A visual perception enhanced instrument pointer reading recognition method and system

Abstract

The present application relates to a kind of visual perception enhanced instrument pointer reading identification method and system, the method comprises the following steps: S1, acquisition and transmission pointer instrument dial image;S2, the pointer instrument dial image extracted by using multiscale Gaussian kernel with the pointer instrument dial image illumination intensity component collected, and reconstruct two-dimensional Gamma function to adaptively correct pointer instrument dial image illumination intensity;S3, the pointer instrument dial image after illumination intensity correction is input to the haze removal elliptical network model that has been trained, removes haze interference and restores clear pointer instrument dial image;S4, centripetal force constraint is added in line segment detector, clear pointer instrument dial image pointer accurate positioning is realized, and the function relationship between the angle of pointer and reading is deduced to realize accurate reading.It is beneficial to be little by light intensity and fog environment Influence, automatic accurate reading meter.

Description

[Technical Field]

[0001] This invention relates to the field of power equipment technology, and specifically to a method and system for recognizing instrument pointer readings with enhanced visual perception. [Background Technology]

[0002] Automatic and accurate reading of pointer-type meters includes pointer positioning and reading recognition. Pointer positioning determines the pointer position using machine vision or image processing technology. Reading recognition establishes a mapping relationship between the pointer deflection angle and the meter reading. For field equipment, automatic and accurate reading of pointer meters presents two challenges: 1) Detection equipment often operates outdoors, requiring pointer meter reading methods to be adaptable to varying light intensities. 2) Due to environmental complexity and pollution, cameras are often exposed to smog interference during meter reading operations. Pointer-type meter reading methods need to be suitable for foggy environments. For adaptive light intensity, the commonly used histogram equalization method achieves illumination correction by changing the grayscale value of each pixel in the histogram. However, it suffers from over-enhancement and color distortion. Partial differential equations correct the contrast field by amplifying the image through solving nonlinear relationships, but this requires the image to be differentiable, which also has certain limitations. Regarding smog interference, researchers have proposed various hypotheses for reconstructing smog images. This mainly involves prior processing and deep learning techniques. Most techniques fail to recognize the heterogeneity of haze, uniformly processing spatial regions during haze removal; this contradicts human perception. Simply accumulating enhancement and dehazing algorithms can easily lead to problems adapting to non-uniform phenomena, thus requiring structural optimization based on existing technologies. This allows the haze removal process to recover a clear image from haze-interferenced images without constructing an optical scattering model, ensuring the independence of optical correction and dehazing processes. On the other hand, in meter reading network projects, the most important thing is to establish a real-time communication system with high transmission efficiency and stable performance. However, with the rapid iteration of technology and the rapid development of the Internet of Things, the communication scenarios are becoming increasingly complex. Traditional wireless transmission technologies such as LoRa, Wi-Fi, and Zigbee can no longer simultaneously achieve low power consumption and long-distance transmission. In a networked transmission system, the gateway is also a crucial component, determining the stability of the wireless communication system, making its optimization particularly important.

[0003] This invention addresses the technical problem that the automatic and accurate readings of pointer instruments are affected by light intensity and foggy environments, and makes technical improvements to the instrument pointer reading recognition method and system. [Summary of the Invention]

[0004] The purpose of this invention is to provide a method for identifying meter pointer readings that is less affected by light intensity and foggy environments and can automatically and accurately read and record meter values.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is a visual perception-enhanced instrument pointer reading recognition method, comprising the following steps:

[0006] S1. Acquire and transmit the image of the pointer instrument panel;

[0007] S2. Extract the illumination intensity component of the acquired pointer instrument dial image using a multi-scale Gaussian kernel, and reconstruct a two-dimensional Gamma function to adaptively correct the illumination intensity of the pointer instrument dial image.

[0008] S3. Input the pointer instrument dial image after light intensity correction into the trained dehazing elliptical network model to remove haze interference and restore a clear pointer instrument dial image.

[0009] S4. Add centripetal force constraints to the line segment detector to achieve clear pointer instrument dial image and accurate pointer positioning, and derive the functional relationship between pointer angle and reading to achieve accurate reading.

[0010] Preferably, step S2 includes the following sub-steps:

[0011] S21. Construct a multi-scale Gaussian model and use it as the illumination component of the original acquired pointer instrument dial image. Where G(x,y) is a Gaussian function, F(x,y) represents the brightness value of the image at coordinate (x,y), I(x,y) represents the illumination intensity component at coordinate (x,y), * represents convolution operation, i = 1, 2, ..., N are the number of Gaussian functions of different scales, and wi represents the illumination intensity component weight of the i-th Gaussian function.

[0012] S22. Construct a two-dimensional Gamma function model as a correction model for the illumination components of an instrument dial image. Where O(x, y) represents the illumination value after the pointer instrument dial image is corrected, r represents the illumination correction parameter, b represents the illumination coefficient, and m (m∈[0, 255]) represents the average value of the illumination components;

[0013] S23. Set the r, b, and m parameters, and output the pointer instrument dial image after light intensity correction.

