General uncalibrated pointer type meter reading abnormity identification method and system

Through the calibration-free method, the image defuzzy and instance segmentation model are used, and the dial blur of the meter is determined by combining the local power spectrum slope. The problem of calibration consumption and low recognition accuracy of meter defect detection in the prior art is solved, and high efficiency and accuracy are achieved The meter reading abnormality is identified.

CN119027927APending Publication Date: 2024-11-26孙瀚 +1
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202411214504.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-01
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

The existing meter defect detection methods rely on complex feature calibration, consume human resources and are difficult to cope with calibration feature failure caused by camera preset position offset, and the AI-based method has low accuracy in identifying unusual pointer tables and cannot Meet the accuracy requirements of practical application scenarios.

Method used

A general calibration-free pointer meter reading abnormal recognition method is proposed. By collecting image samples, training object detection, image defuzzing and instance segmentation models, the dial blur of the meter is judged by the local power spectrum slope, and the defuzzing algorithm is obtained. High-definition image, fusion meter category and range information, intelligently match the range range to judge abnormal readings.

Benefits of technology

No manual calibration is required, which reduces dependence on meter defect samples, improves recognition accuracy, effectively eliminates fuzzy error detection caused by camera imaginary focus, reduces the difficulty of system tuning, and is suitable for general detection of diversity meters.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119027927A_ABST
    Figure CN119027927A_ABST
Patent Text Reader

Abstract

The invention discloses a general uncalibrated pointer type meter reading abnormity identification method and system, and relates to the field of image processing and artificial intelligence, and the method comprises the steps: collecting meter type image samples, carrying out the partial labeling, and building an initial sample library; training an image deblurring model, a target detection model and an instance segmentation model of the meter by using the sample library image; performing semi-automatic labeling by using the basic model and iteratively updating the model; and identifying the position and category of the meter through the target detection model, converting the image to a frequency domain, calculating the local power spectrum slope of the meter area, and judging whether the meter dial plate is fuzzy or not. The method provided by the invention does not need manual calibration, can automatically identify whether the general pointer type meter has the reading anomaly or not, does not depend on a large number of defect samples in model training, has the advantages of strong anti-interference capability, good model generalization and the like, and can be applied to engineering application scenes requiring meter reading anomaly detection.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the fields of image processing and artificial intelligence, and in particular to a general uncalibrated pointer meter reading abnormality recognition method and system. Background Art

[0002] Traditional meter maintenance and defect detection usually rely on manual inspections, which is a cumbersome, time-consuming and prone to missed inspections, posing a great challenge to the stable operation of the entire system. With the rapid development of artificial intelligence technology, the inspection method has shifted from traditional manual inspections to intelligent inspection systems consisting of inspection robots, aerial drones and fixed-point high-definition cameras. These intelligent means have also brought about the need for automated identification of meter defects.

[0003] Existing meter defect detection methods often rely on complex feature calibration, requiring each meter to be individually configured with meter pointer and range information, which consumes huge human resources and is difficult to effectively deal with calibration feature failures caused by camera preset position offset and other reasons. Some detection methods have also introduced artificial intelligence technology to identify meter defects through target detection models, achieving universal meter defect recognition. However, due to the diversity of meter types and the scarcity of defect scenes in real scenarios, this solution has a low recognition accuracy for uncommon pointer meters and is easily affected by factors such as lighting, weather, and shooting angle, and cannot meet the accuracy requirements in actual application scenarios. Summary of the invention

[0004] Purpose of the invention: In view of the above-mentioned problems, the present invention proposes a universal uncalibrated pointer meter reading abnormality identification method and system, which does not require the meter to be calibrated before use, and can greatly reduce the model's dependence on meter defect samples, thereby improving the system's recognition accuracy.

[0005] Technical solution: A general method and system for identifying abnormal readings of pointer-type meters without calibration of the present invention specifically comprises the following steps: Step 1: Collect meter image samples, screen and sort them, and then annotate them to establish an initial sample library; Step 2: Use the sample library to train the meter’s object detection, image deblurring, and instance segmentation models; Step 3: Semi-automatically label the remaining data and iteratively update the model; Step 4: Use the target detection model to locate the meter position and category information, convert the image to the frequency domain and calculate the local power spectrum slope of the meter area to determine whether the meter dial is blurred; Step 5: For the image determined to be blurred, a deblurring algorithm is applied to obtain a high-definition meter image; Step 6: Integrate the meter category with the meter needle, range and other regional mask information, intelligently match the current meter effective range, and determine whether there is an abnormal meter reading based on its intersection with the meter needle.

