Local denoising method and device for oil level image of edge side transformer

Abstract

The invention relates to the technical field of image processing, solves the problems that in the prior art, a wide wireless channel is needed for data transmission of a background computer, and the requirement for the operation performance of the background computer is high, and discloses an edge side transformer oil level image local denoising method and device, and the method comprises the steps: obtaining an original image of a transformer oil level; constructing a BN neural network model; carrying out HSI spatial transformation on the original image to independently extract a brightness component Ix; performing convolution processing and nonlinear mapping on the brightness component Ix by using the BN neural network model, and performing feature learning to obtain residual brightness F (IX); superposing the brightness component IX and the residual brightness F (IX) to obtain a reconstructed image; according to the method, image preprocessing is locally completed, so that the processing efficiency is relatively high, the original data volume is reduced, and the data bandwidth overhead and cost are saved.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to a method and apparatus for local denoising of edge-side transformer oil level images. Background Technology

[0002] During the booming development of the power industry, many power plants were built, and the safe operation of these plants is extremely important. As one of the essential pieces of equipment required for power production within a power plant, the safe and stable operation of power transformers is also crucial for the safety of the entire power grid.

[0003] Oil-immersed power transformers are a particularly common type of transformer. Their interiors are filled with a certain volume of insulating oil, which keeps the internal structure in an insulated state, protecting it and preventing damage from overheating or electric arcing, effectively extending the transformer's service life. Generally, the load and ambient temperature affect the temperature of the transformer's internal equipment, and these internal temperature changes affect the oil level: if the equipment temperature rises, the oil level rises; if the equipment temperature falls, the oil level falls. Besides temperature, water ingress will also raise the oil level, and oil leakage will lower it. Therefore, the transformer oil level is a dynamic indicator that needs to remain within certain limits. Exceeding the maximum value increases the internal oil pressure, potentially causing oil overflow; falling below the minimum value leads to insufficient oil and increased contact with air, reducing insulation performance. Both factors shorten the transformer's lifespan and, in severe cases, can even cause power grid failure, resulting in personal injury and significant economic losses. Over the years, oil leakage incidents in power transformers have occurred frequently, causing serious consequences. This has also served as a warning that it is essential to conduct efficient detection of transformer oil levels in a timely manner to identify any abnormalities and to replenish or drain the oil as needed, thus keeping the transformer oil level consistently within the specified range.

[0004] Previously, the oil level of power transformers was mostly checked manually. Staff would periodically visit the transformer site to visually inspect the oil level and see if it was within the acceptable range. However, this method has certain drawbacks: firstly, it is labor-intensive and requires high manpower costs; secondly, the accuracy is low, mainly because it may not be able to detect problems in a timely manner, leading to missed detections, and because human observation has a certain degree of error, abnormal oil levels may not be accurately detected, resulting in false positives.

[0005] In recent years, with the rapid development of video imaging technology, intelligent monitoring technology based on machine vision has also developed rapidly. The previously laborious, time-consuming, and high-false-detection-rate detection methods have gradually become obsolete. Instead, the approach involves first acquiring image data through video surveillance technology, and then using image processing technology to perform depth processing and recognition on the images to separate the target objects to be detected. In fact, such technology is already commonplace in public places such as banks, supermarkets, road traffic, and airports, and it also plays a significant role in fields such as medicine and the military.

[0006] However, during the process of capturing and automatically identifying outdoor power equipment such as transformer oil levels and ring main unit meters using image vision technology, environmental interference such as rain, snow, and dust can easily occur. This interference can obstruct or distort the background content of the image or cause atmospheric occlusion, resulting in blurring and loss of detail. Therefore, effective preprocessing and denoising of the image before deep image recognition enhances the effectiveness of various image algorithms and is essential for increasing the accuracy of subsequent image recognition. Currently, denoising for transformer oil level images is mainly divided into two categories: video image denoising and single-frame image denoising. Compared to video image denoising, which has continuous inter-frame information as a reference, single-frame image denoising can only recover from degradation caused by the aggregation of background noise lines due to rain, snow, dust, etc., making it more challenging. Current single-frame image denoising algorithms can be mainly divided into two types: single-frame image denoising algorithms based on prior image knowledge, such as wavelet transform, and single-frame image denoising algorithms based on deep learning, such as the RESCAN algorithm. However, both wavelet transform denoising algorithms and deep learning denoising algorithms currently require computation on computers with powerful hardware computing capabilities, which places high demands on hardware resources.