[0014] Preferably, the haze-removing elliptic network model includes an extraction module for extracting low-frequency features, a multi-scale attention network for learning haze features, and a multi-feature fusion module for reconstructing haze-free images; the extraction module includes a ReLU activation layer and a convolutional layer, the multi-scale attention network includes a channel attention subnetwork and a spatial attention subnetwork, and the multi-feature fusion module includes a convolutional layer, an element stacking layer, and a Sigmad activation layer.

[0015] Preferably, the dehazing elliptic network model training in step S3 includes the sub-steps of constructing and preprocessing the pointer instrument dial image interference dataset: collecting pointer instrument dial image data under different scenarios through an image acquisition device to construct the original dataset, then removing the interference data that affects the training and recognition of the neural network, manually adding haze effects to the remaining data, and labeling the pointer instrument dial category.

[0016] Preferably, step S4, precise pointer positioning, includes the following sub-steps:

[0017] S41. The first-order gradient components of each pixel (x, y) in the four directions of the clear pointer dial image were calculated. Among them, is g x (x,y) is the first-order gradient component in the x-direction, g y (x,y) is the first-order gradient component in the y-direction, g 45 (x,y) is the first-order gradient component in the 45° direction, g 135 (x,y) is the first-order gradient component in the 135° direction;

[0018] S42. In the line segment detector, a centroid constraint is added where the angle of the horizontal line is close to the angle of the center of the clear pointer instrument dial image. This reduces the pointer positioning calculation time and improves the pointer positioning accuracy. Centroid constraint mechanism expression. Where (xo, yo) are the coordinates of the center point of the clear pointer instrument panel image, θ(xi, yi) is the angle between the pixel (xi, yi) and the x-axis, and th1 is a set threshold used to filter feature points in the candidate region. Without affecting the result, the smaller the value of th1, the shorter the computation time. The expression for θ(xi, yi) is as follows:

[0019]

[0020] S43. The expression for region growth after adding centripetal constraint prior conditions during the segmentation of feature regions in a clear pointer instrument dial image is as follows: Where I is the set of all pixels in the clear pointer instrument dial image, th2 is the set threshold to remove the uncertainty of small gradient value points; the longest line segment in the segmented pointer and scale segment set is the pointer, realizing precise pointer positioning.

[0021] Preferably, step S4: Divide the pointer instrument dial into four quadrants to calculate the relationship between the pointer angle and the dial scale for accurate reading. When the identified pointer is in the first quadrant, the pointer instrument dial scale reading is x0 + G*(180° + θ) / θ0; when the identified pointer is in the second quadrant, the pointer instrument dial scale reading is x0 + G*θ / θ0; when the identified pointer is in the third quadrant, the pointer instrument dial scale reading is x0 - G*θ / θ0; when the identified pointer is in the fourth quadrant, the pointer instrument dial scale reading is x0 - G*(180° + θ) / θ0. Here, the pointer scale range is G, and the angle range corresponding to the pointer scale range is θ0. Taking the center point of the dial as the center point, the left scale value x0 where the dial intersects the X-axis is used as the configuration parameter.

[0022] Preferably, the visual perception-enhanced instrument pointer reading recognition method includes the following steps:

[0023] S1, The camera sensor collects and transmits images of the pointer instrument dial, which are then transmitted from the camera node to the back-end processing module via the WAPI gateway.

[0024] S2, the back-end processing module's illumination correction submodule extracts the illumination intensity components of the acquired pointer instrument dial image using a multi-scale Gaussian kernel, and reconstructs a two-dimensional Gamma function to adaptively correct the illumination intensity of the pointer instrument dial image.

[0025] S3, the backend processing module's dehazing submodule inputs the light intensity-corrected pointer instrument dial image into the pre-trained dehazing elliptical network model to remove haze interference and restore a clear pointer instrument dial image;

[0026] S4. The pointer positioning submodule of the back-end processing module adds centripetal force constraints to the line segment detector to achieve accurate pointer positioning of the pointer in the clear pointer instrument dial image, and derives the functional relationship between pointer angle and reading to achieve accurate reading.

[0027] S5. The backend processing module transmits the reading results and the processing results of the illumination correction submodule, the dehazing submodule, and the pointer positioning submodule to the client. The client uploads the step processing results to the cloud data storage and provides a unified query and management page for the results.

[0028] Another objective of this invention is to provide an instrument pointer reading identification system that is less affected by light intensity and foggy environments and can automatically and accurately read and record meter values.

[0029] To achieve the aforementioned objective, the present invention provides a visual perception-enhanced instrument pointer reading recognition system, comprising several camera nodes, a back-end processing module, a client module, and a cloud data storage device. The camera nodes connect to and manage several camera sensors, and are used to execute the aforementioned visual perception-enhanced instrument pointer reading recognition method.

[0030] Preferably, the visual perception-enhanced instrument pointer reading recognition system further includes a WAPI gateway and a WAPI verification server; the WAPI gateway and the multiple camera nodes form a star-shaped self-organizing network, the back-end processing module is communicatively connected to the WAPI gateway and the client module respectively, the client module is communicatively connected to the cloud data storage, and the WAPI verification server is used to provide user and terminal authentication keys as well as key management and data encryption protection functions for transmission by each module.