[0006] Furthermore, step 1 specifically includes: Step 1.1: Use intelligent acquisition equipment to obtain image data with meter equipment as the main body; Step 1.2: Use the minimum bounding rectangle to mark the location and category information of the meter equipment and generate a Pasal VOC format annotation file; Step 1.3: Capture the meter device area image, use pixel-level masks to annotate the meter pointer area and meter range area, and generate a Pasal VOC format annotation file; Step 1.4: Use the manual focus mode of the intelligent acquisition device to collect the blurred image of the meter dial and the corresponding clear image of the meter dial; Step 1.5: Divide the labeled image samples into training set and test set in a ratio of 9:1 to build an initial sample library.

[0007] Furthermore, the meter equipment categories described in step 1.2 are divided into three categories according to their universal characteristics, including conventional single-pointer meters, conventional double-pointer meters, multi-color range single-pointer meters, and multi-color range double-pointer meters. A total of 10,000 image samples are collected and 2,000 are annotated to form a data set for training the target detection model.

[0008] Furthermore, the meter range area described in step 1.3 refers to the sector-shaped area on the dial containing the scale. A total of 10,000 image samples are collected and 2,000 are annotated to form a data set for training the instance segmentation model.

[0009] Furthermore, in the image acquisition process described in step 1.4, an intelligent acquisition device is used to adjust the zoom parameters and fix other shooting parameters to capture normal and blurred images of the same meter to form a group of samples, and a total of 10,000 groups of image samples are obtained to form a data set for training the deblurring model.

[0010] Furthermore, the step 2 specifically includes: Step 2.1: Use the YOLO model to train the data set described in step 1.2 to obtain a basic target detection model that can locate the meter category and position; Step 2.2: Use the YOLACT model to train the data set described in step 1.3 to obtain a basic instance segmentation model that can locate the meter range and pointer area; Step 2.3: Use the DeepRFT model to train the data set described in step 1.4 to obtain a deblurring model that can improve image quality.

[0011] Furthermore, the YOLO model described in step 2 adopts the YOLOv8 model, the initial learning rate is set to 0.005, and the optimizer is AdamW; the YOLACT model adopts the YOLACT model, the initial learning rate is set to 0.001, and the optimizer is AdamW; the DeepRFT model, the initial learning rate is set to 0.00005, and the AdamW optimizer is used; Furthermore, the semi-automatic annotation described in step 3 uses the basic target detection and instance segmentation model described in step 2 to infer the unlabeled image samples, and obtains high-quality labeled data through manual review.

[0012] Furthermore, the step 4 specifically includes: Step 4.1: Use the object detection model to identify the meter area in the image; Step 4.2: Calculate the power spectrum slopes of the global image and the meter area image in the frequency domain using discrete Fourier transform; Step 4.3: Calculate the local power spectrum slope of the meter area image relative to the global image to determine whether the current meter has a meter dial blur defect.

[0013] Furthermore, the power spectrum slope calculation process described in step 4.2 is as follows: (1) in It means that the size after Fourier transform to frequency domain is Pictures of and Represents the two-dimensional frequency of the image in polar coordinates, by Directional power spectrum The summation yields the following approximate representation: (2) in is the amplitude proportional factor in each direction. By transforming equation (2), the frequency domain component, i.e., the power spectrum slope, can be calculated. : (3) Furthermore, the calculation process of the local power spectrum slope described in step 4.3 is as follows: (4) in Indicates the frequency spectrum slope of the current meter area, represents the frequency spectrum slope of the global image. Both are calculated by formula (3). The ratio of the two is It is the slope of the local power spectrum. If it is greater than the threshold 1, it is considered that there is a dial blur defect in this meter area.

[0014] Furthermore, the step 6 specifically includes: Step 6.1: Use the instance segmentation model on the meter image to obtain the area information of the meter needle and range; Step 6.2: Filter and detect invalid instances in the needle and range according to the post-processing strategy; Step 6.3: Integrate the meter category detected in step 4.1, intelligently match the normal range of the current meter, and determine whether the meter needle has an intersection with it. If no intersection is detected, the current meter has an abnormal meter reading defect.