[0007] Therefore, the common method is to capture the image of the oil level transformer locally using a front-facing camera, and then send the original image directly to the backend via broadband wired fiber optic or wireless 4G / 5G networks. The backend computer then performs image processing and recognition. The advantage of this method is that it preserves the original data, resulting in high processing accuracy for the backend computer. The disadvantages are that it requires a wide wireless channel, leading to high mobile data usage, and it also suffers from large communication delays. For some time-sensitive transformer anomalies, it cannot provide timely feedback. In addition, this method has high requirements for the computing performance of the backend computer, and the cost is relatively high. Summary of the Invention

[0008] The purpose of this application is to overcome the problems in the prior art that data transmission to the backend computer requires a wide wireless channel and high computing performance of the backend computer, and to provide a method and apparatus for local denoising of oil level images of edge-side transformers.

[0009] Firstly, a local denoising method for edge-side transformer oil level images is provided, including:

[0010] Obtain the raw image of the transformer oil level;

[0011] Construct a BN neural network model, where BN stands for Batch Normalization, which is a standardization process for the original data.

[0012] The luminance component I is extracted separately by performing HSI spatial transformation on the original image. x ;

[0013] The luminance component I is analyzed using the BN neural network model. x Convolutional processing and nonlinear mapping are performed, followed by feature learning to obtain the residual brightness F(I). X );

[0014] Brightness component I X and residual brightness F(I) X The images are superimposed to obtain the reconstructed image;

[0015] The reconstructed image is then subjected to an inverse HSI transform and mapped to the RGB space to obtain a denoised image.

[0016] Furthermore, the BN neural network model comprises N layers, each layer containing a nonlinear transformation operation: H n =(*), and a BN layer is set after each nonlinear transformation operation, where n represents the nth layer in the network, H n = (*) represents a network action, which includes normalization, activation, convolution, or pooling. The output X of the nth layer is... n Represented as: X n =H n ([X0, X1, ..., X n-1 ]), where [X0, X1, ..., X n-1 [] represents the feature map output of the 0th, 1st, ..., n-1st layer.

[0017] Optionally, the BN neural network model can be optimized.

[0018] Furthermore, optimizing the BN neural network model includes:

[0019] Cancel the BN layer after each nonlinear transformation operation;

[0020] Add a BN layer after all nonlinear transformation operations;

[0021] An SE layer is added after the BN layer to enhance the dependency of the feature map in the multi-dimensional channels of the BN layer output image.

[0022] Furthermore, the residual luminance F(I) X )=(I Y -I X ), where I Y This represents the luminance component of the original image without background interference.

[0023] Furthermore, the luminance component I' of the reconstructed image Y =F(I X )+I X .

[0024] Secondly, a local denoising device for edge-side transformer oil level images is provided, comprising:

[0025] A storage unit stores program code for execution by a multi-core processor unit, the program code including steps for performing methods as described in any of the implementations of the first aspect;

[0026] The camera unit is used to acquire raw images of the transformer oil level;

[0027] A multi-core processor unit is used to execute the program code stored in the storage unit;

[0028] The communication unit is used to send the image feature values ​​preprocessed by the multi-core processor unit to the background computer in real time.

[0029] The camera unit is connected to the multi-core processor unit, the multi-core processor unit is connected to the communication unit, and the storage unit is connected to the multi-core processor unit.

[0030] It also includes a power supply unit for supplying power to the camera unit, the multi-core processor unit, and the communication unit, wherein the camera unit, the multi-core processor unit, and the communication unit are all electrically connected to the power supply unit.

[0031] Thirdly, a computer-readable storage medium is provided that stores program code for execution by a device, the program code including steps for performing a method as described in any of the implementations of the first aspect.

[0032] Fourthly, an electronic device is provided, the electronic device including a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the method as in any of the implementations of the first aspect.