[0031] Preferably, the camera node includes a WAPI module and an OLED display screen. The camera node sends ACK request key information to the WAPI verification server, sends network access request and image encoding data to the WAPI gateway through the WAPI module, and simultaneously parses the instructions sent by the WAPI gateway. The OLED screen is used to display the information status of the camera node.

[0032] The beneficial effects of the visual perception enhancement meter pointer reading recognition method and system of the present invention are as follows: (1) First, the light intensity factor of the two-dimensional Gamma function is reconstructed to adaptively correct the light coefficient of different intensities; (2) The proposed elliptical model adopts multi-scale feature extraction, multi-level attention and multi-hop connection, which helps the model recover more details and reduce artifacts in the deblurred image; at the same time, the dehazing part can restore the clear image from the haze interference image without constructing an optical scattering model, so as to ensure the independence of the optical correction and dehazing process; (3) The LSD of the centripetal constraint mechanism is introduced to achieve more accurate pointer positioning, and the function between the pointer deflection angle and the meter reading is derived to achieve automatic and accurate meter reading; (4) The WAPI gateway of the present invention can realize the automatic networking of the dial reading nodes, automatically collect the image frames of the nodes, and decode and convert the images. The processing results are transmitted to the client through the back-end processor module. The client stores and manages the data in a unified manner, which can realize long-distance wireless communication and ultra-low power consumption, improve the stability of the wireless communication system, simplify the communication process in production, reduce network construction time and save operating costs. [Attached Image Description]

[0033] Figure 1 This is a step-by-step diagram of a visual perception-enhanced instrument pointer reading recognition method.

[0034] Figure 2 This is a schematic diagram of the illumination-reflection imaging model referenced in this invention.

[0035] Figure 3 This is a schematic diagram of the enhanced result of the reconstruction of the two-dimensional gamma function correction according to the present invention.

[0036] Figure 4 This is a schematic diagram of the elliptical network model for haze removal in this invention.

[0037] Figure 5 This is a schematic diagram showing the comparison of the standard deviation of image enhancement for all low-light images in the dataset using standard gamma, MSR, and the algorithm of this embodiment.

[0038] Figure 6 This is a schematic diagram showing the comparison of the standard deviation of image enhancement for all high-illumination images in the dataset using standard gamma, MSR, and the algorithm of this embodiment.

[0039] Figure 7 This is a loss function curve used to verify the haze removal effect of the OVAL-NET network.

[0040] Figure 8 This is an architecture diagram of an SF6 instrument pointer reading recognition system with enhanced visual perception.

Detailed Implementation Methods

[0041] The features and exemplary embodiments of various aspects of the present invention will now be described in detail. In the following detailed description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that the invention may be practiced without requiring some of these specific details. The following description of embodiments is merely intended to provide a better understanding of the invention by illustrating examples of the invention. The invention is by no means limited to any specific configurations and algorithms presented below, but covers any modifications, substitutions, and improvements to elements, components, and algorithms without departing from the inventive concept. In the accompanying drawings and the following description, well-known structures and techniques are not shown in order to avoid unnecessarily obscuring the invention.

[0042] Example 1

[0043] This embodiment implements a visual perception-enhanced method for recognizing pointer readings on SF6 instruments.

[0044] Figure 1 This is a step-by-step diagram illustrating a visual perception-enhanced instrument pointer reading recognition method. (For example...) Figure 1As shown, the method in this embodiment includes: S1, acquiring and transmitting an image of the pointer instrument panel; S2, extracting the illumination intensity component of the acquired pointer instrument panel image using a multi-scale Gaussian kernel, and reconstructing a two-dimensional Gamma function to adaptively correct the illumination intensity of the pointer instrument panel image; S3, inputting the illumination intensity-corrected pointer instrument panel image into a pre-trained dehazing elliptical network model to remove haze interference and restore a clear pointer instrument panel image; S4, adding centripetal force constraints to the line segment detector to achieve precise pointer positioning in a clear pointer instrument panel image, and deriving the functional relationship between the pointer angle and the reading to achieve accurate reading.

[0045] Taking the reading of SF6 meters and the construction of an information management system for meter reading as an example, this embodiment of a visual perception-enhanced SF6 meter pointer reading recognition method includes the following steps:

[0046] Step 1: Acquire and transmit SF6 barometer images, which are transmitted from the camera node to the backend processing module via the WAPI gateway;

[0047] Step 2: The illumination correction submodule in the backend processing module extracts the illumination intensity components of the acquired dial image using a multi-scale Gaussian kernel and reconstructs a two-dimensional Gamma function to adaptively correct the illumination intensity of the instrument image.

[0048] Step 3: The dehazing submodule in the backend processing module introduces a branch elliptic network (oval-net) to input the light-corrected dial image into the trained dehazing elliptic network model to restore a clear image from the haze interference image.

[0049] Step 4: The pointer positioning submodule in the backend processing module adds centripetal force constraints to the line segment detector to achieve accurate pointer positioning after the haze removal effect image, and derives the functional relationship between pointer angle and reading to achieve accurate reading;

[0050] Step 5: The backend processing module transmits the reading results and the results processed by each submodule to the client. The client uploads the results processed by each submodule in steps 2-4 to the cloud server and stores them, while providing a unified query and management page for the results.