[0015] Furthermore, the post-processing strategy described in step 6.2 uses image morphological methods to perform contour optimization, pixel connection and outlier region filtering on the segmentation results, and filters out segmentation instances that do not conform to the rules to improve segmentation accuracy.

[0016] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: Compared with the meter defect detection method based on feature calibration, the present invention does not require manual calibration for each meter, thereby improving the application efficiency of the system; Compared with the meter defect detection method based on target detection, the present invention does not require a large number of meter defect image samples and can be efficiently expanded to various general meters; The meter dial blur detection algorithm based on the local power spectrum slope proposed in the present invention can effectively filter out the dial blur misdetection samples caused by the camera out-of-focus, and combined with the deblurring model, can effectively improve the accuracy of meter defect detection; The present invention realizes the calibration-free recognition of abnormal readings of general pointer meters by fusing target detection and instance segmentation, thereby improving the recognition accuracy of the system and reducing the difficulty of system tuning. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a schematic diagram of the process of the present invention; Figure 2 This is a schematic diagram of the needle and range interval marking of the present invention; Figure 3 It is the network structure diagram of the YOLOv8 model of the present invention; Figure 4 It is a schematic diagram of the structure of the Res FFT-Conv module of the present invention; Figure 5 It is the network structure diagram of the YOLACT model of the present invention; Figure 6 It is a schematic diagram of the normal measuring range of the present invention. DETAILED DESCRIPTION

[0018] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments: The present invention provides a universal uncalibrated pointer meter reading abnormality identification method and system. The method first uses the collected image samples to train the meter's target detection, image deblurring and instance segmentation models, uses the target detection model to obtain the meter area and category, determines whether the meter dial is blurred by the local power spectrum slope, and uses the deblurring model to obtain a high-definition meter image. Finally, the segmentation results of the instance segmentation model are integrated to determine whether the meter reading is abnormal based on the intersection of the meter pointer and the normal range interval. Figure 1 As shown, the specific implementation of the present invention includes the following steps: Step 1: Collect meter image samples, screen and sort them, and then annotate them to establish an initial sample library. The specific implementation steps are as follows: Step 1.1: Use intelligent acquisition equipment to obtain image data with meters as the main body, including meter images taken by intelligent patrol drones, patrol robots, fixed-point high-definition cameras, and patrol videos containing meters; Step 1.2: Filter and process the collected images and videos. This step is implemented through the following sub-steps: Get images by taking frames from the captured video, name them according to time and save them in a fixed directory; Use the Structural Similarity (SSIM) matching algorithm to filter out image samples with similarity higher than a threshold of 0.85; Through manual screening, low-quality images caused by exposure, focus, etc. are filtered out; Step 1.3: Use the minimum bounding rectangle to mark the location and category information of the meter equipment of some image samples, and generate a labeling file in Pasal VOC format. In this embodiment, four types of meter equipment are used, and the corresponding labeling labels for each type of meter are shown in Table 1: Table 1: Meter categories and corresponding label tables

[0019] Step 1.4: Capture the meter equipment area image, rename the image in a unified format, and save it to a fixed directory. The naming format is "{original image name}_truncated_{xmin}_{ymin}_{xmax}_{ymax}.jpg", where xmin, ymin, xmax, and ymax represent the coordinate information of the meter equipment in the original image; Step 1.5: Use pixel-level masks to annotate the pointers and meter range areas of some image samples and generate annotation files in PasalVOC format. Figure 2 As shown, when marking the needle, only the area containing the needle is marked with a mask, and when marking the meter range, the entire sector area on the dial containing the scale needs to be marked.

[0020] Step 1.6: Use the intelligent acquisition device to adjust the zoom parameters and fix other shooting parameters under fixed parameter configuration to shoot normal and blurred images of the same meter to form a group of samples.

[0021] Step 1.7: Divide the labeled image samples into training set and test set in a ratio of 9:1 to build the initial sample library. During the division process, the structural similarity matching algorithm is used to ensure that the similarity between the images in the validation set is less than the threshold of 0.5.

