Electric energy meter reading detection method and device

Abstract

The invention is suitable for the technical field of target detection, and provides an electric energy meter reading detection method and device, and the method comprises the steps: obtaining an image data set of electric energy meter reading detection; based on the image data set, training a preset electric energy meter reading detection model to obtain a trained electric energy meter reading detection model; the electric energy meter reading detection model is preset to adopt an IF-CIoU loss function; the preset electric energy meter reading detection model comprises a space depth non-step convolution module and an efficient channel attention module; the IF-CIoU loss function is used for optimizing bounding box loss of a network of the preset electric energy meter reading detection model; and obtaining a to-be-detected electric energy meter reading image, and inputting the to-be-detected electric energy meter reading image into the trained electric energy meter reading detection model to obtain a detection result. According to the method, the problems of missed reading and false detection of the electric energy meter can be solved by solving the problems of inter-channel correlation neglect, redundant information interference and adaptability which still exist in a detection model.

Description

Technical Field

[0001] The present application belongs to the field of target detection technology, and in particular, relates to a method and device for detecting electric energy meter readings. Background Art

[0002] With the rapid development and widespread application of smart grid technology, the automatic detection and identification technology of electric energy meter readings has become the key to achieving efficient management of power resources. The traditional manual reading method is not only time-consuming and labor-intensive, but also easily affected by subjective judgment, resulting in reading errors. Therefore, the research of efficient and accurate automatic electric energy meter reading detection technology is of great significance to improving the efficiency of power system management.

[0003] Object detection has gone through more than two decades of development, from early traditional methods to current deep learning-based methods. Object detection algorithms based on deep learning can be divided into two categories. The first category includes two-stage algorithms represented by Region-based Convolutional Neural Networks (R-CNN) and Fast R-CNN. These algorithms extract region proposals based on object features and then use convolutional neural networks to classify and locate objects. The second category includes single-stage algorithms represented by the YOLO series, SSD, etc. These algorithms use convolutional neural networks to directly predict the location and category of different objects. However, due to the complexity and computational intensity of object detection, many advanced detection algorithms pursue higher accuracy and complex network structures, resulting in a larger number of parameters and computational requirements, and there is still room for improvement in complex environments and small object detection.

[0004] As a typical representative of small target detection in complex environments, the detection of electricity meter readings is still affected by factors such as lighting conditions and the small numbers read. The relevant detection models still have problems such as neglect of channel correlation, redundant information interference and adaptability, which makes it impossible to effectively capture the slight differences in the electricity meter readings, and it is very easy to cause missed detection and false detection of the electricity meter readings. Summary of the invention

[0005] The embodiments of the present application provide a method and device for detecting electric energy meter readings, which can solve the problems of missed detection and false detection of electric energy meter readings by solving the problems of ignoring inter-channel correlation, redundant information interference and adaptability that still exist in the detection model.

[0006] This application is implemented through the following technical solutions:

[0007] In a first aspect, an embodiment of the present application provides a method for detecting an electric energy meter reading, comprising:

[0008] Obtaining an image dataset for electric energy meter reading detection;

[0009] Based on the image dataset of electric energy meter reading detection, the preset electric energy meter reading detection model is trained to obtain the trained electric energy meter reading detection model; the preset electric energy meter reading detection model adopts the IF-CIoU loss function; the preset electric energy meter reading detection model includes a spatial deep non-stepped convolution module and an efficient channel attention module; the IF-CIoU loss function is used to optimize the bounding box loss of the network of the preset electric energy meter reading detection model;

[0010] The electric energy meter reading image to be detected is obtained, and the electric energy meter reading image to be detected is input into the trained electric energy meter reading detection model to obtain the detection result.

[0011] In a possible implementation of the first aspect, the preset electric energy meter reading detection model is a YOLOv8n model improved on the basis of the YOLOv8n model; the spatial depth non-stepped convolution module is arranged in the backbone network and the neck network of the preset electric energy meter reading detection model, and the spatial depth non-stepped convolution module is used to replace the traditional convolution layer in the YOLOv8n model; the efficient channel attention module is arranged in the backbone network of the preset electric energy meter reading detection model, and the efficient channel attention module is used to be added before the SPPF module in the YOLOv8n model and connected to the SPPF module.

[0012] In a possible implementation of the first aspect, a backbone network of a preset electric energy meter reading detection model includes a convolution module, a first spatial depth non-stepped convolution module, a first c2f module, a second spatial depth non-stepped convolution module, a second c2f module, a third spatial depth non-stepped convolution module, a third c2f module, a fourth spatial depth non-stepped convolution module, a fourth c2f module, an efficient channel attention module and an SPPF module connected in sequence;

[0013] The neck network of the preset electric energy meter reading detection model includes a first upsampling module, a first feature fusion module, a fifth c2f module, a second upsampling module, a second feature fusion module, a sixth c2f module, a fifth spatial depth non-step convolution module, a third feature fusion module, a seventh c2f module, a sixth spatial depth non-step convolution module, a fourth feature fusion module and an eighth c2f module which are connected in sequence; wherein the SPPF module is connected to the first upsampling module and the fourth feature fusion module respectively, the third c2f module is connected to the first feature fusion module, the second c2f module is connected to the second feature fusion module; the fifth c2f module is connected to the third feature fusion module;

[0014] The neck network of the preset electric energy meter reading detection model includes a first classifier layer, a second classifier layer and a third classifier layer; wherein the sixth c2f module is connected to the first classifier layer, the seventh c2f module is connected to the second classifier layer, and the eighth c2f module is connected to the third classifier layer.

[0015] In a possible implementation of the first aspect, the spatial-depth non-strided convolution module includes a spatial-to-depth layer and a non-strided convolution layer;

[0016] The space-to-depth layer is used to downsample the first feature map input to the space-to-depth layer into multiple sub-feature maps based on a preset scale factor, and connect the multiple sub-feature maps along the channel dimension to obtain a second feature map, and then input the second feature map to the non-strided convolution layer; the first feature map includes a first preset number of channels; the second feature map includes a second preset number of channels;

[0017] The non-strided convolutional layer is used to output the third feature map.

