Automatic identification method for reading of general meter

Abstract

The invention provides a general meter reading automatic identification method, and relates to the technical field of computer vision and deep learning, and the method comprises the steps: carrying out the image collection of a meter to obtain an original image containing an instrument panel region of the meter, and carrying out the image recognition of the original image, obtaining an integer area and a decimal area in the instrument panel area and corresponding coordinates thereof, and determining physical layout positions of the integer area and the decimal area in the instrument panel area according to the coordinates; segmenting a plurality of numeric character regions from the integer region; and performing number identification on each number character area and decimal area, and sequentially splicing a plurality of numbers obtained by identification according to corresponding physical layout positions to obtain the meter reading. The method has the beneficial effects that a multi-stage cascade detection method is adopted, interference is filtered layer by layer (positioning-segmentation-classification digital recognition) through cooperative work of multiple processes, and the problem that cross-specification universal recognition is difficult to realize by a traditional single detection model is solved.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and deep learning technology, and in particular to an automatic identification method for readings of general measuring instruments. Background Technology

[0002] With the development of smart cities and IoT technologies, the demand for automated reading of water, electricity, and gas meters is becoming increasingly urgent. Traditional manual meter reading methods are inefficient and costly, while existing computer vision-based automatic identification technologies face the challenge of poor compatibility with various meter specifications in practical applications: the interface layouts of different meter models vary significantly, the roller digit arrangement of mechanical counters is diverse, and existing single detection models are difficult to achieve universal identification across specifications.

[0003] Therefore, there is an urgent need for a general-purpose detection method that combines high precision, strong robustness, and low computational cost to meet the actual needs in complex industrial environments. Summary of the Invention

[0004] To address the problems existing in the prior art, the present invention provides an automatic identification method for readings of general measuring instruments, comprising:

[0005] Step S1: Acquire an image of the measuring instrument to obtain an original image of the instrument panel area, and perform image recognition on the original image to obtain the integer area and decimal area in the instrument panel area and their corresponding coordinates. Determine the physical layout position of the integer area and the decimal area in the instrument panel area based on the coordinates.

[0006] Step S2: Divide the integer area into multiple numeric character areas;

[0007] Step S3: Perform digit recognition on each of the digit character areas and the decimal area, and then sequentially splice the recognized digits according to their corresponding physical layout positions to obtain the meter reading.

[0008] Preferably, step S1 further includes an image distortion correction process, comprising:

[0009] Step A1: Before identifying the dashboard area in the original image, perform global distortion correction on the original image;

[0010] And / or step A2, after identifying the integer and decimal regions in the dashboard area of ​​the original image, correcting the original image using a perspective transformation matrix.

[0011] Preferably, in step S1, the original image is subjected to image recognition using a dashboard positioning model. This dashboard positioning model is an improvement upon the YOLO model, and its improvement process includes:

[0012] Step B1 introduces an anchorless mechanism into the initial model of the YOLOv8n architecture, using each pixel on the output feature map of the initial model as a candidate detection point, and marking the distance from the candidate detection point to the boundary of the target box to construct the target box, thereby realizing the target box selection of the dashboard area in the original image.

[0013] Step B2: The detection head of the initial model adopts a classification and regression decoupled structure so that the confidence branch and the target box regression branch in the initial model are processed by independent sub-networks respectively.

[0014] Step B3: Introduce the DFL loss function into the loss function architecture of the initial model to realize a distributed prediction mechanism for each boundary point position during the box regression process, outputting multiple discrete probability distribution points to approximate the continuous regression target.

[0015] Step B4: Replace the box-frame regression function in the loss function architecture of the initial model with the EIoU loss function, and introduce a layout consistency loss term.

[0016] Preferably, when a new original image of a new meter that has not been input before is input into the dashboard positioning model, a new sample set is constructed based on multiple new original images of the same new meter. Then, the dashboard positioning model is subjected to lightweight incremental training based on the new sample set to achieve fine-tuning of the model parameters.

[0017] Preferably, in step S1, the dashboard positioning model is used to perform image recognition on the original image, and the dashboard positioning model is trained using a hybrid dataset, which includes real images captured and synthetic images rendered and synthesized.

[0018] The process for generating the synthesized image includes:

[0019] Create a 3D geometric model of the dashboard, then set the material and anti-texture of the 3D geometric model, set the background of the 3D geometric model, and set the light source to illuminate the 3D geometric model;

[0020] The camera parameters of the camera are randomly set, and attitude perturbation and distortion simulation are added to the camera;

[0021] The camera samples the 3D geometric model image from multiple angles, saves each frame of the sampled image during the image sampling process, and performs image correction and perspective perturbation on the sampled images to obtain an enhanced image. Then, all sampled images and the enhanced image are saved.

[0022] The two-dimensional bounding boxes of the integer and decimal regions in the sampled image and the enhanced image are labeled to obtain corresponding YOLO format labels. The labeled sampled image and the enhanced image are then used as the composite image.

[0023] Preferably, in step S2, a segmentation network model is used to segment multiple numeric character regions from the integer region, and the segmentation network model is the YOLOv8n-Seg segmentation model;

[0024] The backbone network of the YOLOv8n-Seg segmentation model uses CSPDarknet-Lite based on the C2f module.

[0025] The image size input to the segmentation network model is fixed at 640*640;

[0026] An FPN+PAN path is introduced into the neck structure of the YOLOv8n-Seg segmentation model;

[0027] The head structure of the YOLOv8n-Seg segmentation model adopts a three-branch output structure, namely class confidence, location coordinates and mask prototype features, and the size of the mask segmentation result is [B,N,32,32], where B is the target box and N is the number of detected targets;

[0028] The loss function of the YOLOv8n-Seg segmentation model is composed of cross-entropy loss, enhanced IOU loss, mask quality optimization loss, and adversarial robustness loss term.

[0029] Preferably, the segmentation network model is trained using an image augmentation dataset, and the process of constructing the image augmentation dataset includes:

[0030] Multiple original images of measuring instruments are acquired. During the training phase, first-order gradient perturbation is added to the original images to enhance adversarial noise perturbation, and / or simulated reflection interference to generate spot mask for pseudo-reflection mask synthesis, and / or image structure boundaries are enhanced, and the distinction between characters and background is improved for color normalization and contrast enhancement. Finally, the data-enhanced images are added to the image enhancement dataset.

[0031] Preferably, in step S3, a digit recognition model is used to perform digit recognition on each of the digit character areas and the decimal area. The digit recognition model is obtained by pruning and reconstructing the backbone structure of the YOLOv8 model. The pruning and reconstruction process includes:

[0032] The C2F structure and shallow residual connections in the backbone of the YOLOv8 model are retained.

