Pointer type instrument reading identification method based on staged detection and inclination over-limit correction

Abstract

The invention provides a pointer instrument reading identification method based on staged detection and inclination over-limit correction. The method comprises the following steps: firstly, acquiring multi-type pointer instrument images, constructing a staged data set and performing data enhancement; training an instrument panel target detection model by using the first part data set, positioning the input image and cutting an instrument panel area; training a key point detection model based on the second part data set, and extracting a starting scale point, a termination scale point, a center point, a pointer tip point and a termination scale value; carrying out normalization processing on the detected key points, judging whether the detected key points exceed an inclination allowable range or not, and executing key point perspective transformation correction on the instrument panel with the inclination exceeding the limit; and finally, combining an angle method and measuring range information to calculate readings. According to the method, only the inclined overrun sample is subjected to geometric correction, full-amount processing is not needed, the calculation power consumption and the processing time can be remarkably reduced while the reading precision is guaranteed in large-scale batch detection, and high precision and high efficiency are achieved.

Description

Technical Field

[0001] This invention relates to the field of image processing and recognition technology, specifically to a method for recognizing pointer instrument readings based on phased detection and tilt over-limit correction. Background Technology

[0002] Pointer-type instruments are widely used in industrial production, energy monitoring, and transportation due to their low cost and high reliability. Their reading accuracy directly affects equipment operation safety and data accuracy. In traditional manual inspection, operators must read and record instrument values ​​one by one, which is inefficient and susceptible to human error. For the inspection of a large number of distributed instruments, this method is not only time-consuming and labor-intensive but also suffers from data delays and reading deviations. While existing computer vision-based automatic reading methods can improve efficiency to some extent, common solutions have the following shortcomings: Single-stage inspection makes it difficult to maintain high accuracy while simultaneously ensuring instrument panel positioning and precise detection of key points; furthermore, while tilted instruments are a major factor affecting reading accuracy, the impact of a small tilt range on reading error is minimal, and tilt correction is unnecessary within acceptable error ranges. However, existing tilt correction methods typically perform geometric transformations on all samples, resulting in significant computational resource consumption and limited processing speed in batch inspections; some methods show a significant decrease in recognition accuracy under uneven lighting, angle deviations, and partial occlusion. Therefore, the main problem this invention aims to solve is the insufficient accuracy, excessive computational resource consumption, and low processing efficiency of existing automatic reading methods for pointer instruments in batch testing. By proposing a pointer instrument reading recognition method based on staged detection and tilt over-limit correction, unnecessary computational overhead can be significantly reduced, and detection efficiency can be improved without affecting reading accuracy. This method is particularly suitable for large-scale batch instrument rapid testing scenarios such as industrial inspection and energy monitoring. Summary of the Invention

[0003] This invention aims to solve the problems of high computational resource consumption and low processing efficiency in existing methods for instrument testing, especially in scenarios such as large-scale batch instrument rapid testing in industrial inspection and energy monitoring. It proposes a pointer-type instrument reading recognition method based on staged testing and tilt over-limit correction.

[0004] The objective of this invention can be achieved through the following technical solution: This invention proposes a method for identifying pointer instrument readings based on staged detection and tilt over-limit correction, comprising:

[0005] Step 1: Image Acquisition and Dataset Construction. Industrial cameras were used to acquire images of various types of pointer instruments under multiple angles and lighting conditions. The dataset was divided into two parts: a dashboard target detection set and a dashboard keypoint detection set. Labeling tools such as Labelme were used for annotation, and the training and test sets were divided in a ratio of 8:2.

[0006] Step 2: Data Augmentation. Perform various data augmentation operations on the training dataset, such as affine transformation, HSV color space transformation, random flipping, random occlusion, and Mosaic, to increase sample diversity and improve the model's generalization ability.

[0007] Step 3: First-stage dashboard detection and cropping. Based on the first part of the dataset, the dashboard bounding boxes are labeled, and the first-stage object detection model is trained. This model is used to detect objects in the input image and crop the data to obtain a standardized dashboard region image, providing high-quality input for subsequent keypoint detection.

[0008] Step 4: Second stage dashboard key point detection. In the second part of the dataset, the start scale point, end scale point, center point, pointer tip point, and end scale value detection box are labeled, and the key point detection model is trained to accurately locate key points in the cropped dashboard image.

[0009] Step 5: Keypoint Normalization and Tilt Determination. The coordinates of the detected keypoints are normalized according to the image width W and height H to obtain the normalized coordinates (X). A’ ,Y A’ It compares the tilt with a preset tilt threshold to determine whether the dashboard exceeds the allowable tilt range.

[0010] Step 6: Tilt Correction For dashboards with excessive tilt, extract four non-collinear points on the edge of the dial scale, calculate the perspective transformation matrix with the corresponding points of the standard template, and update the key point coordinates.