[0018] In a possible implementation of the first aspect, the efficient channel attention module includes a global average pooling module, a one-dimensional convolution module, and a feature weighting module;

[0019] The global average pooling module is used to convert the fourth feature map input to the global average pooling module into a one-dimensional vector with a constant number of channels, and input the one-dimensional vector to the one-dimensional convolution module;

[0020] The one-dimensional convolution module is used to map the one-dimensional vector based on the adaptive convolution kernel to determine the weight of each channel of the fourth feature map;

[0021] The feature weighting module is used to multiply the fourth feature map channel by channel based on the weight of each channel of the fourth feature map to obtain a weighted fifth feature map.

[0022] In a possible implementation of the first aspect, the adaptive convolution kernel k is expressed as:

[0023]

[0024] where ψ(c) represents a function of the number of channels c, |·| odd represents the odd number closest to ·; γ and b represent preset parameters.

[0025] In a possible implementation manner of the first aspect, a preset electric energy meter reading detection model is trained based on an image data set for electric energy meter reading detection to obtain a trained electric energy meter reading detection model, including:

[0026] The image dataset of the electric energy meter reading detection is input into the backbone network of the preset electric energy meter reading detection model, and multi-scale image features are extracted from the image dataset of the electric energy meter reading detection based on the spatial deep non-strided convolution module and the efficient channel attention module;

[0027] The multi-scale image features are input into the neck network of the preset electric energy meter reading detection model, and the multi-scale image features are fused based on the efficient channel attention module to obtain the fused image features;

[0028] The fused image features are input into the head network of the preset electricity meter reading detection model, and the bounding box loss of the preset electricity meter reading detection model is optimized based on the IF-CIoU loss function to obtain the trained electricity meter reading detection model.

[0029] In a possible implementation of the first aspect, optimizing a bounding box loss of a preset electric energy meter reading detection model based on an IF-CIoU loss function includes:

[0030] Obtain original IoU loss samples; original IoU loss samples include high IoU samples and low IoU samples;

[0031] Based on high IoU samples and low IoU samples, the scaling factor ratio is used to determine auxiliary bounding boxes of different scales;

[0032] Based on the auxiliary bounding box and the real bounding box, the IF-CIoU loss function is used to optimize the bounding box loss of the preset electric meter reading detection model to obtain the IoU reconstruction loss.

[0033] In a possible implementation of the first aspect, the IF-CIoU loss function L IF-CIoU It is expressed as:

[0034] L IF-CIoU =IoU γ (L CIoU +IoU-IoU inner )

[0035] Among them, L CIoU represents the CIoU loss function; IoU represents the IoU value calculated based on the original bounding box; IoU inner represents the value of InnerIoU; γ represents the parameter that controls the degree of outlier suppression.

[0036] In a second aspect, an embodiment of the present application provides an electric energy meter reading detection device, which applies the electric energy meter reading detection method of the first aspect, including:

[0037] A data acquisition module, used to acquire an image data set for detecting electric energy meter readings;

[0038] A model training module is used to train a preset electric energy meter reading detection model based on an image data set for electric energy meter reading detection to obtain a trained electric energy meter reading detection model; the preset electric energy meter reading detection model adopts an IF-CIoU loss function; the preset electric energy meter reading detection model includes a spatial deep non-stepped convolution module and an efficient channel attention module; the IF-CIoU loss function is used to optimize the bounding box loss of the network of the preset electric energy meter reading detection model;

[0039] The detection result output module is used to obtain the reading image of the electric energy meter to be detected, and input the reading image of the electric energy meter to be detected into the trained electric energy meter reading detection model to obtain the detection result.

[0040] Compared with the related art, the embodiments of the present application have the following beneficial effects:

[0041] The electric energy meter reading detection method and device of the embodiment of the present application aims to solve the problem that the subtle differences in the electric energy meter readings cannot be effectively captured. The spatial deep non-strided convolution (SPDConv) is used to replace the strided convolution and pooling layers of the original model, thereby reducing the loss of fine-grained information and enhancing the ability to extract multi-scale features. In addition, in order to highlight key local features and suppress background interference during electric energy meter readings, an efficient channel attention mechanism (ECA) is introduced into the backbone network to adaptively adjust the weights of different feature channels, emphasize important features and suppress unnecessary information, thereby improving the ability to detect small targets. Finally, a loss function IF-CIoU specifically for such small target detection is designed to optimize the bounding box loss of the network, thereby improving the convergence speed and regression accuracy of the model, solving the problems of missed and false detection of electric energy meter readings, and improving the accuracy of electric energy meter reading detection.

[0042] The beneficial effects of the above-mentioned second aspect embodiment refer to the beneficial effects of the first aspect embodiment, which will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments or related technical descriptions will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0044] Figure 1 This is an application scenario diagram of an electric energy meter reading detection method and device provided in an embodiment of the present application;

[0045] Figure 2 It is the overall framework of the electric energy meter reading detection model provided in one embodiment of the present application;

[0046] Figure 3 It is a flowchart of an electric energy meter reading detection method provided in an embodiment of the present application;

[0047] Figure 4 It is the overall framework of the YOLOv8n model provided in one embodiment of the present application;

[0048] Figure 5 is a schematic diagram of the structure of a spatial depth non-strided convolution module provided in one embodiment of the present application;

[0049] Figure 6 is a schematic diagram of the structure of an efficient channel attention module provided in an embodiment of the present application;

[0050] Figure 7 This is an example diagram of an image data set for detecting electric energy meter readings provided in an embodiment of the present application;

[0051] Figure 8 It is a structural schematic diagram of an electric energy meter reading detection device provided in one embodiment of the present application. DETAILED DESCRIPTION

[0052] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application may also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present application.

[0053] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, wholes, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or combinations thereof.

[0054] It should also be understood that the term “and / or” used in the specification and appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0055] In addition, in the description of the present application specification and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

[0056] References to "an embodiment", "one embodiment" or "some embodiments" described in the specification of this application mean that one or more embodiments of the present application include specific features, structures or characteristics described in conjunction with the embodiment. Therefore, the statements "in one embodiment", "in some embodiments", "in some other embodiments", "in some other embodiments", etc. that appear in different places in this specification do not necessarily refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in other ways. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized in other ways.