[0033] This application has the following beneficial effects: It utilizes the edge computing capabilities of embedded multi-core processors to divide image processing into two parts in terms of topology: image acquisition and preprocessing are directly completed at the edge side image acquisition point, eliminating background interference in the original image while delineating the optimal feature values ​​of the image. Then, the feature values ​​are wirelessly reported to the back-end computer, which completes the subsequent recognition work. The advantage of this method is that image preprocessing is completed locally, which has high processing efficiency, reduces the amount of original data, and saves data bandwidth overhead and cost. It is an effective structural improvement to the current automatic transformer oil level identification method and has great market application value. Attached Figure Description

[0034] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an undue limitation of this application.

[0035] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 This is a flowchart of the local denoising method for the edge-side transformer oil level image in Embodiment 1 of this application;

[0037] Figure 2 This is a schematic diagram of the structure of the BN neural network model in the local denoising method for the oil level image of the edge-side transformer in Embodiment 1 of this application;

[0038] Figure 3 This is a schematic diagram of the denoising algorithm model after adding HSI spatial transformation in the local denoising method for edge-side transformer oil level images in Embodiment 1 of this application;

[0039] Figure 4 This is a schematic diagram of the mutual mapping between RGB space and HSI space in the local denoising method for the oil level image of the edge-side transformer in Embodiment 1 of this application;

[0040] Figure 5 This is a schematic diagram of the structure of the improved BN neural network model in the local denoising method for edge-side transformer oil level images in Embodiment 1 of this application;

[0041] Figure 6 This is a schematic diagram of the internal structure of the SE layer in the local denoising method for the edge-side transformer oil level image of Embodiment 1 of this application;

[0042] Figure 7This is a diagram of the original BN neural network model architecture in the local denoising method for edge-side transformer oil level images in Embodiment 1 of this application;

[0043] Figure 8 This is a diagram of the BN neural network model architecture after adding the SE module layer in the local denoising method for edge-side transformer oil level images in Embodiment 1 of this application;

[0044] Figure 9 This is a structural block diagram of the local noise reduction device for the edge-side transformer oil level image in Embodiment 2 of this application.

[0045] Figure label:

[0046] 100. Storage unit; 200. Camera unit; 300. Multi-core processor unit; 400. Communication unit; 500. Power supply unit. Detailed Implementation

[0047] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0048] Example 1

[0049] The method for local denoising of edge-side transformer oil level images according to Embodiment 1 of this application includes:

[0050] Specifically, Figure 1 A flowchart of the local denoising method for the edge-side transformer oil level image in Embodiment 1 of the application is shown, including:

[0051] S100: Obtain the original image of the transformer oil level;

[0052] For example, a high-definition digital CMOS wide-angle camera is used to shoot the image at a certain height, with a straight-line distance of 1.2-1.5 meters from the center of the transformer bushing and an allowable deviation of ±0.2 meters. The CMOS wide-angle camera is required to cover an angle of 0-80 degrees, which can cover the center bushing window of the transformer relatively well, thereby capturing the original image of the transformer oil level.

[0053] S200, Construct the BN neural network model;

[0054] Please see Figure 2The traditional neural network model for single-image denoising is a 5-layer model based on simplified Batch Normalization (BN). The input of each layer is the feature map output by the preceding convolutional network. Assuming the network contains N layers, each layer contains a nonlinear transformation H... n = (*), where n represents the nth layer in the network, H n = (*) represents various network operations, such as normalization, activation, convolution, or pooling. The output of the nth layer is denoted as X. n The nth layer receives feature maps from all previous layers, which can be represented as:

[0055] X n =H n ([X0, X1, ..., X n-1 (1)

[0056] In formula (1), [X0, X1, ..., X n-1 [] represents the feature map output of the 0th, 1st, ..., n-1st layer.