[0033] Remove the path aggregation structure from the YOLOv8 model and remove the detection head, keeping only the classification head;

[0034] Add a global average pooling layer to the end of the pruned backbone structure of the YOLOv8 model;

[0035] Reconstruct the classification header;

[0036] A lightweight attention module is embedded in the backbone structure of the trimmed YOLOv8 model.

[0037] Preferably, the output result of performing digit recognition on the images in each of the digit character areas and the decimal area in step S3 is as follows:

[0038] A probability vector of length 10 is generated, which includes each possible digit and its corresponding probability value. Then, the category with the highest probability value is selected from the probability vector as the final recognition result.

[0039] If any of the probability values ​​is less than a preset confidence threshold, the system either rejects the data or proceeds to the Top-K post-processing logic, which includes outputting K numbers sorted by probability and their confidence sets.

[0040] Preferably, the image of each of the digit character areas has a corresponding digit frame, and the meter reading includes integer area reading and decimal area reading. In step S3, the splicing process of the integer area reading includes:

[0041] The number frames of multiple numbers identified from the images of each of the number character areas are sorted from left to right according to the horizontal position of the corresponding number character area in the physical layout position of the integer area;

[0042] The sorted number boxes are adjusted for consistent row height and width ratio to perform physical constraint filtering.

[0043] The integer region reading is obtained by merging and concatenating the various number frames after physical constraints into a number string.

[0044] The above technical solution has the following advantages or beneficial effects:

[0045] To address the problem of poor compatibility among various meter specifications in the background technology, this invention employs a multi-stage cascaded detection method. Through the coordinated work of meter panel positioning, image segmentation, and digital recognition, it achieves layer-by-layer interference filtering (positioning—segmentation—classification), reducing the false detection rate by 42% compared to single-stage methods. It also demonstrates significant adaptability to differences in interface layout among different meter models. This makes the invention widely applicable to various mechanical roller counters such as water meters, electricity meters, and gas meters, solving the problem of traditional single-detection methods' inability to achieve universal identification across specifications. Attached Figure Description

[0046] Figure 1 A flowchart illustrating an automatic identification method for readings of a general measuring instrument is provided in a preferred embodiment of the present invention.

[0047] Figure 2 A flowchart illustrating the image distortion correction process in a preferred embodiment of the present invention;

[0048] Figure 3 This is a schematic diagram of the improved process of the dashboard positioning model in a preferred embodiment of the present invention. Detailed Implementation

[0049] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. The present invention is not limited to this embodiment; other embodiments that conform to the spirit of the present invention may also fall within the scope of the present invention.

[0050] In a preferred embodiment of the present invention, based on the above-mentioned problems existing in the prior art, an automatic identification method for readings of general measuring instruments is provided, such as... Figure 1 As shown, it includes:

[0051] Step S1: Acquire images of the measuring instruments to obtain the original image of the instrument panel area containing the measuring instruments, and perform image recognition on the original image to obtain the integer area and decimal area in the instrument panel area and their corresponding coordinates. Determine the physical layout position of the integer area and decimal area in the instrument panel area based on the coordinates.

[0052] Step S2: Divide the integer area into multiple numeric character areas;

[0053] Step S3: Perform digit recognition on each digit character area and decimal area, and then sequentially splice the recognized digits according to their corresponding physical layout positions to obtain the meter reading.

[0054] Specifically, addressing the problem of poor compatibility among various instrument specifications in the background technology, this invention employs a multi-stage cascaded detection method. Preferably, it can achieve the coordinated operation of the three processes of instrument panel positioning, image segmentation, and digital recognition through an instrument panel positioning model, a segmentation network model, and a digital recognition model. This enables layer-by-layer filtering of interference in the three stages of (positioning-segmentation-classification), reducing the false detection rate by 42% compared to a single-stage model. Each stage model is independently trained and updated, supporting rapid iteration with small samples (single model training samples <5k), and achieving significant adaptation to differences in the interface layout of different instrument models.

[0055] The application of this series of technical means enables the present invention to be widely applicable to various mechanical roller counters such as water meters, electricity meters, and gas meters, solving the problem that traditional single detection models are difficult to universally identify across specifications.

[0056] In a preferred embodiment of the present invention, step S1 further includes an image distortion correction process, such as... Figure 2 As shown, it includes:

[0057] Step A1: Before identifying the dashboard area in the original image, perform global distortion correction on the original image;

[0058] And / or step A2, after identifying the integer and decimal regions in the dashboard area of ​​the original image, corrects the original image using a perspective transformation matrix.

[0059] Specifically, in this embodiment, an image distortion correction algorithm can be optionally built into the instrument panel positioning model to eliminate short focal length lens distortion through affine transformation; it is mainly used to correct radial distortion (barrel distortion of short focal length lens) and tangential distortion (caused by installation tilt);

[0060] The core method involves pre-calculating camera parameters using Zhang's Camera Calibration method, and then performing real-time correction using OpenCV's undistort() function (cv2.undistort() is the core function in OpenCV used to correct lens distortion; its function is to reverse-correct image distortion caused by the optical characteristics of the camera lens, restoring it to an imaging effect close to that of an ideal pinhole camera model).

[0061] The processing steps can be divided into two parts:

[0062] Step 1 (Preprocessing; optional): Perform global distortion correction on the original input image before locating the dashboard area;

[0063] Step 2 (Post-processing; optional): After obtaining the dashboard ROI (Region of Interest, a core concept in computer vision used to focus on key areas and reduce computational load), a perspective transformation matrix is ​​used for correction in extreme deformation scenarios (such as refraction from curved protective covers).

[0064] In a preferred embodiment of the present invention, in step S1, an instrument panel positioning model is used to perform image recognition on the original image. The instrument panel positioning model is an improvement on the YOLO model, such as... Figure 3 As shown, the improvement process includes:

[0065] Step B1 introduces an anchorless mechanism into the initial model of the YOLOv8n architecture, using each pixel on the output feature map of the initial model as a candidate detection point, and marking the distance from the candidate detection point to the boundary of the target box to construct the target box, thereby achieving target box selection for the dashboard area in the original image.

[0066] Step B2: The detection head of the initial model adopts a classification and regression decoupled structure so that the confidence branch and the target box regression branch in the initial model are processed by independent sub-networks respectively.

[0067] Step B3 introduces the DFL loss function into the loss function architecture of the initial model to realize a distributed prediction mechanism for each boundary point position during the box regression process, outputting multiple discrete probability distribution points to approximate the continuous regression target.

[0068] Step B4: Replace the box-frame regression function in the loss function architecture of the initial model with the EIoU loss function, and introduce a layout consistency loss term.