[0051] Furthermore, the working mode of the camera node in step 1 is as follows:

[0052] (1) After the camera node is powered on, the camera (sensor) node is initialized;

[0053] (2) After the node is initialized, the node generates a random network entry delay. After the delay time is reached, it first sends an ack request key information to the WAPI authentication server, and then sends a network entry request to the WAPI gateway. If no response is received during this period, the network entry request will continue to be sent after a delay.

[0054] (3) After receiving the response frame from the WAPI gateway, the node joins the network and sets its own time.

[0055] (4) When the time slice of the camera (sensor) node is reached, the camera node acquires the instrument image, encodes the image information, and sends it to the WAPI gateway through the WAPI module.

[0056] Furthermore, in step 1, the WAPI gateway is deployed in the control center. The WAPI gateway includes a main control MCU module, a WAPI module, and an OLED display. The WAPI gateway sends ACK request key information to the WAPI authentication server through the WAPI module, receives information sent by camera (sensor) nodes, decodes and parses it, identifies newly joined camera nodes and camera (sensor) nodes that have reconnected after power failure, and performs protocol conversion on the data packets. The WAPI module uploads the collected raw dial image frames to the client and cloud data storage through the backend processor. The data packets include network access request data packets and sensor image frame packets.

[0057] Furthermore, in step 1, the working mode of the WAPI gateway is as follows:

[0058] (1) After the gateway is powered on, initialize the gateway;

[0059] (2) After initialization is complete, the WAPI gateway begins to attempt to connect to the authentication server and sends an ack request key information to the WAPI authentication server.

[0060] (3) After successfully connecting to the network, it automatically connects to the backend processor module and subscribes to topics for communication with the client;

[0061] (4) After connecting to the backend processor module, the WAPI gateway begins to enter the node information listening mode;

[0062] (5) When the WAPI module receives an image frame, the WAPI gateway will perform CRC verification on the frame. After the CRC verification is correct, the data frame type is determined according to the frame header. If the frame information is a network access request, the camera node is added to the network, an ACK response frame is sent to the camera node, a time slice and ID are allocated to the node, and the clocks of all nodes are synchronized. If the data frame is image encoding information sent by the camera node, the frame is decoded, a JSON data packet is formed, and the WAPI module sends it to the backend processing module.

[0063] (6) When the time synchronization time is reached, time synchronization data is sent to synchronize the time across the entire network.

[0064] The network is established by connecting to each node through a WAPI gateway. The collected meter reading images are transmitted to the backend processor module. First, the illumination correction submodule performs illumination correction processing on the original collected SF6 barometer images.

[0065] Figure 2 This is a schematic diagram of the illumination-reflection imaging model referenced in this invention. (See diagram below.) Figure 2 As shown, according to the imaging principle, the brightness value of an image can be written in the following form:

[0066] F(x,y)=I(x,y)·R(x,y)

[0067] Where F(x,y) represents the brightness value of the image at coordinates (x,y), I(x,y) is the illumination intensity component, and R(x,y) is the reflection component.

[0068] The illumination component represents the low-frequency characteristics of an image, determining the dynamic range of a pixel. The reflection component represents the high-frequency characteristics of the image (details), determined by the dial itself. The goal of adaptive illumination correction is to minimize the influence of the illumination component while preserving the reflection component. The illumination component expression can be rewritten as:

[0069] I(x,y)=F(x,y)*G(x,y)

[0070] Where * denotes convolution operation, and G(x,y) is a Gaussian function.

[0071] Therefore, the processing method of the illumination correction submodule in step 2 further includes the following:

[0072] (1) Construct a multi-scale Gaussian model and use it as the illumination component in the original acquired image:

[0073]

[0074]

[0075] Where i (i = 1, 2, ..., N) is the number of Gaussian functions at different scales, and wi represents the weight of the illumination intensity component of the i-th Gaussian function. To balance extraction accuracy and time, this parameter is set to N = 3, c1 = 15, c2 = 80, c3 = 250 and w1 = w2 = w3 = 1 / 3.

[0076] (2) Construct a two-dimensional Gamma model as a correction model for the illumination component:

[0077]

[0078]

[0079] Where O(x, y) represents the corrected illumination value, r represents the illumination correction parameter, b represents the illumination coefficient, and m (m∈[0, 255]) represents the average value of the illumination components.

[0080] (3) Set the parameters and output the corrected image result. Set b = m / 255. Figure 3 This is a schematic diagram illustrating the enhanced results of the reconstruction of the two-dimensional gamma function correction according to the present invention. (See diagram below.) Figure 3 As shown, the value of r is not fixed, but varies with the illumination component of each pixel. It typically ranges from 0 to 255.

[0081] Figure 4 This is a schematic diagram of the elliptical network model for haze removal in this invention. (See diagram below.) Figure 4 As shown, the dehazing submodule further includes a dehazing elliptical network model, which has three parts: an extraction module for extracting low-frequency features, a multi-scale attention network for learning haze features, and a multi-feature fusion module for reconstructing haze-free images.