[0057] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present application will be clearly and completely described below in conjunction with the accompanying drawings and specific implementation methods. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0058] As mentioned in the background technology, the relevant target detection algorithm uses a convolutional neural network to directly predict the location and category of different objects. The main steps are to extract features from the image and then select a classifier to classify the image. However, due to the complexity and computational intensity of target detection, many advanced detection algorithms pursue higher accuracy and complex network structures, resulting in a larger number of parameters and computational requirements. Therefore, they cannot meet the needs of low-computing edge devices. Compared with two-stage networks, it is generally believed that one-stage networks show faster detection speeds, but the accuracy of one-stage networks may be slightly lower. In the one-stage network category, the YOLO series of networks show a relatively better balance between speed and accuracy. Therefore, compared with other networks, the YOLO series of networks are more commonly used in actual scenarios. Related studies have improved the YOLO series of networks using lightweight backbone networks such as GhostNet and MobileNet, spatial pyramid pooling, adaptive column mixer (ACmix) and convolutional block attention module (CBAM). Although these studies have brought significant improvements to the field of target detection, considering the actual situation of electric meter reading detection, they still lack advantages in complex environments and small target detection.

[0059] In view of the actual situation of small target detection in complex environments such as electric energy meter reading detection, considering the influence of factors such as lighting conditions and the small numbers read, it is easy to cause problems such as missed detection and false detection. The YOLOv8n model is used for detection. Although the detection effect is better than other target detection models, the original YOLOv8n model still has many shortcomings in the field of small target detection in such complex environments:

[0060] First, due to the strided convolution and pooling layers in the CNN architecture, the YOLOv8n model is prone to losing fine-grained information during feature extraction, which will result in the inability to effectively capture the subtle differences in the meter readings, thus affecting the detection accuracy. Second, the YOLOv8n model treats different feature channels equally and ignores the correlation between channels. In the detection of meter readings, different channels may contain different useful information (such as numbers, background, etc.), and ignoring these differences leads to a decrease in model performance. Third, because the YOLOv8n model treats different feature channels equally, redundant information interference occurs, which reduces the model performance. In the detection of meter readings, this interference may come from complex backgrounds or irrelevant features, affecting the accuracy of the detection results. Fourth, although YOLOv8n is a lightweight model, it is still limited by computing resources in practical applications. In the detection of meter readings, if computing resources are limited, the performance of the model cannot be fully utilized, affecting the detection speed and accuracy. Fifth, the YOLOv8n model shows insufficient adaptability when facing complex and changing actual scenarios. For example, the detection effect of the model will be affected under different lighting conditions, shooting angles or electricity meter types.

[0061] Therefore, this application proposes an improved method of the YOLOv8n model for electricity meter reading detection. While reducing the calculation amount and parameter amount of the model, it effectively improves the accuracy and efficiency of the model in electricity meter reading detection, which has important practical significance and application value.

[0062] Reference Figure 1 , which is an application scenario diagram of the electric energy meter reading detection method and device of the present invention. The prediction method and device run in the server 102. The server 102 obtains image data and related data of the electric energy meter reading detection from the database 101, and the data covers the electric energy meter images taken under different lighting conditions and various interference conditions.

[0063] The image acquisition device 101 captures the image of the electric energy meter and uploads it to the cloud 102. For example, when the terminal 103 manually or automatically uploads an electric energy meter image, the model of the cloud 102 can extract fine-grained information in the image and return the information of the electric energy meter reading in the database to the terminal 103.

[0064] The model of the cloud 102 is the electric energy meter reading detection model proposed by the present invention. The overall framework of the model is as follows: Figure 2 shown. Figure 3 is a flow chart of a method for detecting electric energy meter readings provided in an embodiment of the present application, referring to Figure 2 and Figure 3 The electric energy meter reading detection method comprises:

[0065] Step 201: Acquire an image data set for detecting electric energy meter readings.

[0066] Exemplarily, the image dataset for electricity meter reading detection comes from a self-made independently collected electricity meter dataset and the public dataset NEU-DET. The dataset covers electricity meter images taken under different lighting conditions and various interference conditions, which effectively enhances the robustness of the improved YOLOv8n model, that is, the electricity meter reading detection model. The dataset is divided into training set, validation set and test set. The training set is used to fit the weight parameters of the model, the validation set is used to fine-tune the parameters of the model to obtain the optimal model, and the test set is used to use the obtained optimal model for final prediction and evaluation. The dataset contains 10 detection categories: frame (reading frame) and 0, 1, 2, 3, 4, 5, 6, 7, 8, 9 (specific readings).

[0067] Step 202: Based on the image data set of electric energy meter reading detection, a preset electric energy meter reading detection model is trained to obtain a trained electric energy meter reading detection model.

[0068] The preset electricity meter reading detection model adopts the IF-CIoU loss function; the preset electricity meter reading detection model includes a spatial deep non-stepped convolution module and an efficient channel attention module; the IF-CIoU loss function is used to optimize the bounding box loss of the network of the preset electricity meter reading detection model.

[0069] Among them, the spatial deep non-strided convolution module is used to extract multi-scale features of small target images and low-resolution images in the image dataset of electricity meter reading detection; the efficient channel attention module is used to extract the key local features of the image dataset of electricity meter reading detection and suppress the background environment interference of the image dataset of electricity meter reading detection.

[0070] Step 203, obtaining the electric energy meter reading image to be detected, and inputting the electric energy meter reading image to be detected into the trained electric energy meter reading detection model to obtain the detection result.

[0071] In order to solve the problem that subtle differences in the readings of the electric energy meter cannot be effectively captured, this embodiment replaces the strided convolution and pooling layers of the original model with spatial depth non-strided convolution, reduces the loss of fine-grained information, and enhances the ability to extract multi-scale features. In addition, in order to highlight key local features and suppress background interference when reading the electric energy meter, an efficient channel attention mechanism is introduced into the backbone network to adaptively adjust the weights of different feature channels, emphasize important features and suppress unnecessary information, thereby improving the ability to detect small targets. Finally, a loss function IF-CIoU specifically for such small target detection is designed to optimize the bounding box loss of the network, thereby improving the convergence speed and regression accuracy of the model, solving the problems of missed and false detection of the readings of the electric energy meter, and improving the accuracy of the electric energy meter reading detection.

[0072] In one embodiment, see Figure 2 and Figure 4 , Figure 4 The preset electric energy meter reading detection model is an improved YOLOv8n model based on the YOLOv8n model; the spatial depth non-step convolution module is set in the backbone network and the neck network of the preset electric energy meter reading detection model, and the spatial depth non-step convolution module is used to replace the traditional convolution layer in the YOLOv8n model; the efficient channel attention module is set in the backbone network of the preset electric energy meter reading detection model, and the efficient channel attention module is used to be added before the SPPF module in the YOLOv8n model and connected to the SPPF module.