[0069] Specifically, in the visual inspection of unattended metering equipment, the YOLO (You Only Look Once) series of single-stage inspection models has become the mainstream solution due to its end-to-end and real-time characteristics. Since its initial introduction in 2015, YOLO has continuously improved its detection accuracy while maintaining high inference speed through continuous iteration (v1→v2→v3→v4→v5→v6→v7→v8). YOLOv8, released by Ultralytics in January 2023, introduced an anchorless, separate head, an advanced backbone and neck architecture, and an optimized accuracy-speed balance strategy, enabling it to achieve industry-leading performance in multiple tasks (detection, segmentation, classification, pose estimation, etc.).

[0070] More specifically, step S1 in this embodiment mainly extracts the instrument panel area from the input meter image based on the instrument panel positioning model, including roller mechanical dials and LCD dials. The instrument panel area positioning module utilizes a lightweight YOLOv8n model, combined with EIoU Loss and layout consistency constraints, to effectively improve the positioning accuracy of small dial areas and support adaptive layout of multiple instrument sizes.

[0071] The dashboard positioning model is based on an improved version of the YOLOv8n architecture, and the specific improvements include:

[0072] 1. Improved YOLOv8n architecture

[0073] To improve the robustness and versatility of target localization in multi-specification instrument images, a YOLOv8n architecture is introduced at the model level with key structural improvements. These improvements are based on the YOLOv8 detection head proposed by Ultralytics, optimizing issues such as complex anchor management, large localization errors, and the tendency to miss small targets during target detection.

[0074] (1) Anchor-Free Mechanism

[0075] Traditional YOLOv1-v5 series models employ an anchor-based detection mechanism, which predefines several anchors on each output feature map unit. Each anchor represents a set of fixed aspect ratios (w / m). a ,h a Then regress its offset:

[0076]

[0077] Where (x) a ,y a ) represents the anchor center point; Δ represents the offset prediction output by the network; Δx, Δy, are the offsets, and Δw, Δh are the size scaling factors; The coordinates of the center of the prediction box; This is the predicted frame size.

[0078] However, the Anchor-Based model has the following drawbacks:

[0079] It is sensitive to anchor settings and requires K-Means clustering to obtain the set of anchors that fit the target.

[0080] Performance degradation when detecting non-standard scale targets (such as long rollers on a mechanical dial).

[0081] Introducing multiple anchors increases the number of regression targets, leading to gradient competition issues.

[0082] To address this, an Anchor-Free strategy is introduced, directly using each pixel position on the output feature map as a candidate detection point. The network outputs four offsets (l, t, r, b) representing the distances from that pixel to the left, top, right, and bottom boundaries of the target box.

[0083]

[0084] Thus constructing the target box This method eliminates the need for preset anchors, avoids the tedious hyperparameter tuning process, and is more adaptable to targets of any size and shape, making it particularly suitable for detection tasks involving small digital panes and irregularly shaped dials.

[0085] (2) Decoupled Head Design

[0086] The YOLOv8 detection head employs a decoupled classification and regression structure, meaning that the confidence score (class probability) and the bounding box regression branch are processed using independent sub-networks, as shown in the following structure:

[0087] Head ds =Conv ds (FPN out ), Head reg =Conv reg (FPN out )

[0088] Among them FPN out This represents the fused feature value from the feature pyramid.

[0089] This design has the following advantages:

[0090] To avoid gradient interference in regression and classification tasks and improve their respective prediction accuracy.

[0091] Improving the network's ability to distinguish between different types of targets, especially in multi-scale dashboards, facilitates the localization of multiple targets within the same frame.

[0092] The decoupled structure supports flexible expansion, such as adding auxiliary regression terms (IoU, confidence scores) or multi-task outputs (such as rotating boxes, masks, etc.).

[0093] (3) Detection frame quality estimation (Distribution Focal Loss)

[0094] To further improve the accuracy of bounding box localization, YOLOv8n introduces the Distribution Focal Loss (DFL) loss function. During the box regression process, a distributed prediction mechanism is used for the location of each boundary point, outputting multiple discrete probability distribution points p. i To approximate the continuous regression objective:

[0095] L DFL =∑ i＝1 n Softmax(p i SmoothL1(x) i -x*)

[0096] Where x* is the true boundary value, x i For discrete boundary points, Softmax ensures a smooth transition in the distribution.

[0097] This method makes the model smoother and more stable when fine-tuning the target boundary, effectively alleviating numerical jitter and prediction jitter problems, and is suitable for fine-level localization such as digit pane detection.

[0098] 2. Loss Function Optimization

[0099] (1) Optional, use EIoU Loss instead of CIOU Loss

[0100] The default loss architecture of YOLOv8n uses a three-component weighted loss, and its overall training loss is:

[0101] L=λ box L CIoU (Box-frame regression) + λ dfl L DFL (Boundary distribution decoding loss) + λ cls L BCE (Multi-class cross-entropy loss)

[0102] To improve the positioning accuracy of small targets, the EloU loss function is adopted, taking into account the center point distance, width and height differences, and overlapping areas.

[0103] The EloU loss function is defined as:

[0104] L EloU =1-IOU+ρ 2 (b,b gt ) / c 2 +(ww gt ) 2 / w c 2 +(hh gt ) 2 / h c 2

[0105] Where ρ represents the Euclidean distance between the center points, c is the diagonal length, and w and h are the width and height of the prediction box, respectively. gt h gt w represents the width and height of the actual bounding box. c h c The width and height of the smallest closure box that encloses the predicted box and the ground truth box.

[0106] (2) Add layout consistency constraints

[0107] To ensure a reasonable layout for both whole and fractional regions, a layout consistency loss term is introduced to penalize prediction results that do not conform to the expected layout.

[0108] Let the center points of the integer region and the decimal region be (x1, y1) and (x2, y2) respectively, and define the layout consistency loss as:

[0109]

[0110] in, The indicator function is λ, which is the loss term when the positional relationship between the integer and decimal regions does not conform to expectations, and 0 otherwise.

[0111] (3) Optionally, add multi-scale and sample weights.

[0112] In the YOLOv8 decoupling head, the outputs at different scales are weighted:

[0113] L box =∑ s∈{P3,P4,P5} α s L (s) EloU ,∑ s α s =1

[0114] The smaller target layer (P3) can be assigned higher weights α3>α4, α5

[0115] In the YOLOv8n model and its improvements, P3, P4, and P5 typically refer to feature maps at different levels in the Feature Pyramid Network (FPN). These maps are used to detect targets of different sizes. In loss function optimization, by setting different weight coefficients for feature maps of different scales, the model's detection performance on targets of different scales can be adjusted.