[0082] (1) The extraction module for extracting low-frequency features mainly includes two ReLU activation layers and two convolutional layers;

[0083] (2) The multi-scale attention network consists of two parts: a channel attention subnetwork and a spatial attention subnetwork. The structures of the two subnetworks (channel attention subnetwork and spatial attention subnetwork) are as follows:

[0084] The first layer is the combined convolutional module A, which consists of four ReLU activation layers, one Sigmad activation layer, one pooling layer, one element-wise concatenation layer, one attention mechanism layer, and five convolutional layers.

[0085] The second layer is the combined convolutional module B, which consists of four ReLU activation layers, one pooling layer, one element-wise concatenation layer, one attention mechanism layer, one transposed convolutional layer, and six convolutional layers.

[0086] The third layer is the combined convolutional module C, which consists of four ReLU activation layers, one pooling layer, one element-wise concatenation layer, one attention mechanism layer, one transposed convolutional layer, and six convolutional layers.

[0087] The fourth layer is the combined convolutional module D, which consists of four ReLU activation layers, one pooling layer, one element-wise concatenation layer, one attention mechanism layer, one transposed convolutional layer, and six convolutional layers.

[0088] The fifth layer is the combined convolutional module E, which consists of two ReLU activation layers, one transposed convolutional layer, and three convolutional layers.

[0089] (3) The multi-feature fusion module structure for reconstructing haze-free images consists of two convolutional layers, an element stacking layer, and a sigmoid activation layer.

[0090] Secondly, the smog removal submodule in this implementation is loaded with a pre-trained offline network model. The offline training process can be as follows:

[0091] (1) Construction and preprocessing of the dial interference dataset;

[0092] This involves collecting pointer instrument image data in different scenarios using image acquisition equipment to construct a raw dataset, then removing interfering data that affects neural network training and recognition, including data with extreme angles and missing dials, and manually adding smog effects to the remaining data, as well as labeling the instrument dial categories.

[0093] (2) Based on the characteristics of the identification objects and application scenarios, construct a haze removal elliptical structure neural network model;

[0094] (3) Load the training parameters into the designed neural network model and train it. After training, the dial desmog model is obtained.

[0095] To verify the dehazing effect of the elliptic network, this embodiment uses a training dataset consisting of pairs of hazy-clear images. Hazy-interferenced images serve as input, and the corresponding clear images as label outputs. This embodiment uses 12,000 image pairs for training and testing the model, all with a size adjusted to 256*256*3. The hardware configuration is an Intel Xeon @ 2.30GHz workstation with 12GB RAM. After experiments and multiple adjustments to different hyperparameters, this invention selects a learning rate of 0.0001, 50 epochs, a batch size of 16, and the Adam optimizer for training. Since the human visual system can easily identify structural differences between two images, this embodiment selects the Structural Similarity Index (SSIM) as the loss function for updating weights.

[0096] (4) Finally, set the training completion flag, set the validation set interval to detect the training accuracy, and set the training completion flag to reach the maximum number of iterations and meet the accuracy requirements. After training is completed, save the model structure and parameters.

[0097] After loading the trained elliptic network model, the dial image output by the illumination correction submodule is processed to obtain a clear image free of haze. This image is then input into the pointer positioning submodule.

[0098] Furthermore, the processing method of the pointer positioning submodule in step 4 includes the following: This embodiment introduces a centripetal constraint mechanism and combines it with a line-segment-detection LSD to achieve accurate pointer positioning and derive the relationship between the pointer deflection angle and the meter reading.

[0099] Pointer positioning detection is a key factor affecting instrument readings. Based on the characteristic that the pointer rotates along the dial's axis, a centripetal constraint is added to the LSD to accurately position the pointer. The first-order gradient components of each pixel in four directions are calculated according to methods in relevant literature. The gradient formulas are as follows:

[0100]

[0101] Among them, is g x (x,y) is the first-order gradient component in the x-direction, g y (x,y) is the first-order gradient component in the y-direction, g 45 (x,y) is the first-order gradient component in the 45° direction, g 135 (x,y) is the first-order gradient component in the 135° direction.

[0102] Based on the characteristics of the pointer in the SF6 barometer, adding a centroid constraint in the LSD (Low Scaling) where the horizontal angle is close to the angle relative to the image center can reduce the pointer positioning calculation time and improve pointer positioning accuracy. The expression for the centroid constraint mechanism is:

[0103]

[0104] Where (xo, yo) are the coordinates of the image center point, θ(xi, yi) is the angle between the pixel (xi, yi) and the x-axis, and th1 is the set threshold. th1 is used to filter feature points in candidate regions. Without affecting the result, the smaller the value of th1, the shorter the computation time. The expression for θ(xi, yi) is as follows:

[0105]

[0106] The expression for region growth after adding centripetal constraint prior conditions during feature region segmentation is as follows:

[0107]

[0108] Where I is the set of all pixels in the image, and th2 is a set threshold to remove the uncertainty of small gradient values.

[0109] In the pointer table, the shortest pointer is longer than the longest tick mark. Based on this property, the longest segment in the divided set of pointers and tick marks is the pointer itself.

[0110] Furthermore, the derivation of the functional relationship between the pointer angle and the reading in step 4 enables accurate reading.