[0073] Exemplarily, the backbone network of the preset electricity meter reading detection model includes a convolution module Conv, a first spatial depth non-stepped convolution module SPDConv1, a first c2f module c2f1, a second spatial depth non-stepped convolution module SPDConv2, a second c2f module c2f2, a third spatial depth non-stepped convolution module SPDConv3, a third c2f module c2f3, a fourth spatial depth non-stepped convolution module SPDConv4, a fourth c2f module c2f4, an efficient channel attention module ECA and an SPPF module SPPF, which are connected in sequence.

[0074] The neck network of the preset electricity meter reading detection model includes a first upsampling module Upsample1, a first feature fusion module Concat1, a fifth c2f module c2f5, a second upsampling module Upsample2, a second feature fusion module Concat2, a sixth c2f module c2f6, a fifth spatial depth non-stepped convolution module SPDConv5, a third feature fusion module Concat3, a seventh c2f module c2f7, a sixth spatial depth non-stepped convolution module SPDConv6, a fourth feature fusion module Concat4 and an eighth c2f module, which are connected in sequence; wherein the SPPF module is connected to the first upsampling module Upsample1 and the fourth feature fusion module Concat4 respectively, the third c2f module c2f3 is connected to the first feature fusion module Concat1, the second c2f module c2f2 is connected to the second feature fusion module Concat2; the fifth c2f module c2f5 is connected to the third feature fusion module Concat3.

[0075] The neck network of the preset electricity meter reading detection model includes a first classifier layer Detect1, a second classifier layer Detect2 and a third classifier layer Detect3; wherein, the sixth c2f module c2f6 is connected to the first classifier layer Detect1, the seventh c2f module c2f7 is connected to the second classifier layer Detect2, and the eighth c2f module is connected to the third classifier layer Detect3.

[0076] The spatial depth non-stepped convolution module may include the first to sixth spatial depth non-stepped convolution modules SPDConv1 to SPDConv6. The first to fourth spatial depth non-stepped convolution modules SPDConv1 to SPDConv4 are arranged in the backbone network of the preset electric energy meter reading detection model, and the fifth to sixth spatial depth non-stepped convolution modules SPDConv5 to SPDConv6 are arranged in the neck network of the preset electric energy meter reading detection model.

[0077] The c2f module may include first to eighth c2f modules c2f1-c2f8. The first to fourth c2f modules c2f1-c2f4 are arranged in the backbone network of the preset electric energy meter reading detection model, and the fifth to eighth c2f modules c2f5-c2f8 are arranged in the neck network of the preset electric energy meter reading detection model.

[0078] In one embodiment, the specific structure and function of the spatial depth non-strided convolution module are introduced. The spatial depth non-strided convolution module includes a spatial to depth layer and a non-strided convolution layer.

[0079] Considering that the YOLOv8n model is based on the traditional CNN architecture for target detection, the CNN architecture network is prone to fine-grained information loss and insufficient learning of effective feature representation due to the presence of strided convolution and pooling layers during feature extraction. Therefore, this embodiment is improved on the basis of the YOLOv8n model, and a spatial depth non-stepped convolution (SPDConv) composed of a spatial to depth (SPD) layer and a non-stepped convolution layer is used to replace the strided convolution and pooling layers of the original model, thereby optimizing the model's processing capabilities for small targets and low-resolution images.

[0080] The space-to-depth layer is used to downsample the first feature map input to the space-to-depth layer into multiple sub-feature maps based on a preset scale factor, and connect the multiple sub-feature maps along the channel dimension to obtain a second feature map, and then input the second feature map to the non-strided convolution layer; the first feature map contains a first preset number of channels; the second feature map contains a second preset number of channels. The non-strided convolution layer is used to output a third feature map.

[0081] See also Figure 5 , taking the preset scale factor of 2 as an example. The SPD layer downsamples the first input feature map (S, S, C1) into multiple sub-feature maps f x,y When the scale factor is 2, the first feature map X is downsampled and cut to obtain 4 sub-feature maps f of size S / 2×S / 2×C1. 0,0 、f 0,1 、f 1,0 、f 1,1 . These four sub-feature maps are connected along the channel dimension to obtain the second feature map X1 (S / 2, S / 2, 4C1). Then the feature map X1 is input into the non-strided convolution layer with C2 filters, and the third feature map X2 (S / 2, S / 2, C2) is output. Through this processing, the model extracts richer fine-grained information from the input feature map and learns more effective feature representations, which helps to improve the accuracy of electric energy meter reading detection.

[0082] This embodiment converts the spatial dimension of the feature map into a depth dimension through the SPD layer, and retains more information by increasing the number of channels. The non-strided convolution layer reduces the number of channels while maintaining the spatial dimension, thereby avoiding information loss and enabling the network to capture more refined features.

[0083] In one embodiment, the specific structure and function of the efficient channel attention module are introduced. The efficient channel attention module includes a global average pooling module, a one-dimensional convolution module and a feature weighting module.

[0084] Considering that the YOLOv8n model treats different feature channels equally, it ignores the correlation between channels, resulting in redundant information interference, thereby reducing the model performance. Therefore, this embodiment introduces an efficient channel attention mechanism ECA in the backbone part of the original model. The ECA module enables the network to better focus on important features by adaptively adjusting the weights of different feature channels, thereby improving the feature extraction capability. In addition, the ECA module performs cross-channel interaction through one-dimensional convolution, avoiding complex dimensionality reduction operations.

[0085] Exemplarily, the global average pooling module is used to convert the fourth feature map X3 input to the global average pooling module into a one-dimensional vector with a constant number of channels, and input the one-dimensional vector to the one-dimensional convolution module, such as Figure 6 As shown in , the fourth feature map X3 is changed from a matrix of size H×W×C to a vector of 1×1×C after channel global average pooling (GAP) without dimensionality reduction.

[0086] The one-dimensional convolution module is used to map the one-dimensional vector based on the adaptive convolution kernel to determine the weight of each channel of the fourth feature map X3. At this time, the ECA module uses one-dimensional convolution to realize cross-channel information interaction, and the size k of the convolution kernel is adaptively determined by mapping the channel dimension C.