[0116] In summary, the optimized overall training loss is:

[0117] L=λ box L box +λ dfl L DFL +λ cls L BCE +λ layout L layout

[0118] (4) Optionally, adjust the λ weights to optimize the model's performance in locating scenes in the dashboard area. For example...

[0119] λ box =3.0; λ dfl =1.0; λ cls =1.0; λ layout =0.5

[0120] In a preferred embodiment of the present invention, when a new original image of a new meter that has not been input before is input into the dashboard positioning model, a new sample set is constructed based on multiple new original images of the same new meter. Then, the dashboard positioning model is subjected to lightweight incremental training based on the new sample set to achieve fine-tuning of the model parameters.

[0121] Specifically, in this embodiment, the dashboard positioning model supports rapid fine-tuning using a small number of samples to adapt to a new dashboard layout when a new layout is detected. Further improvements to the dashboard positioning model include adaptive deployment.

[0122] The dashboard area positioning module supports online fine-tuning. During the deployment phase, when a new dashboard layout is detected, it supports rapid fine-tuning using a small number of samples to adapt to the new layout.

[0123] Let the original model parameters be θ, the new sample set be Dnew, and the fine-tuning objective be to minimize the loss function:

[0124] θ'=argmin θ ∑ (x,y)∈Dnew L(f(x;θ),y)

[0125] Where f(x; θ) is the predicted output of the model, L is the loss function, and θ' is the fine-tuned model parameters.

[0126] In a preferred embodiment of the present invention, a hybrid dataset is used to train the dashboard positioning model, the hybrid dataset including real images captured and synthetic images rendered and synthesized.

[0127] The process of generating a composite image includes:

[0128] Create a 3D geometric model of the dashboard, then set the materials and anti-textures of the 3D geometric model, set the background of the 3D geometric model, and set the light source to illuminate the 3D geometric model;

[0129] Randomly set the camera parameters and add attitude perturbations and distortion simulations to the camera;

[0130] The 3D geometric model image is sampled from multiple angles by a camera. Each frame of the sampled image is saved during the image sampling process. The sampled images are then image-corrected and perspective-perturbed to obtain enhanced images. Finally, all sampled images and enhanced images are saved.

[0131] The two-dimensional bounding boxes of integer and decimal regions in the sampled and enhanced images are labeled to obtain the corresponding YOLO format labels. The labeled sampled and enhanced images are then used as the composite image.

[0132] Specifically, the improved dashboard positioning model still needs to be trained, and a hybrid dataset is used for model training during the training process.

[0133] The process of building a hybrid dataset includes:

[0134] 1. The hybrid dataset consists of two parts: real data and synthetic data.

[0135] Authentic data: High-resolution (no less than 800x600) images of multiple brands and models (no fewer than 15 mainstream brands and models) of mechanical water meters, electricity meters, and gas meters (including common roller-type mechanical dials and a small number of LCD dials) taken on-site. Images were captured against different backgrounds (indoor and outdoor), at different installation angles and distances to capture perspective effects caused by installation position deviations and uneven shooting surfaces.

[0136] Synthetic data: Automatically generated based on 3D models and simulation environments using rendering engines such as Blender / BlenderProc, covering more rare layouts, different lighting, reflections and distortions.

[0137] The two sets of data are mixed proportionally (e.g., 70% real + 30% synthetic) to ensure that the model can learn details from the real world while maintaining stable performance in extreme or rare scenarios.

[0138] 2. Apply augmentation strategies to the data in the mixed dataset (before labeling):

[0139] Layout disturbance: Randomly scale / translate the dial area to simulate installation position deviation.

[0140] Adversarial texturing: Add dial-like texture noise to non-digital areas to prevent the model from overfitting to specific backgrounds.

[0141] 3. Label the data in the mixed dataset, and the labeling specifications are as follows:

[0142] All samples adopt the YOLO format (class, x_center, y_center, width, height), and the center of the dial is used as the origin of the normalized coordinates to ensure consistent layout across specifications (avoiding absolute coordinate dependency).

[0143] Label the integer group area (integer digit scroll wheel cluster) and the decimal area (decimal digit scroll wheel cluster) to ensure accurate input for downstream segmentation and classification modules.

[0144] Existing technologies suffer from insufficient robustness to complex imaging interference: due to factors such as distortion from short focal length lenses, reflection and refraction from transparent curved protective covers, and environmental water mist contamination, digital areas in images are prone to deformation, blurring, or partial occlusion, leading to an increased false detection rate in traditional OCR technology.

[0145] To address the issue of insufficient robustness to complex imaging interference, this invention innovates at the data-level strategy level. Through hybrid dataset construction and enhancement strategies, the stability of the model under complex imaging conditions is significantly improved. The hybrid dataset combines real and synthetic data, covering multiple brands and models of mechanical water meters, electricity meters, and gas meters, ensuring that the model can learn details from the real world while maintaining stable performance in extreme or rare scenarios. Simultaneously, enhancement strategies, including layout perturbation and adversarial texture addition, simulate installation position deviations and background interference, effectively preventing the model from overfitting to specific backgrounds. The application of these data-level strategies enables this invention to effectively cope with complex imaging interference such as short-focal-length lens distortion, reflections and refractions from transparent curved protective covers, and environmental water mist contamination, significantly reducing the false detection rate.

[0146] More specifically, the process of generating synthetic data includes:

[0147] 1. 3D Modeling and Scene Construction

[0148] (1) Dashboard and component modeling

[0149] Using CAD tools or public resource libraries (such as ShapeNet, Sketchfab), construct a 3D geometric model of a typical instrument structure, including:

[0150] M = {Dial 3D Wheel 3D Cover 3D Base 3D}

[0151] Each object is a mesh g = (V, E), where V is the set of vertices and E is the set of edges.

[0152] (2) Material and Confrontation Texture Design

[0153] Apply physically true PBR material functions to each geometric surface Si:

[0154] R i =f r (ω i ,ω o ,n;θ)

[0155] Among them, f r It is a BRDF function, ω i ω is the direction of the incident light. o Here, n is the direction of the emitted light, θ is the surface normal, and θ is the material parameter (roughness, metallicity, reflectivity, etc.).

[0156] To prevent the model from overfitting to the background region, in the non-digit region Ω bg Add adversarial texture noise:

[0157]

[0158] in The noise is Gaussian, and α controls the disturbance intensity.

[0159] (3) Background and lighting construction

[0160] Background image B is randomly sampled from industrial or residential scene sets: B ~ μ(D) bg )

[0161] The lighting L is a combination of multiple light sources, including directional light, point light, and ambient light, simulating natural light and indoor light sources:

[0162] I total =∑ k＝1 K (I k ·max(0,n·l k ))

[0163] Among them, I k (uppercase i) represents the direction of the k-th light source, l k (Lowercase L) represents its intensity, and K represents the number of light sources.