[0111] The calculation method for pointer instrument readings ultimately boils down to calculating the relationship between the pointer angle and the dial scale. This embodiment divides the dial into four quadrants. Based on the relationship between each quadrant and the pointer scale, the pointer scale value is calculated using the formula for each quadrant. Assuming the pointer scale range is G, and the corresponding angle range is θ0, with the center point of the dial as the center point, the left scale value x0 where the dial intersects the X-axis is used as a configuration parameter for input by engineering personnel.

[0112] When the pointer is identified in the first quadrant, the dial scale is calculated as follows:

[0113] x0+G*(180°+θ) / θ0

[0114] When the pointer is in the second quadrant, the dial scale is calculated as follows:

[0115] x0+G*θ / θ0

[0116] When the pointer is in the third quadrant, the dial scale is calculated as follows:

[0117] x0-G*θ / θ0

[0118] When the recognized pointer is in the fourth quadrant, the dial scale is calculated as follows:

[0119] x0-G*(180°+θ) / θ0

[0120] Furthermore, in step 5, the client module, cloud data storage, and WAPI verification server, from an operator's perspective, primarily utilize the client's front-end page for step-by-step analysis and processing. The client is built using the Vue framework, comprising a homepage module, an image processing layered module, an API interface module, and a backend management module. The homepage module displays the latest uploaded data from each camera node in real time. The image processing layered module displays historical uploaded data from nodes and provides separate query and filtering functions for image frames from different nodes. The API interface module receives JSON data packets transmitted from the backend processor module. The backend management module manages the image results processed by different sub-modules. The cloud storage uses Huawei Cloud's backend as the data repository for interaction and access with the client. The WAPI verification server manages the system's network access authentication requirements, providing user and terminal authentication keys, key management for transmissions from various modules, and data encryption protection.

[0121] To verify the validity of the results, on-site analysis and application were conducted.

[0122] To verify whether the reconstructed 2D gamma function can perform adaptive illumination correction, images from an sf6 barometer under both low and high illumination conditions were selected in this embodiment. Adaptive illumination correction was compared between the standard 2D gamma function, the multi-scale Retinex (MSR), and the improved 2D gamma function proposed in this embodiment.

[0123] By comparing (a) the original image, (b) the output image of the standard 2D gamma function, (c) the output image of the MSR function, and (d) the output image of the improved 2D gamma function, it was found that in low-light conditions, the standard 2D gamma function improves the overall brightness, but some details in low-light areas are not preserved, artifacts are introduced, and some noise appears. The MSR algorithm has a poor correction effect; the overall image tone is greater than the original image, the image sharpness is not ideal, and direct nonlinear operations on pixels produce ghosting and blurring effects. The algorithm in this embodiment improves the brightness in low-light conditions, enhances the image sharpness and contrast, and significantly improves the image quality in low-light conditions. In high-light-intensity conditions, the standard gamma function reduces the light intensity, but a slight color cast occurs. Both the MSR algorithm and the algorithm in this embodiment can reduce the light intensity of the image and have better color preservation, but the image processed by this embodiment is sharper than that processed by the MSR algorithm.

[0124] Figure 5 This is a schematic diagram showing the comparison of the standard deviation of image enhancement for all low-light images in the dataset using standard gamma, MSR, and the algorithm of this embodiment. Figure 6This diagram illustrates the comparison of image enhancement standard deviations for all high-illumination images in the dataset using standard gamma, MSR, and the algorithm of this embodiment. Figure 5 , Figure 6 As shown, image enhancement was performed on all low-light and high-light images in the dataset using standard gamma, MSR, and the algorithm of this embodiment, respectively, and their various indices were calculated. Due to space limitations, only 10 samples are given. It can be seen that the standard deviation of the enhanced images obtained after processing by each algorithm was improved compared with that before processing. The improvement effect of the algorithm of this embodiment is better than the other two algorithms, and it is closer to the evaluation indices of images with suitable illumination.

[0125] Figure 7 This is a loss function curve used to verify the haze removal effect of the OVAL-NET network. (Example) Figure 7 As shown in the figure, the effect of OVAL-NET network in removing haze was verified. The structural similarity index (SSIM) was selected as the loss function for updating weights, and the loss function curve is shown in the figure.

[0126] To achieve engineering applications and rapid computation capabilities, this embodiment directly uses on-site images with fog interference to verify the effect. The images were primarily collected from a substation in Shanghai during rainy weather and from a substation in Beijing under foggy conditions. By comparing (a) the original image, (b) the HAZERD result image, (c) the DenseHAZE result image, and (d) the result image from this embodiment, it can be seen that the method in this embodiment has a better subjective visual effect in removing fog. HAZERD failed to remove fog in the actual environment and also exhibited artifacts at the edges. The DenseHZE method can achieve some visual improvement in areas with mild fog, but overall it fails to achieve the desired effect.

[0127] The pointer recognition reading method in this embodiment was compared with DC-HT, Retinex, and Adaptive Bilateral Filtering (RABF). To ensure a fair comparison, all ten selected images were corrected images. To verify the accuracy of the readings, the mean relative error (ARE) was used as the evaluation metric, and the comparison results are shown in the table below.

[0128] Table 1. Pointer positioning results of different algorithms: (a) RABF, (b) DC-HT, (c) CLSD.