[0087] Among them, the adaptive convolution kernel k is expressed as:

[0088]

[0089] where ψ(c) represents a function of the number of channels c, |·| odd represents the odd number closest to ·; γ and b represent preset parameters.

[0090] The feature weighting module is used to obtain a weighted fifth feature map X4 by multiplying the fourth feature map X3 channel by channel based on the weight of each channel of the fourth feature map X3.

[0091] For example, the ECA module obtains the weight of each channel of the fourth feature map X3 by performing a one-dimensional convolution with a size of k = 5. Finally, the normalized weights are multiplied by the original input feature map channel by channel to generate the weighted fifth feature map X4.

[0092] This embodiment enhances the model's ability to capture features without increasing the amount of complex calculations by adding an ECA module to the backbone network of the model, thereby improving detection accuracy.

[0093] Exemplarily, a process of training a preset electric energy meter reading detection model based on the structure of the preset electric energy meter reading detection model is introduced, and step 202 includes:

[0094] The image dataset of electricity meter reading detection is input into the backbone network of the preset electricity meter reading detection model, and multi-scale image features are extracted from the image dataset of electricity meter reading detection based on the spatial deep non-stepped convolution module and the efficient channel attention module.

[0095] The multi-scale image features are input into the neck network of the preset electricity meter reading detection model, and the multi-scale image features are fused based on the efficient channel attention module to obtain the fused image features.

[0096] The fused image features are input into the head network of the preset electricity meter reading detection model, and the bounding box loss of the preset electricity meter reading detection model is optimized based on the IF-CIoU loss function to obtain the trained electricity meter reading detection model.

[0097] For example, in the YOLOv8n model, the traditional CIoU (Complete Intersection over Union) adds the aspect ratio of the bounding box as a penalty term to the bounding box loss function. Although it can speed up the regression convergence process of the prediction box to a certain extent, the CIoU loss function still has the problems of aspect ratio failure and imbalance of difficult and easy samples. At the same time, the CIoU loss function has limitations in adaptability. When dealing with occlusion problems, the CIoU loss function is easily affected by the occlusion area, resulting in inaccurate evaluation results.

[0098] In this regard, the inventors proposed to integrate the ideas of InnerIoU and FocalerIoU into CIoU respectively, and designed a loss function IF-CIoU for such small target detection. IF-CIoU provides a more accurate evaluation of the overlapping area by focusing on the core part of the bounding box rather than the whole. In addition, IF-CIoU also analyzes and considers the impact of the distribution of difficult samples and simple samples in bounding box regression on the regression results. By focusing on different regression samples, the detection performance in different detection tasks is improved.

[0099] Specifically, the bounding box loss of the preset electric energy meter reading detection model is optimized based on the IF-CIoU loss function, including:

[0100] Obtain the original IoU loss samples; the original IoU loss samples include high IoU samples and low IoU samples. Based on the high IoU samples and the low IoU samples, determine the auxiliary bounding boxes of different scales using the scaling factor ratio. Based on the auxiliary bounding box and the true bounding box, use the IF-CIoU loss function to optimize the bounding box loss of the preset electric energy meter reading detection model to obtain the IoU reconstruction loss.

[0101] Considering that Inner-IoU is a loss function used to optimize bounding box regression (BBR) in the field of object detection, it aims to calculate the intersection-over-union ratio (IoU) loss through auxiliary bounding boxes, thereby improving the performance of the standard IoU loss function. The core of the above steps is to use auxiliary bounding boxes of different scales to calculate the loss for high IoU and low IoU samples, so as to effectively accelerate the bounding box regression process and improve the generalization ability and positioning accuracy of the model. Inner-IoU Loss controls the size of the auxiliary bounding box by introducing a scaling factor ratio to calculate the loss for different datasets and detectors.

[0102] Among them, the size of the auxiliary bounding box is appropriately scaled compared to the original bounding box, and the scaling factor ratio can be set to 0.70. The IoU reconstruction loss is to calculate the intersection and union between the auxiliary bounding box and the real bounding box.

[0103] Also considering that the Focaler-IoU loss function is an efficient loss function for bounding box regression in object detection, the IoU loss is reconstructed through linear interval mapping to adapt to the focus on simple / difficult samples in different tasks. By focusing on different regression samples, the detector performance in different detection tasks is improved. This is to reconstruct the IoU loss through linear interval mapping to achieve the focus on different samples. It analyzes and considers the impact of the distribution of difficult and simple samples in bounding box regression on the regression results, which is an aspect that is often overlooked in traditional IoU loss functions. Focaler-IoU makes up for the shortcomings of existing bounding box regression methods through its unique method, thereby further improving detection performance in different detection tasks.

[0104] Exemplary, IF-CIoU loss function L IF-CIoU It is expressed as:

[0105] L IF-CIoU =IoU γ (L CIoU +IoU-IoU inner ) (2)

[0106] Among them, L CIoU represents the CIoU loss function; IoU represents the IoU value calculated based on the original bounding box; IoU inner represents the value of InnerIoU; γ represents the parameter that controls the degree of outlier suppression.

[0107] Preferably, γ can be set to 0.5, which has a better effect.

[0108] The IF-CIoU loss function integrates the auxiliary bounding box concept of Inner-IoU, the dynamic focusing mechanism of different samples of Focaler-IoU, and the CIoU loss function. The IF-CIoU loss function provides a more accurate evaluation of the overlapping area by focusing on the core part of the bounding box rather than the whole, and at the same time analyzes and considers the impact of the distribution of difficult samples and simple samples in bounding box regression on the regression results. By focusing on different regression samples, the detection performance in different detection tasks can be improved.

[0109] This embodiment uses the IF-CIoU loss function to combine the advantages of the above three loss functions, connects their core functions in series, and has excellent performance on the electric energy meter dataset for reading detection. This loss function can optimize the bounding box loss of the network, thereby improving the convergence speed and regression accuracy of the model.

[0110] In one embodiment, in order to verify the effectiveness of the electric energy meter reading detection method, a model experiment analysis is carried out, and the specific implementation method of the present invention is further analyzed through the electric energy meter reading detection results.

[0111] The image dataset for the electric energy meter reading detection in this embodiment is collected independently, containing a total of 1050 images, each with a resolution of 640×640 pixels. The dataset covers images of electric energy meters taken under different lighting conditions and various interference conditions, which effectively enhances the robustness of the improved YOLOv8n model. The dataset is divided into training set, validation set and test set in a ratio of 6:2:2. The dataset contains 10 detection categories: frame (reading frame) and 0, 1, 2, 3, 4, 5, 6, 7, 8, 9 (specific readings). Figure 7 A sample of this dataset is shown.