[0022] The initial sample library constructed in this embodiment includes three data sets: Dataset 1: Contains 2,000 image samples labeled with meter locations and categories and 8,000 unlabeled image samples, used to train the initial object detection model; Dataset 2: Contains 2,000 image samples annotated with meter pointer and range mask information and 8,000 unlabeled image samples, used to train the initial instance segmentation model; Dataset 3: 10,000 sets of images containing blurred images and corresponding clear images, used to train the deblurring model. Step 2: Use the sample library to train the meter's object detection, instance segmentation, and image deblurring models. The specific implementation steps are as follows: Step 2.1: Train the YOLOv8 model on Dataset 1 to identify meter categories and locate locations. In this embodiment, the initial learning rate is set to 5e-3 and the optimizer is AdamW.

[0023] Step 2.2: Train the YOLACT model on Dataset 2 to locate the meter range and pointer area. In this embodiment, the initial learning rate is set to 1e-3 and the optimizer is AdamW.

[0024] Step 2.3: Train the DeepRFT model on dataset 3 to clarify the blurred image of the meter dial. In this example, the initial learning rate is set to 5e-5 and the optimizer is AdamW.

[0025] Step 3: Use the trained object detection and example segmentation model to semi-automatically label the remaining unlabeled data, and use the updated data set to iteratively update the model, which can improve labeling efficiency and reduce labor costs. The specific implementation steps are as follows: Step 3.1: Model reasoning and data verification: Use the basic model trained in steps 2.1 and 2.2 to infer the remaining unlabeled data. After the labelers review and correct the reasoning results, the label files are added to datasets one and two.

[0026] Step 3.2: Model iterative update: Use datasets 1 and 2 containing 10,000 labeled data to train the model to obtain a YOLOv8 target detection model and a YOLACT instance segmentation model with better recognition accuracy.

[0027] Step 4: Use the target detection model to locate the meter position and category information, convert the image to the frequency domain and calculate the local power spectrum slope of the meter area to determine whether there is a defect of blurred meter dial. The specific implementation steps are as follows: Step 4.1: Use the improved YOLOv8 target detection model on the image to identify the meter area and meter category. The model structure diagram is as follows Figure 3 As shown in the figure, it mainly includes backbone network module, feature fusion module and result output module. The specific details of the model are as follows: The backbone network module of the YOLOv8 model described in the present invention is based on the CSP Darknet-53 network, which is composed of a DarkNet-53 network and a CSP structure.

[0028] The DarkNet-53 network consists of 53 convolutional layers. The deeper network path is conducive to capturing complex features in the image, and the introduction of residual connections effectively solves the gradient vanishing problem encountered during deep network training, improving the performance and stability of the network.

[0029] The CSP (Cross Stage Partial) structure is introduced on the basis of DarkNet-53, and cross-stage partial links are introduced into the network. The idea of ​​the CSP structure is to divide the input feature map into two branches, one branch is processed through a deeper network path, and the other branch is directly short-circuited. The features of the two branches are then fused to form an output feature map, which realizes the information exchange between features of different dimensions in the network and gives the network stronger nonlinear expression capabilities.

[0030] The present invention specifically uses the C2F (CSP Bottleneck with 2 convolutions) structure for CSP cross-stage linking. This module merges two branches through an additional Split operation and transmits the post-joining features in the form of residual links, thereby reducing the computational complexity and memory consumption of the model.

[0031] The feature fusion module of the YOLOv8 model described in the present invention is based on a spatial pyramid pooling layer (Spatial Pyramid Pooling, SPP) and a path aggregation network (Path Aggregation Network, PAN).

[0032] (2.1) Pooling operations are performed on the feature maps of three different scales at the same target scale, and the high-order features extracted by pooling are concatenated to achieve the function of condensing features of different levels into a single feature map. Through the spatial pyramid pooling layer, the model makes up for the underlying pixel features lost in the high-order features for small target detection, improving the model's ability to detect small targets.

[0033] (2.2) The original target detection algorithm is to detect on the feature map of the last convolution output. When the network is deep, the underlying semantic information will be lost, resulting in the problem of low recognition accuracy of small target objects. Therefore, the model introduces the PAN feature fusion module to solve this problem. In view of the complex scenes in which general pointer meters are located and the small differences in features between different categories, in order to improve the model's detail perception ability and the recognition effect of small meters, the PAN module in the present invention processes four different levels of features at the same time. This module is a deeper feature fusion structure based on the SPP layer. The difference between it and the SPP layer is that it uses a bottom-up feature flow path to perform layer-by-layer feature fusion, and after fusing the four layers of feature maps, it uses a top-down feature path to refine the features layer by layer, so as to deeply fuse semantic information of different scales and generate four high-order feature maps of different scales, further strengthening the high-order features used to detect the output head structure.