[0129] Manual meter reading RABF DCHT CLSD 1 1.430 1.374 1.323 1.422 2 0.56 0.594 0.678 0.567 3 0.87 0.881 0.826 0.858 4 11.450 11.321 11.015 11.427 5 7.900 7.778 7.655 7.915 ARE 2.74％ 4.45％ 1.29％

[0130] It can be seen that the RABF method is prone to mistaking horizontal lines for pointers when identifying table pointers, while the DCHT method does not accurately identify pointers. The method in this implementation can effectively extract the centripetal pointer line.

[0131] The ARE of this embodiment is 3.49% higher than that of DC-HT, and it is also superior to that of RABF. This indicates that the method of this embodiment can read meters more accurately compared to other methods. In addition, the ARE of this meter reading method is 1.17% in 50 test sets, which also verifies that it has good universality to a certain extent.

[0132] Example 2

[0133] This embodiment implements a visual perception-enhanced SF6 instrument pointer reading recognition system.

[0134] WAPI (Wireless LAN Authentication and Privacy Infrastructure) is a wireless LAN authentication and privacy infrastructure, a security protocol, and also a mandatory wireless LAN security standard in China. It was first proposed by the State Key Laboratory of Integrated Services Network Theory and Key Technologies at Xi'an University of Electronic Science and Technology.

[0135] Figure 8 This is an architecture diagram of an SF6 instrument pointer reading recognition system with enhanced visual perception. (For example...) Figure 8 As shown, the system in this embodiment includes a camera group composed of camera nodes, a WAPI gateway, a backend processing module, a client module, a cloud data storage (server), and a WAPI verification server. The WAPI gateway forms a star-shaped self-organizing network with multiple camera (sensor) nodes and transmits data via a proprietary WAPI protocol and image frame format. The backend processing module communicates with both the WAPI gateway and the client module. The backend processing module mainly includes several sub-modules embedded with the method of Embodiment 1: a lighting correction sub-module, a dehazing sub-module, and a pointer positioning and reading sub-module. The client module communicates with the cloud database (cloud data storage), collects and encodes information from images of different camera nodes through the WAPI gateway, packages it into a unified format, sends it to the backend processing module, and then uploads it to the client module. The client module stores the data in the cloud database (cloud data storage) and manages it. The WAPI verification server manages the network access authentication requirements of this embodiment system, providing user and terminal authentication keys, key management for transmissions between modules, and data encryption protection.

[0136] Furthermore, the camera node includes a main control MCU module, a WAPI module, a CMOS module, and an OLED display screen. The camera node sends ACK request key information to the WAPI verification server through the WAPI module, sends network access request and image encoding data to the WAPI gateway, and parses the instructions sent by the WAPI gateway. The OLED screen is used to display the information status of the camera (sensor) node.

[0137] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM).

[0138] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several improvements and additions without departing from the principle of the present invention, and these improvements and additions should also be considered within the scope of protection of the present invention.

Claims

1. A method for recognizing instrument pointer readings with enhanced visual perception, characterized in that... Includes the following steps: S1. Acquire and transmit the image of the pointer instrument panel; S2. Extract the illumination intensity component of the acquired pointer instrument dial image using a multi-scale Gaussian kernel, and reconstruct a two-dimensional Gamma function to adaptively correct the illumination intensity of the pointer instrument dial image. S3. Input the pointer instrument dial image after light intensity correction into the trained dehazing elliptical network model to remove haze interference and restore a clear pointer instrument dial image. S4. Add centripetal force constraints to the line segment detector to achieve clear pointer instrument dial image and accurate pointer positioning, and derive the functional relationship between pointer angle and reading to achieve accurate reading.

2. The method for recognizing instrument pointer readings with enhanced visual perception according to claim 1, characterized in that... Step S2 includes the following sub-steps: S21. Construct a multi-scale Gaussian model and use it as the illumination component of the original acquired pointer instrument dial image. Where G(x,y) is a Gaussian function, F(x,y) represents the brightness value of the image at coordinate (x,y), I(x,y) represents the illumination intensity component at coordinate (x,y), * represents convolution operation, i = 1, 2, ..., N are the number of Gaussian functions of different scales, and wi represents the illumination intensity component weight of the i-th Gaussian function. S22. Construct a two-dimensional Gamma function model as a correction model for the illumination components of an instrument dial image. Where O(x, y) represents the illumination value after the pointer instrument dial image is corrected, r represents the illumination correction parameter, b represents the illumination coefficient, and m (m∈[0, 255]) represents the average value of the illumination components; S23. Set the r, b, and m parameters, and output the pointer instrument dial image after light intensity correction.

3. The method for recognizing instrument pointer readings with enhanced visual perception according to claim 2, characterized in that: The haze removal elliptic network model includes an extraction module for extracting low-frequency features, a multi-scale attention network for learning haze features, and a multi-feature fusion module for reconstructing haze-free images. The extraction module includes a ReLU activation layer and a convolutional layer. The multi-scale attention network includes a channel attention subnetwork and a spatial attention subnetwork. The multi-feature fusion module includes a convolutional layer, an element stacking layer, and a Sigmad activation layer.