[0164] 2. Camera and Rendering Randomization

[0165] (1) Randomization of camera intrinsic parameters

[0166] The camera model follows the pinhole projection model:

[0167] x 2D ~K·[R|t]·X 3D

[0168] in:

[0169] K = [f x 0c x ;0f y c y [;001] is the intrinsic parameter matrix, and the focal length f x ,f y ~μ(20,50)mm

[0170] [R|t] is the rotation and translation matrix, simulating the camera pose.

[0171] X3D is the coordinate system for the 3D world.

[0172] (2) Attitude perturbation and distortion simulation

[0173] Camera extrinsic parameters (attitude) introduce perturbation terms:

[0174] R'=R·R δ ,t'=t+Δt

[0175] in,

[0176] R δ =R x (δ x )R y (δ y )R z (δ z )

[0177] δ x ,δ y ,δ z ~(0,σ θ 2 ) indicates a tiny rotation

[0178] Add barrel and tangential distortion simulation:

[0179] x dist =x(1+k1r) 2 +k2r 4 +k3r 6 )+2p1xy+p2(r 2 +2x 2 )

[0180] y dist =y(1+k1r 2 +k2r 4 +k3r 6 )+2p2xy+p1(r 2 +2y 2 )

[0181] in,

[0182] k1, k2, k3 are radial distortion coefficients

[0183] p1 and p2 are tangential distortion coefficients.

[0184] r 2 =x 2 +y 2

[0185] (3) Image enhancement and dual-version sampling

[0186] Two versions are generated for each frame of the image:

[0187] Iraw: Uncorrected original image

[0188] Iundist: Image after OpenCV cv2.undistort() correction

[0189] Optionally, apply perspective perturbation.

[0190] x'=H·x

[0191] Used to simulate projection offset during the shooting process.

[0192] 3. Automatic Label Generation

[0193] (1) Built-in annotation generation

[0194] Using Blender's rendering engine and object coordinate transformations, output integer region B. int With decimal zone B frac Two-dimensional bounding box:

[0195] B={(x min ,y min ,x max ,y max )}

[0196] After coordinate normalization, convert to YOLO format labels:

[0197] YOLO = (c, (x) min +x max ) / 2W,(y min +y max ) / 2H,(x max -x min ) / W,(y max -y min ) / H)

[0198] Where W and H are the image width and height, and c is the category number.

[0199] (2) Multi-view enhancement

[0200] For the same dial model, perform n camera view samplings R i ,t i+1 n Obtain a multi-angle image set:

[0201]

[0202] Used to enhance the robustness of the model under different viewing angles.

[0203] In a preferred embodiment of the present invention, the segmentation network model in step S2 adopts the YOLOv8n-Seg segmentation model;

[0204] The backbone network of the YOLOv8n-Seg segmentation model uses CSPDarknet-Lite based on the C2f module.

[0205] The image size input to the segmentation network model is fixed at 640*640;

[0206] An FPN+PAN path is introduced into the neck structure of the YOLOv8n-Seg segmentation model;

[0207] The YOLOv8n-Seg segmentation model adopts a three-branch output structure, which consists of class confidence, location coordinates, and mask prototype features. The size of the mask segmentation result is [B,N,32,32], where B is the bounding box and N is the number of detected targets.

[0208] The loss function of the YOLOv8n-Seg segmentation model is composed of cross-entropy loss, enhanced IOU loss, mask quality optimization loss, and adversarial robustness loss term.

[0209] Specifically, in this embodiment, the digital segmentation and enhancement part of step S2 adopts the YOLOv8n-Seg segmentation model, enhances the context awareness capability through the FPN+PAN structure, and introduces adversarial noise perturbation, pseudo-reflection mask synthesis and color normalization and contrast enhancement strategies to further improve the segmentation robustness.

[0210] More specifically, the segmentation network model is used to accurately segment each numeric character region from the integer region ROI of the dashboard area output in step S1. To address common interference problems in dashboard images such as reflection, occlusion, and blurring, a lightweight segmentation network structure and a multi-level enhancement mechanism are proposed to ensure pixel-level localization accuracy and robustness.

[0211] More specifically, the segmentation network model adopts the YOLOv8n-Seg segmentation model as the backbone network, with a fixed input image size of 640×640, balancing model performance and edge deployment efficiency. The improvement process includes:

[0212] To improve the accuracy and stability of character encoding segmentation, the YOLOv8n-Seg model structure and loss design are as follows:

[0213] Backbone: CSPDarknet-Lite based on the CZf module, extracting multi-scale feature maps;

[0214] Neck: Introduces FPN+PAN path to enhance context awareness;

[0215] Head: Outputs three branches: class confidence p, location coordinates (x, y, w, h), and mask prototype features.

[0216] Segmentation Head output size: [B, N, 32, 32], where N is the number of detected targets. The masking result is obtained through prototype decoding.

[0217]

[0218] Where α i,k Let P be the mask coefficient. k For the k-th prototype;

[0219] The segmentation network model uses a joint loss function (Segmentation Loss):

[0220] The loss from task segmentation consists of four parts:

[0221] L seg =λ ds ·L BCE +λ box ·L CloU +λ mask ·L Dice +λ enh ·L adv

[0222] Wherein, LBCE: Cross Entropy Loss, is used to optimize category classification;

[0223] LCloU: Enhanced IOU loss (can be replaced with EIOU) optimizes position regression;

[0224] LDice: Used for pixel mask quality optimization, defined as:

[0225]

[0226] Ladv: An adversarial robust loss term that promotes consistent model performance under perturbed images.

[0227] In a preferred embodiment of the present invention, the segmentation network model is trained using an image augmentation dataset. The process of constructing the image augmentation dataset includes:

[0228] Multiple original images of measuring instruments are acquired. During the training phase, first-order gradient perturbation is added to the original images to enhance adversarial noise perturbation, and / or simulated reflection interference to generate spot mask for pseudo-reflection mask synthesis, and / or image structure boundaries are enhanced, and the distinction between characters and background is improved for color normalization and contrast enhancement. Finally, the data-enhanced images are added to the image enhancement dataset.