[0112] In this experiment, the size of the input image is set to 640×640 pixels, the number of epochs in the training process is 200, the batch size is 16, and the patience is 100. In the training stage, stochastic gradient descent (SGD) is used as the optimizer, the initial learning rate is set to 0.01, and the momentum factor is 0.937. The specific experimental environment parameters are shown in Table 1.

[0113] Table 1 Experimental environment

[0114]

[0115] In order to objectively evaluate the performance of the trained electric energy meter reading detection model, this paper uses precision (Precision, P), recall (Recall, R), mean average precision (mean average precision, mAP), model parameter quantity (Params), computational effort (GFLOPs), frames per second (Frames Per Second, FPS) and model size (Size) as evaluation indicators. The calculation formulas for each evaluation indicator are as follows:

[0116]

[0117] Among them, TP, FP and FN represent the number of true positives, false positives and missed positives detected correctly, respectively; AP represents the area enclosed by the PR curve and the coordinate axis; N represents the number of categories of the electric energy meter reading images to be detected; Framenum represents the total number of electric energy meter reading images to be detected; ElapsedTime represents the total time required for detection.

[0118] In order to verify the effect of introducing the IF-CIoU loss function, SPDConv module, and ECA attention mechanism on the performance improvement of the original YOLOv8n model, this embodiment conducted a reading detection experiment on a self-built electric energy meter dataset. On the basis of the original model, the IF-CIoU loss function was first added to form model1; then the SPDConv module was introduced to form model2; and finally the ECA attention mechanism was introduced to form model3. The ablation experiment was designed for comparison and verification, and the results are shown in Table 2.

[0119] Table 2 Ablation experiment

[0120]

[0121] As shown in Table 2, after replacing the original CIoU loss function with the IF-CIoU loss function on the original model, the detection accuracy of the model is improved to 72.2%. At this time, after introducing the SPDConv module on the original model, the calculation amount, parameter amount and size of the model are reduced by 7.4%, 10.6% and 10% respectively, and the real-time performance of the original model is well maintained. After further introducing the ECA attention mechanism, the model can more effectively focus on the information of small targets, so that the mAP50 reaches 74.2%, and does not affect the lightweight performance of the model.

[0122] An attention mechanism comparison experiment is conducted to illustrate the advantages of adding the ECA module to the improved YOLOv8n model over other attention modules. First, the traditional convolutional layer of the original YOLOv8n model is replaced with the SPDConv module, and then the CIoU of the original model is replaced with IF-CIoU. Then, seven different attention mechanisms, namely the ECA module, CBAM module, GAM module, CA module, EMA module, iRMB module, and ACmix module, are introduced into the backbone network for comparative experiments. The results are shown in Table 3. As can be seen from the table, the ECA module performs best among all the attention mechanisms involved in the comparative experiment. While maintaining the good computational complexity, parameter quantity, and FPS of the electric energy meter reading detection model, the mAP50 is improved by 3.0%.

[0123] Table 3 Comparative experiments of different attention mechanisms

[0124]

[0125] In order to verify the advancedness of the electric energy meter reading detection model, the improved YOLOv8n model was compared with a variety of currently popular detection algorithms under the condition that the experimental environment and parameters remain unchanged. The experimental results are shown in Table 4. It can be observed that the accuracy, mAP50 and mAP50-95 of the improved YOLOv8n model proposed in this paper reached 71.8%, 74.2% and 42.2% respectively, which is significantly better than other popular detection algorithms. At the same time, under the premise of maintaining a low amount of calculation, parameter amount and model size, the model can still maintain high real-time performance and detection accuracy.

[0126] Table 4 Comparative experiments of different algorithms

[0127]

[0128] This example uses the public dataset NEU-DET to perform a generalization verification experiment on the improved YOLOv8n algorithm. The experimental results are shown in Table 5. Compared with the original YOLOv8n model, except for a slight decrease in recall, mAP50 increased by 1.2%, mAP50-95 increased by 0.9%, and precision increased by 0.4%. These results show that the improved YOLOv8n model has good generalization.

[0129] Table 5 Model generalization verification experiment

[0130]

[0131] The embodiment of the present application proposes an improved YOLOv8n model specifically for electricity meter reading detection. By replacing the traditional convolutional layer of the original model with the SPDConv module, while reducing the amount of model parameters, it focuses on fine-grained information and improves the network's ability to extract features of various scales, thereby enhancing the model's detection accuracy for electricity meter readings. In addition, the introduction of a lightweight channel attention mechanism ECA increases the model's attention to key local features, further improving the model's detection accuracy. This model reduces network parameters and computational complexity, and also improves the accuracy of electricity meter reading detection. A loss function IF-CIoU for small target detection is also designed. By integrating the ideas of InnerIoU and FocalerIoU into CIoU, an IF-CIoU loss function is designed to optimize the network's bounding box loss, thereby improving the convergence speed and regression accuracy of the model.

[0132] Finally, experiments were conducted on a self-made electricity meter dataset and a public dataset NEU-DET, respectively, and the best results were achieved, verifying its effectiveness and generalization.

[0133] See also Figure 8 An embodiment of the present application provides an electric energy meter reading detection device, which executes the electric energy meter reading detection method as described in the above embodiment, including a data acquisition module 301, a model training module 302 and a detection result output module 303.

[0134] The data acquisition module 301 is used to acquire an image data set for detecting electric energy meter readings.

[0135] The model training module 302 is used to train the preset electric energy meter reading detection model based on the image data set of electric energy meter reading detection to obtain a trained electric energy meter reading detection model; the preset electric energy meter reading detection model adopts the IF-CIoU loss function; the preset electric energy meter reading detection model includes a spatial deep non-stepped convolution module and an efficient channel attention module; the IF-CIoU loss function is used to optimize the bounding box loss of the network of the preset electric energy meter reading detection model.

[0136] The detection result output module 303 is used to obtain the electric energy meter reading image to be detected, and input the electric energy meter reading image to be detected into the trained electric energy meter reading detection model to obtain the detection result.