[0034] (3) The result output module of the YOLOv8 model described in the present invention uses a decoupling head, which is composed of a classification head for implementing the classification of candidate boxes and a regression head for implementing the regression task.

[0035] (3.1) Each classification head consists of two 3x3 convolutions and two 1x1 convolutions, and the error between the predicted category of the candidate box and the true category is minimized by the BCE loss function (Binary Cross Entropy) shown in formula (1): (1) in Indicates that the model predicts that the candidate box is the current The probability of the categories, Indicates Class binary labels are 0 or 1.

[0036] (3.2) Each regression head consists of two 3x3 convolutions and two 1x1 convolutions. The loss function is obtained by summing the two loss functions of equations (3) and (4) to minimize the error between the predicted boundary of the candidate box and the actual boundary: (2) (3) (4) In formula (3) The DFL loss (Distribution Focal Loss) of the regression box is a loss function that assumes that the distribution of the regression box should be close to the true value, and is used to make the model learn that the probability density of the bounding box is as large as possible at the true value. represents the bounding box probability distribution learned by the model, represents the bounding box coordinates predicted by the model, and Represents the distance in the probability distribution The two closest points.

[0037] In formula (4) The CIoU loss (Complete Intersection over Union Loss) of the regression box is a loss function that takes into account three geometric factors: overlapping area, center point distance, and aspect ratio. It aims to more comprehensively evaluate the difference between the predicted box and the true box and provide more accurate regression results. and They represent the intersection of the predicted box and the true box and the Euclidean distance of the center point respectively. Represents the diagonal distance of the minimum closure area that can contain both the predicted box and the true box. and Represents the width and height of the box respectively.

[0038] Step 4.2: Use discrete Fourier transform to calculate the power spectrum slope of the global image and the meter area image in the frequency domain. The calculation process is as follows: (5) in It means that the size after Fourier transform to frequency domain is Pictures of and Represents the two-dimensional frequency of the image in polar coordinates, by Directional power spectrum The summation yields the following approximate representation: (6) in is the amplitude proportional factor in each direction. By transforming equation (2), the frequency domain component, i.e., the power spectrum slope, can be calculated. : (7) Step 3.3: Calculate the local power spectrum slope of the meter area image relative to the global image to determine whether the current meter has a meter dial blur defect. The local power spectrum slope calculation process is as follows: (8) in Indicates the frequency spectrum slope of the current meter area, Represents the frequency spectrum slope of the global image, the ratio of the two It is the slope of the local power spectrum. If it is greater than the threshold 1, it is considered that there is a dial blur defect in this meter area.

[0039] Step 5: For the blurred image area of ​​the meter dial, apply the DeepRFT image deblurring model to obtain the meter image with improved clarity. If the meter dial blur defect is not detected in step 3, there is no need to perform step 4. The blur in the meter defect recognition scene is usually caused by camera defocus. Since the conventional image deblurring algorithm is built based on the motion blur dataset, the effect of directly applying it to the defocused and blurred scene is poor. Therefore, the present invention trains the DeepRFT model (Deep Residual Fourier Transformation) on dataset three to obtain better meter image deblurring effect. The model realizes the end-to-end deblurring function of a single image by inserting the Res FFT-Conv module into the MIMO-UNet architecture. The specific details of the model are designed as follows: The MIMO-UNet architecture is a U-Net architecture that supports multiple inputs and multiple outputs. The encoder and decoder are composed of three encoder blocks (Multi-Input Single Encoder, MISE) and three decoder blocks (Multi-Output Single Decoder, MOSD), respectively. Each block is designed to process image features of a specific scale, and then the multi-scale information flow between the encoder and decoder is merged through the asymmetric feature fusion mechanism (Asymmetric Feature Fusion, AFF) to improve the model performance.

[0040] The structure of the Res FFT-Conv module is as follows Figure 4As shown in the figure, a frequency domain coding branch is added on the basis of the ResBlock module to convert the input feature vector from the spatial domain to the frequency domain. By utilizing the characteristics of the fast Fourier transform, the global information of the image is encoded in the frequency component, which makes it easier for the network to capture the global difference between blurred images and clear images and improve the model performance.