[0229] Specifically, in this embodiment, adversarial enhancement and interference suppression are introduced into the training data of the segmentation network model during the training process to enhance the model's adaptability to real-world environmental disturbances such as specular reflections and changes in shooting angles. To address complex lighting changes, specular reflections, and slight occlusion issues in real-world scenes, the following data augmentation schemes are proposed to improve segmentation robustness from the training level, including:

[0230] 1. Adversarial Noise Injection

[0231] During the training phase, a first-order gradient perturbation is added to the original image x to achieve adversarial training:

[0232]

[0233] Where ε controls the disturbance amplitude (e.g., ε = 4 / 255);

[0234] adversarial example x adv The model is trained with the same weights as the original image to improve its insensitivity to local perturbations.

[0235] 2. Synthetic Reflection Masks

[0236] To simulate common reflection interference in real-world photography, a light spot mask of the following form is generated:

[0237] I bland =(1-α)·I+α·R,α∈[0.2,0.4]

[0238] R is a specular reflection image generated by Gaussian blur, which is synthesized using strip, circular, and spot distributions;

[0239] α controls the fusion intensity, giving the training images a diverse reflective distribution.

[0240] 3. Color Normalization and Contrast Enhancement (CLAHE + HSV Normalization)

[0241] Enhance image structure boundaries to improve the distinction between characters and background:

[0242] CLAHE (Contrast Limited Adaptive Histogram Equalization) enhances the local contrast of grayscale or luminance channels.

[0243] HSV Normalize converts the image from RGB to HSV space, performing normalization and gamma adjustment on the V channel:

[0244] V'=((V-μ V ) / σ V )·γ,clip to[0,255]

[0245] Where μV and σV are the mean and standard deviation of brightness, and γ∈[0.8,1.2] controls the brightness stretching amplitude.

[0246] In a preferred embodiment of the present invention, in step S3, a digit recognition model is used to perform digit recognition on each digit character area and the decimal area. The digit recognition model is obtained by trimming and reconstructing the main structure of the YOLOv8 model. The trimming and reconstruction process includes:

[0247] The C2F structure and shallow residual connections in the backbone of the YOLOv8 model are retained.

[0248] Remove the path aggregation structure from the YOLOv8 model and remove the detection head, keeping only the classification head;

[0249] Add a global average pooling layer to the end of the pruned backbone structure of the YOLOv8 model;

[0250] Reconstruct the category header;

[0251] A lightweight attention module is embedded in the backbone structure of the trimmed YOLOv8 model.

[0252] Specifically, in this embodiment, the digit recognition part enhances the ability to extract fuzzy digit features by using a customized YOLOv8 classification network, combined with backbone pruning, global average pooling layers and attention mechanisms, and achieves high-precision 0-9 digit classification.

[0253] More specifically, step S3 is used to perform digit recognition on the digit character area segmented in step S2. The recognition process is based on a digit recognition model, which performs robust digit classification and recognition for low-resolution digit images (such as mechanical instruments and dial screenshots).

[0254] The core of the digit recognition model is a customized YOLOv8 backbone network, incorporating a lightweight attention mechanism to enhance fuzzy digit recognition capabilities. The model takes a 64×64 image as input and outputs the digit category c. i∈[0,1,...,9] and its confidence level conf. The modifications include:

[0255] 1. Backbone Pruning and Adaptation

[0256] Based on the YOLOv8 backbone, a lightweight convolutional classification network is designed.

[0257] Where θ represents the model parameters, and C represents the number of input channels (usually 1 or 3). The pruning and reconstruction strategies are as follows:

[0258] 1.1 Retain the underlying feature extraction module:

[0259] The C2f module structure (Conv→BN→SiLU→DWConv) is repeatedly stacked;

[0260] Preserve shallow residual connections to enhance gradient propagation and local structure representation;

[0261] 1.2 Remove large target detection branches and PAN / FPN layers:

[0262] Remove the detection-oriented path aggregation structure (Neck) from YOLOv8.

[0263] Only the classification head is retained for outputting the predicted probability vector.

[0264] 1.3 Introduce a global average pooling layer:

[0265] In the main output characteristics Apply GAP:

[0266] z d =(1 / H'W')∑ i＝1 H' ∑ j＝1 W' F i,j,d ,d=1,...,D

[0267] 2. Classification Head Structure

[0268] The construction format is as follows:

[0269]

[0270] in:

[0271] Output vector for GAP;

[0272] The dropout probability is p = 0.2, used to suppress overfitting;

[0273] Softmax is defined as follows:

[0274] Softmax(y k ) = e y k / ∑ j＝1 10 e y j ,k=0,...,9.

[0275] 3. Attention-Based Feature Refinement

[0276] To enhance the ability to distinguish between blurry and poorly edged digital images, a lightweight attention module A is introduced into the middle layer of the backbone, including channel attention and spatial attention:

[0277] 3.1 Optionally, the SE (Squeeze-and-Excitation) or CBAM (Convolutional Block Attention Module) module can be embedded into the intermediate layer. The insertion position is usually after the last two C2f blocks of the main body to extract high-level semantic features. The implementation method is to scale the intermediate feature map channel by channel to enhance the robustness of the model to distortion scenarios such as blurred strokes and broken boundaries.

[0278] Use the SE module (Squeeze-and-Excitation)

[0279] The input feature map is First, perform global compression:

[0280] s c =(1 / HW)∑ i＝1 H ∑ j＝1 W F i,j,c c = 1, ..., C

[0281] 3.2 Generating scaling factors through activation functions:

[0282] α=σ(W2·δ(W1·s))

[0283] Where δ is ReLU, σ is Sigmoid, and the final output is a channel-scaled feature map:

[0284] F' i,j,c =αc ·F i,j,c

[0285] Using the CBAM module:

[0286] Spatial attention mechanisms further enhance attention to strokes / boundaries:

[0287] M s =σ(f 7×7 ([AvgPool(F);MaxPool(F)]))

[0288] The final output is:

[0289] F”=F'⊙M s , where ⊙ represents element-wise multiplication.

[0290] In a preferred embodiment of the present invention, in step S3, a digit recognition model is used to perform digit recognition on each digit character area and the decimal area. The output result of the digit recognition model after performing digit recognition on the images in each digit character area and the decimal area is as follows:

[0291] A probability vector of length 10 is generated, which includes each possible digit to be identified and its corresponding probability value. Then, the category with the highest probability value is selected from the probability vector as the final recognition result output.

[0292] If any probability value is less than the preset confidence threshold, the system will either reject the data or proceed to the Top-K post-processing logic. The Top-K post-processing logic includes outputting the K numbers sorted by probability and their confidence sets.

[0293] Specifically, the output format of the digit recognition model is defined as follows:

[0294] Final prediction results Let be a probability vector of length 10, where the k-th term represents the probability that the image belongs to the digit k:

[0295]

[0296] in:

[0297] c i ∈[0,1,...,9]; represents the final classification result (output value) of the i-th number, such as 8, 3, etc.