[0137] Exemplarily, the preset electricity meter reading detection model is a YOLOv8n model improved on the basis of the YOLOv8n model; the spatial depth non-stepped convolution module is arranged in the backbone network and the neck network of the preset electricity meter reading detection model, and the spatial depth non-stepped convolution module is used to replace the traditional convolution layer in the YOLOv8n model; the efficient channel attention module is arranged in the backbone network of the preset electricity meter reading detection model, and the efficient channel attention module is used to be added before the SPPF module in the YOLOv8n model and connected to the SPPF module.

[0138] Exemplarily, the backbone network of the preset electric energy meter reading detection model includes a convolution module, a first spatial depth non-stepped convolution module, a first c2f module, a second spatial depth non-stepped convolution module, a second c2f module, a third spatial depth non-stepped convolution module, a third c2f module, a fourth spatial depth non-stepped convolution module, a fourth c2f module, an efficient channel attention module and an SPPF module connected in sequence;

[0139] The neck network of the preset electric energy meter reading detection model includes a first upsampling module, a first feature fusion module, a fifth c2f module, a second upsampling module, a second feature fusion module, a sixth c2f module, a fifth spatial depth non-step convolution module, a third feature fusion module, a seventh c2f module, a sixth spatial depth non-step convolution module, a fourth feature fusion module and an eighth c2f module which are connected in sequence; wherein the SPPF module is connected to the first upsampling module and the fourth feature fusion module respectively, the third c2f module is connected to the first feature fusion module, the second c2f module is connected to the second feature fusion module; the fifth c2f module is connected to the third feature fusion module;

[0140] The neck network of the preset electric energy meter reading detection model includes a first classifier layer, a second classifier layer and a third classifier layer; wherein the sixth c2f module is connected to the first classifier layer, the seventh c2f module is connected to the second classifier layer, and the eighth c2f module is connected to the third classifier layer.

[0141] Exemplarily, the spatial-to-depth non-strided convolution module includes a spatial-to-depth layer and a non-strided convolution layer;

[0142] The space-to-depth layer is used to downsample the first feature map input to the space-to-depth layer into multiple sub-feature maps based on a preset scale factor, and connect the multiple sub-feature maps along the channel dimension to obtain a second feature map, and then input the second feature map to the non-strided convolution layer; the first feature map includes a first preset number of channels; the second feature map includes a second preset number of channels;

[0143] The non-strided convolutional layer is used to output the third feature map.

[0144] Exemplarily, the efficient channel attention module includes a global average pooling module, a one-dimensional convolution module, and a feature weighting module;

[0145] The global average pooling module is used to convert the fourth feature map input to the global average pooling module into a one-dimensional vector with a constant number of channels, and input the one-dimensional vector to the one-dimensional convolution module;

[0146] The one-dimensional convolution module is used to map the one-dimensional vector based on the adaptive convolution kernel to determine the weight of each channel of the fourth feature map;

[0147] The feature weighting module is used to multiply the fourth feature map channel by channel based on the weight of each channel of the fourth feature map to obtain a weighted fifth feature map.

[0148] Exemplarily, the adaptive convolution kernel k is expressed as:

[0149]

[0150] where ψ(c) represents a function of the number of channels c, |·| odd represents the odd number closest to ·; γ and b represent preset parameters.

[0151] Exemplarily, the model training module 302 is specifically used for:

[0152] The image dataset of the electric energy meter reading detection is input into the backbone network of the preset electric energy meter reading detection model, and multi-scale image features are extracted from the image dataset of the electric energy meter reading detection based on the spatial deep non-strided convolution module and the efficient channel attention module;

[0153] The multi-scale image features are input into the neck network of the preset electric energy meter reading detection model, and the multi-scale image features are fused based on the efficient channel attention module to obtain the fused image features;

[0154] The fused image features are input into the head network of the preset electricity meter reading detection model, and the bounding box loss of the preset electricity meter reading detection model is optimized based on the IF-CIoU loss function to obtain the trained electricity meter reading detection model.

[0155] Exemplarily, in the model training module 302, the bounding box loss of the preset electric energy meter reading detection model is optimized based on the IF-CIoU loss function, including:

[0156] Obtain original IoU loss samples; original IoU loss samples include high IoU samples and low IoU samples;

[0157] Based on high IoU samples and low IoU samples, the scaling factor ratio is used to determine auxiliary bounding boxes of different scales;

[0158] Based on the auxiliary bounding box and the real bounding box, the IF-CIoU loss function is used to optimize the bounding box loss of the preset electric meter reading detection model to obtain the IoU reconstruction loss.

[0159] Among them, the IF-CIoU loss function L IF-CIoU It is expressed as:

[0160] L IF-CIoU =IoU γ (L CIoU +IoU-IoU inner ) (2)

[0161] Among them, L CIoU represents the CIoU loss function; IoU represents the IoU value calculated based on the original bounding box; IoU inner represents the value of InnerIoU; γ represents the parameter that controls the degree of outlier suppression.

[0162] The beneficial effects of the electric energy meter reading detection device in the embodiment of the present application refer to the beneficial effects of the electric energy meter reading detection method in the above embodiment, which will not be repeated here.

[0163] It should be noted that, although several units / modules or sub-units / modules of the electric energy meter reading detection device are mentioned in the above detailed description, such division is merely exemplary and not mandatory. In fact, according to the embodiments of the present application, the features and functions of two or more units / modules described above may be embodied in one unit / module. Conversely, the features and functions of one unit / module described above may be further divided to be embodied by multiple units / modules.

[0164] In addition, although the operations of the method of the present application are described in a specific order in the drawings, this does not require or imply that the operations must be performed in this specific order, or that all the operations shown must be performed to achieve the desired results. Additionally or alternatively, some steps may be omitted, multiple steps may be combined into one step, and / or one step may be decomposed into multiple steps.

[0165] The present application also provides a computer-readable storage medium, in which computer-executable instructions are stored. When a processor executes the computer-executable instructions, the electric energy meter reading detection method provided in the above embodiment of the present application is implemented.

[0166] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0167] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. A method for detecting electric energy meter readings, characterized in that: include: Obtaining an image dataset for electric energy meter reading detection; Based on the image data set of the electric energy meter reading detection, a preset electric energy meter reading detection model is trained to obtain a trained electric energy meter reading detection model; the preset electric energy meter reading detection model adopts an IF-CIoU loss function; the preset electric energy meter reading detection model includes a spatial deep non-stepped convolution module and an efficient channel attention module; the IF-CIoU loss function is used to optimize the bounding box loss of the network of the preset electric energy meter reading detection model; The electric energy meter reading image to be detected is obtained, and the electric energy meter reading image to be detected is input into the trained electric energy meter reading detection model to obtain a detection result.