[0041] Step 6: Fusion of meter category with regional mask information of the needle and range, intelligent matching of the current effective range of the meter, and judging whether there is an abnormal meter reading defect based on its intersection with the needle. Conventional meter reading recognition algorithms based on instance segmentation use the needle and dial scale as segmentation targets. While model recognition is difficult, it is necessary to configure corresponding templates for various types of dials to achieve reading matching. The present invention converts the more difficult reading matching task into a range interval detection task through step 5, solving the low recognition accuracy of conventional algorithms and dependence on manual configuration, and uses a lightweight YOLACT model framework to improve reasoning speed. The specific implementation steps are as follows: Step 6.1: Use the YOLACT instance segmentation model on the meter image to obtain the area information of the meter needle and the range. The YOLACT (You Only Look At CoefficienTs) model structure of the present invention is as follows: Figure 5 As shown in the figure, two parallel sub-branches, prototype mask generation and mask coefficient prediction, are added to the YOLOv8 model described in step 3.1 to realize the instance segmentation function. The specific details of the model are designed as follows: The prototype mask generation branch is implemented through a fully convolutional network (FCN), which receives the feature vector from the P3 layer after feature fusion and sets the number of output channels of the last convolution layer. Generate a corresponding number of prototype masks, which have the same spatial resolution as the input feature map. In the present invention, is 32.

[0042] The mask coefficients prediction head is parallel to the classification head and regression head. The convolution layer predicts the mask coefficients for each prototype mask. Each coefficient ranges from -1 to 1 and is used to represent the weight of the prototype mask during mask synthesis. Then, the mask coefficients and the prototype mask are linearly combined through formula x to obtain the instance mask: (9) in express dimensional prototype mask vector, express dimensional mask coefficient vector, Indicates the number of objects predicted after NMS, Represents the sigmoid operation.

[0043] Step 6.2: Use the logic post-processing strategy to optimize the mask boundary and filter out invalid instances in the needle and range, reduce the false positive rate of the recognition result, and improve the recognition accuracy. The post-processing strategies mainly include the following: The mask of small-area outlier instances is filtered by area threshold. In a specific embodiment of the present invention, let the image area be , set the needle area threshold to , the dial area threshold is .

[0044] The image morphological method is used to optimize the contour of the segmentation result. First, the mask boundary is expanded through the dilation operation, which can be used to fill small holes in the mask and connect broken edges. Then, the mask boundary is shrunk using the erosion operation to restore the overall shape and size of the mask.

[0045] Step 6.3: Combine the meter categories detected by the target detection model in step 3.1, and intelligently match the normal range intervals of each type of meter based on the existing prior knowledge, such as Figure 6 As shown, the matching normal range interval in this embodiment is set to 5% to 95% of the dial range, which is used to detect the situation where the meter reading is 0 or the limit value. By comparing the needle and its center extension line with the normal range interval, it is determined whether the two have an intersection. If no intersection is detected, the meter has an abnormal reading defect.

[0046] The above contents are merely examples and explanations of the technical solution of the present invention. The technicians in this technical field may make various modifications or additions to the specific embodiments described or replace them in a similar manner. As long as they do not deviate from the technical solution of the invention or exceed the scope defined by the claims, they should all fall within the protection scope of the present invention.

Claims

1. A general non-calibrated pointer meter reading abnormality identification method and system, the technical characteristics of which are that it includes the following steps: Step 1: Collect meter image samples, filter and sort them, annotate some data, and establish an initial sample library; Step 2: Use the sample library to train the meter’s object detection, image deblurring, and instance segmentation models; Step 3: Semi-automatically label the remaining data and iteratively update the model; Step 4: Use the target detection model to locate the meter position and category information, convert the image to the frequency domain and calculate the local power spectrum slope of the meter area to determine whether the meter dial is blurred; Step 5: For the image determined to be blurred, a deblurring algorithm is applied to obtain a high-definition meter image; Step 6: Integrate the meter category with the meter needle, range and other regional mask information, intelligently match the current meter effective range, and determine whether there is an abnormal meter reading defect based on its intersection with the meter needle.