[0011] Step 7: Angle method reading calculation. Using the corrected coordinates of the starting point, ending point, center point, and pointer tip, calculate the angle between ∠start-middle-end and ∠start-middle-point to obtain the angle ratio value, and combine it with the instrument range to calculate the final reading result.

[0012] This invention significantly improves the targeting and accuracy of feature extraction by separating dashboard detection from key point detection; it introduces normalized tilt determination and perspective transformation correction to effectively reduce reading deviations caused by changes in shooting posture; and it combines angle method calculation to ensure reading accuracy under different ranges and scale distributions. Attached Figure Description

[0013] Figure 1 This is an overall flowchart of the method of the present invention; Figure 2This is a schematic diagram of the dashboard inspection results in the first phase.

[0014] Figure 3 This is a schematic diagram of the key point detection results for the dashboard in the second phase. Figure 4 This is a schematic diagram illustrating the principle of angle-based reading calculation. Detailed Implementation

[0015] The present invention will be further described below with reference to embodiments. The description of the embodiments below is only for the purpose of helping to understand the present invention. It should be noted that those skilled in the art can make several modifications to the present invention without departing from the principle of the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

[0016] The overall process of the method of this invention is as follows: Figure 1 As shown. Specifically includes:

[0017] Step one: First, use an industrial camera to collect images of various pointer instruments and build a phased dataset to ensure that it includes samples from different angles, lighting conditions and environmental conditions.

[0018] Step two: Perform data augmentation on the images in the training set to expand the amount of data and feature categories of the instrument images in the training set.

[0019] The data was then labeled. During the labeling process, the first part of the dataset was labeled with the dashboard bounding box, and the second part of the dataset was labeled with the key points related to the readings.

[0020] Step 3: Train the model on the first part of the training dataset to obtain the dashboard detection model, which specifically includes:

[0021] The first stage, dashboard detection, uses an improved lightweight YOLO11 model to detect and locate the dashboard position, and then crops a standardized image to effectively reduce background interference. The main model improvements include the following:

[0022] 1) GhostConv replaces shallow convolutions and fused convolutions to minimize computational load in high-resolution stages;

[0023] 2) Modules such as DSConv, DS-Bottleneck, and DS-C3k were constructed using depthwise separable convolutions to replace traditional large-kernel convolutions. For example, the DS-C3k2 module was used as a lightweight backbone network to extract multi-scale features, which significantly reduced the number of parameters and computational cost while maintaining the receptive field, thus improving the computational efficiency of the model.

[0024] 3) Scales limit the width / depth and the maximum number of channels to avoid wasting computing power due to excessive width and facilitate horizontal migration based on device performance.

[0025] Step four involves training the model on the second part of the training dataset to obtain the dashboard keypoint detection model, specifically including:

[0026] The second stage involves key point detection on the dashboard. The YOLO11-Pose model is used to perform fine positioning of local areas of the dashboard and output the coordinates of key points for reading calculation.

[0027] The model incorporates a keypoint regression branch after the detection head, which can simultaneously output the target bounding box and the corresponding keypoint coordinates, enabling multi-task joint learning. Compared to traditional two-stage pose estimation methods, YOLO11-Pose offers advantages such as fast inference speed, low latency, and easy deployment.

[0028] Step 5: After normalizing the keypoint coordinates, a tilt threshold is used for judgment. For samples exceeding the threshold, perspective transformation correction is performed to obtain the corrected keypoint coordinates. This specifically includes:

[0029] Each second-stage instrument image has a width of W and a height of H. The starting point, ending point, center point, pointer tip, and the detection frame for the ending scale value are marked, and related points are depicted and their coordinate information is recorded. The detection frame is a rectangle, and the normalized coordinates of the key points on the instrument panel are (X...). A’ ,Y A’ Let the coordinates of the key points identified by the model be (X). A ,Y A ),but .

[0030] Based on the normalized coordinates of the key points, calculate the distance ratio from the center point to the starting point and the ending point, as well as ∠. 起-中-终 Let the size be T. A T B Therefore, a tilt threshold is set. If both conditions are met, the tilt effect can be ignored; otherwise, the critical point correction stage begins. The threshold setting is determined by the tolerance range of the instrument reading error.

[0031] Step six, based on the principle of the angle method, such as Figure 4 As shown, M1 represents the line connecting the starting point and the center point of the instrument, M2 represents the line connecting the ending point and the center point of the instrument, M3 represents the pointer, θ1 represents the angle between the pointer pixel and the starting point pixel, and θ2 represents the angle between the pointer pixel and the ending point pixel. Assuming the instrument range is L, the reading R can be calculated as: .

[0032] This method demonstrates high reading stability and applicability in industrial field testing and can be widely applied to scenarios such as automated inspection, remote monitoring, and intelligent detection.