[0298] conf i ∈[0,1],conf i =P(c i |input), which is the highest probability term in the softmax output; supports confidence-based circuit breaking mechanism for judging credibility in the post-processing stage.

[0299] argmax indicates that the parameter corresponding to the maximum value is returned (i.e., the variable that maximizes the function value is returned);

[0300] This represents the probability value of the k-th class output by Softmax.

[0301] Optional, also includes:

[0302] Confidence threshold filtering: If confidence threshold filtering is used... i <τ, can be transferred to Top-K post-processing logic.

[0303] Top-K candidate output (K = 2 or 3, corresponding to Top-2 / Top-3 inference mode): Output the sorted K numbers and their confidence scores:

[0304]

[0305] It can be used for candidate voting or confidence comparison in subsequent modules (such as multi-frame fusion or character context correction).

[0306] In a preferred embodiment of the present invention, the image of each numeric character area has a corresponding numeric frame, and the meter reading includes integer area reading and decimal area reading. In step S3, the process of splicing the integer area readings includes:

[0307] The number frames obtained from the images of each number character area are sorted from left to right according to the horizontal position of the corresponding number character area in the physical layout of the integer area.

[0308] The sorted number boxes are adjusted for consistent row height and width ratio to perform physical constraint filtering.

[0309] The readings of the integer region are obtained by merging and concatenating the number boxes after physical constraints into a number string.

[0310] Specifically, in this embodiment, the splicing process uses a post-processing module to splice multiple digits, combining higher-order single digits into a complete reading according to the physical layout. The specific process includes:

[0311] 1. Spatial sorting

[0312] Each detected digit box is sorted from left to right according to its horizontal position in the integer region:

[0313] sorted_digits = Sort({(x i ,y i ,w i ,h i ,c i byxi

[0314] Where (x) i ,y i Let be the coordinates of the center point of the i-th digit frame, and c be the coordinates of the center point of the i This is the classification result.

[0315] 2. Physical constraint filtering

[0316] Row height consistency: Vertical center difference between adjacent boxes Δy < τ y (e.g., 5 pixels) Ensure they are in the same row.

[0317] Width ratio: Width w of each box i They should be approximately equal, (max w - min w) / min w < α (e.g., 10%), to exclude false positives.

[0318] 3. Merging and splicing

[0319] Concatenate the sorted category labels into a string in order:

[0320] reading = [c (1) ||c (2) ||...||c (L) ]

[0321] Where L is the number of integer digits (e.g., 5 or 6 digits).

[0322] Set up verification rules to apply a series of business and statistical rules to the splicing results, in order to eliminate unreasonable results caused by false detections or extreme conditions. The verification rules include:

[0323] 1. Continuity check (Preventing Mechanical Jump)

[0324] Numerical jump threshold: The difference between adjacent numbers should be less than the threshold Δvmax (e.g., 5).

[0325] |int(c i )-int(c i-1 )|≤Δv max

[0326] Roller angle consistency: If angle detection is implemented, the maximum rotation angle difference θmax between adjacent rollers can be limited.

[0327] 2. Business Logic Validation Table

[0328]

[0329] 3. Confidence Fusion

[0330] Single-digit confidence level: If a certain conf i <T low If the value is 0.9, then the marking needs to be manually verified.

[0331] Overall confidence level:

[0332] conf=(1 / L)∑ i＝1 L conf i 2

[0333] If conf <T overall If the value is 0.85, then reject.

[0334] When a confidence level failure event occurs (resulting in "Reject"), a resampling event will be triggered, requiring the terminal acquisition device to re-acquire images. If a resampling event is triggered three times consecutively or multiple times within a period of time (e.g., five times within a week), the system will flag the corresponding device and push the information to maintenance personnel.

[0335] Existing technologies also face the problem of balancing algorithm efficiency and accuracy: solutions based on complex post-processing (such as perspective transformation and image enhancement) can partially alleviate the interference problem, but the computational overhead is large and it is difficult to meet the real-time requirements of embedded devices; while lightweight models often perform poorly in low-quality images due to insufficient feature extraction capabilities.

[0336] To address the challenge of balancing efficiency and accuracy in existing algorithms, this invention achieves a dual improvement in both through lightweight model design and post-processing module optimization. Regarding lightweight model design, this invention employs lightweight models such as YOLOv8n, significantly reducing computational overhead. Simultaneously, by optimizing the loss function and enhancing feature extraction capabilities, detection accuracy is maintained. In terms of the post-processing module, this invention introduces a multi-digit concatenation and confidence-based circuit breaking mechanism. This concatenates discrete single digits into a complete reading according to their physical layout. By setting a confidence threshold, when the confidence of a particular digit or the overall score falls below the threshold, a resampling event is triggered, or the result is marked for manual review.

[0337] In summary, this invention effectively solves the problems of poor compatibility with multiple instrument specifications, insufficient robustness to complex imaging interference, and difficulty in balancing algorithm efficiency and accuracy in traditional visual inspection of measuring instruments by comprehensively applying a multi-stage cascaded detection method, data-level enhancement strategy, post-processing module optimization, and production-level data management module. Experimental results show that this invention exhibits significant advantages in terms of accuracy, efficiency, cost, and robustness, achieving an F1-score of 98.7%, reducing deployment costs by 70%, and achieving an error reading rate of less than 0.3%. Furthermore, it supports hybrid cloud / edge computing deployment and has broad prospects for industrial applications.

[0338] Furthermore, it also includes a production-level data management module. This module serves as the core of the overall system's data and model management, responsible for supporting highly reliable file storage, version control, API service integration, and structured data publishing under large-scale deployments, and has the ability to seamlessly integrate with industrial IoT platforms.

[0339] The system architecture of the production-grade data management module is as follows: It is a distributed object storage system built on the open-source MinIO platform, compatible with the Amazon S3 protocol, providing a unified solution for managing detection data, model files, and logs. The core structure is as follows:

[0340]

[0341]

[0342] The naming conventions for object buckets are as follows:

[0343]

[0344] Each object is stored using a uniform path format for easy retrieval and version control. The storage path is as follows:

[0345] detection-data / (device_id) / (date) / (timestamp).jpg

[0346] model-weights / (model_name) / (version) / (file_name).pt

[0347] logs / (component) / (date) / (log_id).txt

[0348] Supports model deployment and rollback based on object version number:

[0349] Model_v_t=model-weights / model_name / v_t / model.pt

[0350] Supports data consistency synchronization in multi-node deployments (Erasure Coding + Hash ConsistencyCheck).