2. According to the method and system for identifying abnormal readings of universal non-calibrated pointer meters according to claim 1, the technical feature is that the step 1 specifically include: Step 1.1: Use intelligent acquisition equipment to obtain image data with meter equipment as the main body; Step 1.2: Use the minimum bounding rectangle to mark the location and category information of the meter equipment and generate a Pasal VOC format annotation file; Step 1.3: Capture the meter device area image, use pixel-level masks to annotate the meter pointer area and meter range area, and generate a Pasal VOC format annotation file; Step 1.4: Use the manual focus mode of the intelligent acquisition device to collect the blurred image of the meter dial and the corresponding clear image of the meter dial; Step 1.5: Divide the labeled image samples into training set and test set in a ratio of 9:1 to build an initial sample library.

3. According to the method and system for identifying abnormal readings of pointer meters without calibration described in claim 2, its technical feature is that the meter equipment categories described in step 1.2 are divided into 4 categories according to universal characteristics, including conventional single-pointer meters, conventional double-pointer meters, multi-color range single-pointer meters and multi-color range double-pointer meters. The data used in steps 1.2-1.4 total 10,000 image samples and 2,000 of them are annotated to form a data set for training target detection, instance segmentation, and defuzzification models.

4. According to the method and system for identifying abnormal readings of universal non-calibrated pointer meters according to claim 1, the technical feature is that the step 2 specifically include: Step 2.1: Use the YOLO model to train the data set described in step 1.2 to obtain a basic target detection model that can locate the meter category and position; Step 2.2: Use the YOLACT model to train the data set described in step 1.3 to obtain a basic instance segmentation model that can locate the meter range and pointer area; Step 2.3: Use the DeepRFT model to train the data set described in step 1.4 to obtain a deblurring model that can improve image quality.

5. According to a general uncalibrated pointer meter reading abnormality identification method and system according to claim 1, its technical feature is that the YOLO model described in step 2 adopts the YOLOv8 model, the initial learning rate is set to 0.005, and the optimizer is AdamW; the YOLACT model adopts the YOLACT model, the initial learning rate is set to 0.001, and the optimizer is AdamW; the DeepRFT model, the initial learning rate is set to 0.00005, and the AdamW optimizer is used.

6. According to a general uncalibrated pointer meter reading abnormality identification method and system as described in claim 1, its technical feature is that the semi-automatic labeling described in step 3 uses the basic target detection and instance segmentation model described in step 2 to infer the unlabeled image samples, and obtains high-quality labeled data through manual review.

7. The method and system for identifying abnormal readings of universal non-calibrated pointer meters according to claim 1, wherein the technical feature is that step 4 specifically include: Step 4.1: Use the object detection model to identify the meter area in the image; Step 4.2: Calculate the power spectrum slopes of the global image and the meter area image in the frequency domain using discrete Fourier transform; Step 4.3: Calculate the local power spectrum slope of the meter area image relative to the global image to determine whether the current meter has a meter dial blur defect.

8. A general method and system for identifying abnormal readings of pointer-type meters without calibration according to claim 7, wherein the power spectrum slope calculation process described in step 4.2 is as follows: (1) in It means that the size after Fourier transform to frequency domain is Pictures of and Represents the two-dimensional frequency of the image in polar coordinates, by Directional power spectrum The summation yields the following approximate representation: (2) in is the amplitude proportional factor in each direction. By transforming equation (2), the frequency domain component, i.e., the power spectrum slope, can be calculated. : (3)。 9. A general method and system for identifying abnormal readings of pointer-type meters without calibration according to claim 7, wherein the local power spectrum slope calculation process described in step 4.3 is as follows: (4) in Indicates the frequency spectrum slope of the current meter area, represents the frequency spectrum slope of the global image. Both are calculated by formula (3). The ratio of the two is It is the slope of the local power spectrum. If it is greater than the threshold 1, it is considered that there is a dial blur defect in this meter area.

10. The method and system for identifying abnormal readings of universal non-calibrated pointer meters according to claim 1, wherein the technical feature is that step 6 specifically include: Step 6.1: Use the instance segmentation model on the meter image to obtain the area information of the meter needle and range; Step 6.2: Filter and detect invalid instances in the needle and range according to the post-processing strategy; Step 6.3: Integrate the meter category detected in step 4.1, intelligently match the normal range of the current meter, and determine whether the meter needle has an intersection with it. If no intersection is detected, the current meter has an abnormal meter reading defect.

Citation Information

Cited By

  • Image segmentation and target detection integrated method for cross-country road

    CN120635530A

  • Multi-modal large model body perception method and system based on double-path ring

    CN121505241A