Claims

1. A method for identifying pointer instrument readings based on staged detection and tilt over-limit correction, characterized in that, The method includes: Step 1-1: Obtain images of various types of pointer-type instruments; Steps 1-2: Perform data augmentation on the images in the training dataset; Steps 1-3: Label the first part of the instrument image dataset and train the first stage instrument target detection model, and crop the model recognition results; Steps 1-4: Label the second part of the instrument image dataset and perform the second stage of instrument panel key point model training and detection; Steps 1-5: Normalize the coordinates of the corresponding key points and determine whether the instrument exceeds the allowable tilt range. Steps 1-6: Perform tilt correction on instruments that are outside the tilt range; Steps 1-7: Apply the angle method to calculate the readings and obtain the instrument reading results.

2. The pointer instrument reading recognition method based on phased detection and tilt over-limit correction according to claim 1, characterized in that, In step 1-1, the dataset is divided into two parts: the first part is the dashboard target detection dataset, and the second part is the dashboard key point detection dataset. The data is captured using an industrial camera, and the captured images are labeled using Labelme annotation software. The labeling results are divided into training set and test set in an 8:2 ratio.

3. The pointer instrument reading recognition method based on phased detection and tilt over-limit correction according to claim 1, characterized in that, In steps 1-2, the captured images undergo affine transformation, HSV transformation, random flipping, random occlusion, and Mosaic algorithm data augmentation.

4. The pointer instrument reading recognition method based on phased detection and tilt over-limit correction according to claim 1, characterized in that, Steps 1-3, the first part of the dashboard image dataset annotation, includes: Each image in the first part of the instrument panel image target detection dataset is labeled to obtain the first part of the instrument panel training dataset, wherein the labeling process is used to determine the bounding box of each target instrument panel in the image; the first part of the instrument panel training dataset is used to train a model to obtain the first stage instrument panel target detection model.

5. Model training using the first part of the dashboard training dataset according to claim 4, characterized in that, The model described is an improved lightweight YOLO11, comprising: GhostConv replaces shallow convolutions and fused convolutions to minimize computational load in high-resolution stages. DSConv, DS-Bottleneck, and DS-C3k modules were constructed using depthwise separable convolutions to replace traditional large-kernel convolutions. For example, the DS-C3k2 module was used as a lightweight backbone network to extract multi-scale features, which significantly reduced the number of parameters and computational cost while maintaining the receptive field, thus improving the computational efficiency of the model. Scales limit the width / depth and the maximum number of channels to avoid wasting computing power due to excessive width, and facilitate horizontal migration based on device performance.

6. The pointer instrument reading recognition method based on phased detection and tilt over-limit correction according to claim 1, characterized in that, The annotation process for the second part of the dashboard image dataset in steps 1-4 includes: The starting point, ending point, center point, pointer tip point, and termination scale value detection box are marked, and the relevant points are drawn and their coordinate information is recorded; the model is trained on the second part of the dashboard training dataset to obtain the second stage dashboard target detection model.

7. Model training using the second part of the dashboard training dataset according to claim 1, characterized in that, The training model mentioned is YOLO11-Pose, which includes: The model incorporates a keypoint regression branch after the detection head, which can simultaneously output the target bounding box and the corresponding keypoint coordinates, enabling multi-task joint learning. Compared to traditional two-stage pose estimation methods, YOLO11-Pose offers advantages such as fast inference speed, low latency, and easy deployment.

8. The pointer instrument reading recognition method based on phased detection and tilt over-limit correction according to claim 1, characterized in that, Steps 1-5, the normalization process, include: Let the width of each second-stage instrument image be W and the height be H. Mark the starting point, ending point, center point, pointer tip, and the detection frame for the ending scale value, and draw the relevant points, recording their coordinate information; wherein, the detection frame is a rectangle, and the normalized coordinates of the key points on the instrument panel are (X... A’ ,Y A’ Let the coordinates of the key points identified by the model be (X). A ,Y A ),but .

9. The pointer instrument reading recognition method based on phased detection and tilt over-limit correction according to claim 1, characterized in that, The tilt correction described in steps 1-6 also includes the following steps: Step 9-1: Select four non-collinear points on the edge of the instrument panel scale, label the second part of the instrument panel image data, and train the model. Step 9-2: Select a standard dashboard image as a template, detect four points from the second part of the dashboard dataset, and calculate the transmission transformation matrix by comparing it with the four points of the standard dashboard image. Step 9-3: Calculate the correction coordinates of the corresponding key points on the dashboard based on the matrix.

10. The pointer instrument reading recognition method based on phased detection and tilt over-limit correction according to claim 1, characterized in that, Steps 1-7 describe the angle method for calculating readings, including: Based on the key point detection, the coordinates of key points such as the starting point, center point, ending point, and pointer tip are obtained, and ∠ is calculated for each. 起-中-终 With ∠ 起-中-指 The angle size is determined to obtain the angle ratio value, and finally, the reading result is calculated based on the range.

Images