deep model. The recognition information of the deep model is learned hierarchically by the shallow model, including: After the deep model training is completed, the shallow model layer uses the output data of the deep model layer as the learning target output data, and feeds back to adjust the network parameters of the shallow model layer until the deep model layer and the shallow model layer have the same data input, and the output data error loss of the two is within the error threshold range, which means that the shallow model has completed the hierarchical learning of the deep model.

4. The method for identifying power audit evidence based on data enhancement retraining according to claim 1, characterized in that: The YOLOv8 network recognition model is trained based on the power audit supporting material recognition sample library. The YOLOv8 network recognition model includes a shallow model and a deep model. The recognition information of the deep model is learned by the shallow model in layers to obtain the trained YOLOv8 network recognition model, including: The input and output definition of the target recognition model takes the image in the recognition sample library as input, divides the image into S×S grid cells, and outputs B prediction boxes for each grid cell. Each prediction box contains the position, size, confidence, and conditional probability of different categories of the prediction box. The confidence of the target positioning is calculated based on the predicted position and predicted size of each prediction box of each grid unit, and the confidence probability value of the i-th type of character is calculated based on the obtained confidence of the target positioning, and the character recognition result is further output; According to the coordinate size error between the predicted box and the true box, the confidence error between the predicted box and the true box, and the character recognition classification error, the loss function error of the YOLOv8 network recognition model is calculated. According to the loss function error feedback gradient, each weight value in the convolution kernel is adjusted and the weight is updated until the error is less than the set threshold. Then the model parameters are fixed and the YOLOv8 network recognition model is output.

5. The method for identifying power audit evidence based on data enhancement retraining as claimed in claim 4, characterized in that: The calculation formula of the confidence of the target positioning is: The calculation formula of the confidence probability value of the i-th type of character is: s i =p i ×c Wherein, p(Object) indicates whether there is a center point of the target (Object) in the grid unit. If so, p(Object) = 1; if not, p(Object) = 0. represents the intersection-over-union ratio of the predicted box and the true box area; area(t) represents the area of ​​the true box; area(p) represents the area of ​​the predicted box, p i It indicates the conditional probability that the feature value in the current prediction box meets the conditional probability of the i-th type of character.

6. The method for identifying power audit evidence based on data enhancement retraining according to claim 4, characterized in that: The loss function error of the YOLOv8 network recognition model is: loss = loss1 + loss2 + loss3, Among them, loss1 is the coordinate size error between the predicted box and the real box, λ coord Represents the set position coefficient of the prediction box, Used to determine whether the jth predicted box in the i-th grid unit is responsible for the target obj, and the real box of the target obj The largest prediction box is responsible for the target obj. If it is responsible, then Otherwise, 0; Indicates the position and size of the prediction box; (x i ,y i ), (w i ,h i ) represents the position and size of the real frame; loss2 is the confidence error between the predicted box and the real box, λ noord Indicates the coefficient when there is no target obj in the grid cell. Used to determine whether the jth prediction box in the i-th grid unit is not responsible for the target obj. If not, Otherwise 0; c i Represents the confidence of the real box; Represents the confidence of the prediction box; loss3 is the character recognition classification error, p i (C) represents the true probability that the i-th grid cell belongs to class C; The prediction for the i-th grid cell is the conditional probability of class C.

7. The method for identifying power audit evidence based on data enhancement retraining according to claim 1, characterized in that: The compressing the trained YOLOv8 network recognition model includes: The network structure and weight parameters of the first layer of the YOLOv8 network recognition model are retained to extract the feature information of the image. The feature information of the image is output by the next layer of the first layer, and a four-layer shallow network is constructed to learn the implicit features of the outputs of the second to fifth layers. In the process of extracting the implicit features of the outputs from the second to the fifth layers as shallow network learning parameters, the output implicit features are decomposed into multiple CBL shallow network representations to compress the network structure from the second to the fifth layers.

8. A power audit evidence recognition system based on data enhancement and retraining, characterized in that: include: The sample classification module is used to obtain power inspection evidence pictures, classify them according to the power text recognition type, and obtain multiple types of power inspection evidence real training sample sets; A sample expansion module, which is used to expand the real training sample set of each type of power audit evidence to generate multiple types of power audit evidence simulation training sample sets; Combine multiple types of power audit evidence real training sample sets and multiple types of power audit evidence simulation training sample sets to obtain a power audit evidence material identification sample library; A recognition model retraining module, which is used to train a YOLOv8 network recognition model based on a sample library of power audit supporting materials recognition, wherein the YOLOv8 network recognition model includes a shallow model and a deep model, and the recognition information of the deep model is learned by the shallow model in layers to obtain a trained YOLOv8 network recognition model; The model compression module is used to compress the trained YOLOv8 network recognition model, and obtain the recognition result of the mobile power inspection evidence image recognition based on the compressed YOLOv8 network recognition model.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps in a method for identifying power audit evidence based on data enhancement and retraining as described in any one of claims 1 to 7 are implemented.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the method for identifying power audit evidence based on data enhancement and retraining as described in any one of claims 1 to 7 are implemented.

Images