4. The method for visual perception enhancement of instrument pointer reading recognition according to claim 3, characterized in that... Step S3, the training of the dehazing elliptic network model, includes the sub-steps of constructing and preprocessing the pointer instrument dial image interference dataset: collecting pointer instrument dial image data under different scenarios through image acquisition devices to construct the original dataset, then removing the interference data that affects the training and recognition of the neural network, manually adding haze effects to the remaining data, and labeling the category of the pointer instrument dial.

5. The method for visual perception enhancement of instrument pointer reading recognition according to claim 2, characterized in that... Step S4, precise pointer positioning, includes the following sub-steps: S41. The first-order gradient components of each pixel (x, y) in the four directions of the clear pointer dial image were calculated. Among them, is g x (x,y) is the first-order gradient component in the x-direction, g y (x,y) is the first-order gradient component in the y-direction, g 45 (x,y) is the first-order gradient component in the 45° direction, g 135 (x,y) is the first-order gradient component in the 135° direction; S42. In the line segment detector, a centroid constraint is added where the angle of the horizontal line is close to the angle of the center of the clear pointer instrument dial image. This reduces the pointer positioning calculation time and improves the pointer positioning accuracy. Centroid constraint mechanism expression. Where (xo, yo) are the coordinates of the center point of the clear pointer instrument panel image, θ(xi, yi) is the angle between the pixel (xi, yi) and the x-axis, and th1 is a set threshold used to filter feature points in the candidate region. Without affecting the result, the smaller the value of th1, the shorter the computation time. The expression for θ(xi, yi) is as follows: S43. The expression for region growth after adding centripetal constraint prior conditions during the segmentation of feature regions in a clear pointer instrument dial image is as follows: Where I is the set of all pixels in the clear pointer instrument dial image, th2 is the set threshold to remove the uncertainty of small gradient value points; the longest line segment in the segmented pointer and scale segment set is the pointer, realizing precise pointer positioning.

6. The method for recognizing instrument pointer readings with enhanced visual perception according to claim 5, characterized in that... Step S4: Divide the pointer instrument dial into four quadrants and calculate the relationship between the pointer angle and the dial scale for accurate reading. When the pointer is identified in the first quadrant, the pointer instrument dial scale reading is x0 + G*(180° + θ) / θ0; when the pointer is identified in the second quadrant, the pointer instrument dial scale reading is x0 + G*θ / θ0; when the pointer is identified in the third quadrant, the pointer instrument dial scale reading is x0 - G*θ / θ0; when the pointer is identified in the fourth quadrant, the pointer instrument dial scale reading is x0 - G*(180° + θ) / θ0. Here, the pointer scale range is G, and the angle range corresponding to the pointer scale range is θ0. Taking the center point of the dial as the center point, the left scale value x0 where the dial intersects the X-axis is used as the configuration parameter.

7. The method for recognizing instrument pointer readings with enhanced visual perception according to claim 1, characterized in that... Includes the following steps: S1, The camera sensor collects and transmits images of the pointer instrument dial, which are then transmitted from the camera node to the back-end processing module via the WAPI gateway. S2, the back-end processing module's illumination correction submodule extracts the illumination intensity components of the acquired pointer instrument dial image using a multi-scale Gaussian kernel, and reconstructs a two-dimensional Gamma function to adaptively correct the illumination intensity of the pointer instrument dial image. S3, the backend processing module's dehazing submodule inputs the light intensity-corrected pointer instrument dial image into the pre-trained dehazing elliptical network model to remove haze interference and restore a clear pointer instrument dial image; S4. The pointer positioning submodule of the back-end processing module adds centripetal force constraints to the line segment detector to achieve accurate pointer positioning of the pointer in the clear pointer instrument dial image, and derives the functional relationship between pointer angle and reading to achieve accurate reading. S5. The backend processing module transmits the reading results and the processing results of the illumination correction submodule, the dehazing submodule, and the pointer positioning submodule to the client. The client uploads the step processing results to the cloud data storage and provides a unified query and management page for the results.

8. A visual perception-enhanced instrument pointer reading recognition system, characterized in that: It includes several camera nodes, a backend processing module, a client module, and a cloud data storage device. The camera nodes are connected to and manage several camera sensors. It is used to execute a visual perception enhancement instrument pointer reading recognition method as described in any one of claims 1 to 7.

9. A visual perception-enhanced instrument pointer reading recognition system according to claim 8, characterized in that: It also includes a WAPI gateway and a WAPI verification server; the WAPI gateway and the multiple camera nodes form a star-shaped self-organizing network, the back-end processing module is communicatively connected to the WAPI gateway and the client module respectively, the client module is communicatively connected to the cloud data storage, and the WAPI verification server is used to provide user and terminal authentication keys as well as key management and data encryption protection functions for transmission by each module.

10. A visual perception-enhanced instrument pointer reading recognition system according to claim 9, characterized in that: The camera node includes a WAPI module and an OLED display. The camera node sends ACK request key information to the WAPI verification server, sends network access request and image encoding data to the WAPI gateway through the WAPI module, and parses the instructions sent by the WAPI gateway. The OLED screen is used to display the information status of the camera node.

Images