[0351] It also supports automatic generation of HTML / JSON reports:

[0352] JSON: Standard structure output for secondary parsing by the IIoT platform;

[0353] HTML: Automatically generates human-readable reports using a template engine, supporting features such as image embedding and confidence chart visualization.

[0354] The above are merely preferred embodiments of the present invention and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should recognize that any equivalent substitutions and obvious changes made using the content of this specification and illustrations should be included within the protection scope of the present invention.

Claims

1. An automatic identification method for readings of general measuring instruments, characterized in that, include: Step S1: Acquire an image of the measuring instrument to obtain an original image of the instrument panel area, and perform image recognition on the original image to obtain the integer area and decimal area in the instrument panel area and their corresponding coordinates. Determine the physical layout position of the integer area and the decimal area in the instrument panel area based on the coordinates. Step S2: Divide the integer area into multiple numeric character areas; Step S3: Perform digit recognition on each of the digit character areas and the decimal area, and then sequentially splice the recognized digits according to their corresponding physical layout positions to obtain the meter reading.

2. The automatic identification method according to claim 1, characterized in that, The execution of step S1 also includes an image distortion correction process, including: Step A1: Before identifying the dashboard area in the original image, perform global distortion correction on the original image; And / or step A2, after identifying the integer and decimal regions in the dashboard area of ​​the original image, correcting the original image using a perspective transformation matrix.

3. The automatic identification method according to claim 1, characterized in that, In step S1, the original image is subjected to image recognition using a dashboard positioning model. The dashboard positioning model is an improvement upon the YOLO model, and its improvement process includes: Step B1 introduces an anchorless mechanism into the initial model of the YOLOv8n architecture, using each pixel on the output feature map of the initial model as a candidate detection point, and marking the distance from the candidate detection point to the boundary of the target box to construct the target box, thereby realizing the target box selection of the dashboard area in the original image. Step B2: The detection head of the initial model adopts a classification and regression decoupled structure so that the confidence branch and the target box regression branch in the initial model are processed by independent sub-networks respectively. Step B3: Introduce the DFL loss function into the loss function architecture of the initial model to realize a distributed prediction mechanism for each boundary point position during the box regression process, outputting multiple discrete probability distribution points to approximate the continuous regression target. Step B4: Replace the box-frame regression function in the loss function architecture of the initial model with the EIoU loss function, and introduce a layout consistency loss term.

4. The automatic identification method according to claim 1, characterized in that, When a new original image of a new meter that has not been input before is input into the dashboard positioning model, a new sample set is constructed based on multiple new original images of the same new meter. Then, the dashboard positioning model is subjected to lightweight incremental training based on the new sample set to achieve fine-tuning of the model parameters.

5. The automatic identification method according to claim 1, characterized in that, In step S1, the dashboard positioning model is used to perform image recognition on the original image, and the dashboard positioning model is trained using a hybrid dataset, which includes real images captured by the camera and synthetic images rendered by the camera. The process for generating the synthesized image includes: Create a 3D geometric model of the dashboard, then set the material and anti-texture of the 3D geometric model, set the background of the 3D geometric model, and set the light source to illuminate the 3D geometric model; The camera parameters of the camera are randomly set, and attitude perturbation and distortion simulation are added to the camera; The camera samples the 3D geometric model image from multiple angles, saves each frame of the sampled image during the image sampling process, and performs image correction and perspective perturbation on the sampled images to obtain an enhanced image. Then, all sampled images and the enhanced image are saved. The two-dimensional bounding boxes of the integer and decimal regions in the sampled image and the enhanced image are labeled to obtain corresponding YOLO format labels. The labeled sampled image and the enhanced image are then used as the composite image.

6. The automatic identification method according to claim 1, characterized in that, In step S2, a segmentation network model is used to segment multiple numeric character regions from the integer region. The segmentation network model is the YOLOv8n-Seg segmentation model. The backbone network of the YOLOv8n-Seg segmentation model uses CSPDarknet-Lite based on the C2f module. The image size input to the segmentation network model is fixed at 640*640; An FPN+PAN path is introduced into the neck structure of the YOLOv8n-Seg segmentation model; The head structure of the YOLOv8n-Seg segmentation model adopts a three-branch output structure, namely class confidence, location coordinates and mask prototype features, and the size of the mask segmentation result is [B,N,32,32], where B is the target box and N is the number of detected targets; The loss function of the YOLOv8n-Seg segmentation model is composed of cross-entropy loss, enhanced IOU loss, mask quality optimization loss, and adversarial robustness loss term.

7. The automatic identification method according to claim 6, characterized in that, The segmentation network model is trained using an image augmentation dataset, and the process of constructing the image augmentation dataset includes: Multiple original images of measuring instruments are acquired. During the training phase, first-order gradient perturbation is added to the original images to enhance adversarial noise perturbation, and / or simulated reflection interference to generate spot mask for pseudo-reflection mask synthesis, and / or image structure boundaries are enhanced, and the distinction between characters and background is improved for color normalization and contrast enhancement. Finally, the data-enhanced images are added to the image enhancement dataset.

8. The automatic identification method according to claim 1, characterized in that, In step S3, a digit recognition model is used to perform digit recognition on each of the digit character areas and the decimal area. The digit recognition model is obtained by pruning and reconstructing the main structure of the YOLOv8 model. The pruning and reconstruction process includes: The C2F structure and shallow residual connections in the backbone of the YOLOv8 model are retained. Remove the path aggregation structure from the YOLOv8 model and remove the detection head, keeping only the classification head; Add a global average pooling layer to the end of the pruned backbone structure of the YOLOv8 model; Reconstruct the classification header; A lightweight attention module is embedded in the backbone structure of the trimmed YOLOv8 model.

9. The automatic identification method according to claim 1, characterized in that, The output result of performing digit recognition on the images in each of the digit character areas and the decimal area in step S3 is as follows: A probability vector of length 10 is generated, which includes each possible digit and its corresponding probability value. Then, the category with the highest probability value is selected from the probability vector as the final recognition result. If any of the probability values ​​is less than a preset confidence threshold, the system either rejects the data or proceeds to the Top-K post-processing logic, which includes outputting K numbers sorted by probability and their confidence sets.

10. The automatic identification method according to claim 1, characterized in that, Each of the aforementioned digit character areas has a corresponding digit frame. The meter reading includes integer area readings and decimal area readings. In step S3, the process of splicing the integer area readings includes: The number frames of multiple numbers identified from the images of each of the number character areas are sorted from left to right according to the horizontal position of the corresponding number character area in the physical layout position of the integer area; The sorted number boxes are adjusted for consistent row height and width ratio to perform physical constraint filtering. The integer region reading is obtained by merging and concatenating the various number frames after physical constraints into a number string.

Images