2. The method for detecting electric energy meter readings according to claim 1, characterized in that: The preset electric energy meter reading detection model is a YOLOv8n model improved on the basis of the YOLOv8n model; the spatial depth non-stepped convolution module is arranged in the backbone network and the neck network of the preset electric energy meter reading detection model, and the spatial depth non-stepped convolution module is used to replace the traditional convolution layer in the YOLOv8n model; The efficient channel attention module is arranged in the backbone network of the preset electric energy meter reading detection model, and the efficient channel attention module is used to be added before the SPPF module in the YOLOv8n model and connected to the SPPF module.

3. The method for detecting electric energy meter readings according to claim 1, characterized in that: The backbone network of the preset electric energy meter reading detection model includes a convolution module, a first spatial depth non-stepped convolution module, a first c2f module, a second spatial depth non-stepped convolution module, a second c2f module, a third spatial depth non-stepped convolution module, a third c2f module, a fourth spatial depth non-stepped convolution module, a fourth c2f module, an efficient channel attention module and an SPPF module connected in sequence; The neck network of the preset electric energy meter reading detection model includes a first upsampling module, a first feature fusion module, a fifth c2f module, a second upsampling module, a second feature fusion module, a sixth c2f module, a fifth spatial depth non-step convolution module, a third feature fusion module, a seventh c2f module, a sixth spatial depth non-step convolution module, a fourth feature fusion module and an eighth c2f module connected in sequence; wherein the SPPF module is connected to the first upsampling module and the fourth feature fusion module respectively, the third c2f module is connected to the first feature fusion module, the second c2f module is connected to the second feature fusion module; the fifth c2f module is connected to the third feature fusion module; The neck network of the preset electric energy meter reading detection model includes a first classifier layer, a second classifier layer and a third classifier layer; wherein the sixth c2f module is connected to the first classifier layer, the seventh c2f module is connected to the second classifier layer, and the eighth c2f module is connected to the third classifier layer.

4. The method for detecting electric energy meter readings according to claim 2, characterized in that: The spatial-depth non-strided convolution module includes a spatial-to-depth layer and a non-strided convolution layer; The space-to-depth layer is used to downsample the first feature map input to the space-to-depth layer into multiple sub-feature maps based on a preset scale factor, and connect the multiple sub-feature maps along the channel dimension to obtain a second feature map, and then input the second feature map to the non-strided convolutional layer; the first feature map includes a first preset number of channels; the second feature map includes a second preset number of channels; The non-strided convolutional layer is used to output a third feature map.

5. The method for detecting electric energy meter readings according to claim 2, characterized in that: The efficient channel attention module includes a global average pooling module, a one-dimensional convolution module and a feature weighting module; The global average pooling module is used to convert the fourth feature map input to the global average pooling module into a one-dimensional vector with a constant number of channels, and input the one-dimensional vector to the one-dimensional convolution module; The one-dimensional convolution module is used to map the one-dimensional vector based on an adaptive convolution kernel to determine the weight of each channel of the fourth feature map; The feature weighting module is used to multiply the fourth feature map channel by channel based on the weight of each channel of the fourth feature map to obtain a weighted fifth feature map.

6. The method for detecting electric energy meter readings according to claim 5, characterized in that: The adaptive convolution kernel k is expressed as: where ψ(c) represents a function of the number of channels c, |·| odd represents the odd number closest to ·; γ and b represent preset parameters.

7. The method for detecting electric energy meter readings according to claim 2, characterized in that: Based on the image data set for the electric energy meter reading detection, a preset electric energy meter reading detection model is trained to obtain a trained electric energy meter reading detection model, including: Inputting the image data set of the electric energy meter reading detection into the backbone network of the preset electric energy meter reading detection model, and extracting multi-scale image features from the image data set of the electric energy meter reading detection based on the spatial depth non-stepped convolution module and the efficient channel attention module; Inputting the multi-scale image features into the neck network of the preset electric energy meter reading detection model, fusing the multi-scale image features based on the efficient channel attention module to obtain fused image features; The fused image features are input into the head network of the preset electric energy meter reading detection model, and the bounding box loss of the preset electric energy meter reading detection model is optimized based on the IF-CIoU loss function to obtain the trained electric energy meter reading detection model.

8. The method for detecting electric energy meter readings according to claim 7, characterized in that: The optimizing the bounding box loss of the preset electric energy meter reading detection model based on the IF-CIoU loss function includes: Obtaining an original IoU loss sample; the original IoU loss sample includes a high IoU sample and a low IoU sample; Determining auxiliary bounding boxes of different scales using a scaling factor ratio based on the high IoU sample and the low IoU sample; Based on the auxiliary bounding box and the real bounding box, the IF-CIoU loss function is used to optimize the bounding box loss of the preset electric energy meter reading detection model to obtain the IoU reconstruction loss.

9. The method for detecting electric energy meter readings according to any one of claims 1 to 8, characterized in that: The IF-CIoU loss function L IF-CIoU It is expressed as: L IF-CIoU =IoU γ (L CIoU +IoU-IoU inner ) Among them, L CIoU represents the CIoU loss function; IoU represents the IoU value calculated based on the original bounding box; IoU inner represents the value of InnerIoU; γ represents the parameter that controls the degree of outlier suppression.

10. An electric energy meter reading detection device, characterized in that: The method for detecting electric energy meter readings according to any one of claims 1 to 9 comprises: A data acquisition module, used to acquire an image data set for detecting electric energy meter readings; A model training module, used to train a preset electric energy meter reading detection model based on the image data set of the electric energy meter reading detection to obtain a trained electric energy meter reading detection model; the preset electric energy meter reading detection model adopts an IF-CIoU loss function; the preset electric energy meter reading detection model includes a spatial deep non-stepped convolution module and an efficient channel attention module; the IF-CIoU loss function is used to optimize the bounding box loss of the network of the preset electric energy meter reading detection model; The detection result output module is used to obtain the reading image of the electric energy meter to be detected, and input the reading image of the electric energy meter to be detected into the trained electric energy meter reading detection model to obtain the detection result.

Images