[0057] S300. Perform HSI spatial transformation on the original image to extract the luminance component I separately. x ;

[0058] exist Figure 2 In this context, for original images containing strong background noise such as rain, snow, and dust, the output feature map of the BN neural network model not only suffers from gradient problems but also from oversmoothing. If the original image is directly input into the BN neural network model, the background can easily cause color distortion. Therefore, this application proposes: before preprocessing a single original image into the BN network, perform a spatial transformation on the color and brightness of the single image, and use the RGB->HSI spatial transformation to extract the brightness component I of the original image. x By preprocessing the original image using HSI spatial transformation, a large amount of original image brightness information is obtained, reducing the probability of color distortion during image processing. A single-image rain removal algorithm model incorporating HSI spatial transformation is shown below. Figure 3 As shown.

[0059] S400, Using the BN neural network model to process the luminance component I x Convolutional processing and nonlinear mapping are performed, followed by feature learning to obtain the residual brightness F(I). X );

[0060] exist Figure 3 In this process, the input original image Im(X) is first subjected to HSI spatial transformation to extract the luminance component I. X Denoising is performed, and then a simplified BN neural network model is used to analyze the luminance component I. xConvolution processing and nonlinear mapping are performed, and its relationship with the luminance component I of the original image Im(Y) of the transformer oil level without background interference is learned. Y The residual component, i.e., F(I) X )=(I Y -I X Obviously, the learned residual brightness F(I) X Compared to learning I Y The entire network is easier to optimize and its efficiency is improved.

[0061] S500, luminance component I X and residual brightness F(I) X The images are superimposed to obtain the reconstructed image;

[0062] Specifically, the luminance component I X and residual brightness F(I) X By superimposing the components, the luminance component I' of the reconstructed image is obtained. Y The formula is shown in (2) below:

[0063] I' Y =F(I X )+I X (2)

[0064] Among them, I X For the luminance component, F(I) X () represents the residual brightness.

[0065] S600: Perform an inverse HSI transformation on the reconstructed image and map it to the RGB space to obtain a denoised image;

[0066] Specifically, by performing an inverse HSI transform on the reconstructed image and mapping it to the RGB space, we can obtain Im(Y) after removing the rain lines. The HSI space transformation maps the RGB color model to the HSI space, and the mapping between the two space variables is as follows: Figure 4 As shown.

[0067] exist Figure 4 In the diagram, (a) is the RGB model, where the X, Y, and Z axes of the Cartesian coordinate system represent the blue, red, and green color components, respectively. (b) is the HSI model, where the angle represents hue (H), the radius represents saturation (S), and the height represents brightness (I). Using the HSI color model to describe the colors of an image is more in line with the characteristics of human visual perception. Furthermore, considering that transformer oil level images mainly present brightness information in the presence of strong background interference (rain, snow, dust, etc.), removing the influence of rain lines from rainy images is essentially processing the brightness information in the rain image. The HSI space transformation can separate the brightness and chromaticity of the image. The conversion steps from RGB to HSI are as follows:

[0068] 1. First, normalize the RGB values.

[0069]

[0070] 2. Calculate the values ​​of H, S, and I.

[0071]

[0072] s = 1 - 3 × min(r, g, b), s ∈ [0, 1]

[0073] i=(R+G+B) / (3×255),i∈[0,1]

[0074] 3. The conversion from HSI color space to RGB color space is shown in the following formula.

[0075] x=i×(1-s) (5)

[0076] Among them, h=H×π / 180, s=S / 100, i=I / 255.

[0077] S700. Optimize the BN neural network model to reduce the amount of computation, so as to adapt to data computation under the condition of limited hardware resources on the edge side;

[0078] because Figure 2 In the BN neural network model constructed in this paper, each convolutional layer contains a BN layer, which has a normalization characteristic for data processing. Therefore, its presence weakens the absolute difference between the background of images with and without background interference. In addition, using the BN network in every convolution increases the consumption of computational resources and reduces the training efficiency of the network, making it unsuitable for model deployment under the current conditions of limited hardware resources on the edge side. Therefore, this application addresses... Figure 2 The BN neural network model constructed in the original paper was appropriately optimized by eliminating the BN layer after each operation and adding a BN training step after six convolutions, thereby reducing computational overhead. The specific structure of the improved BN neural network model is as follows: Figure 5 As shown.

[0079] exist Figure 2In this paper, assuming that the output of each convolutional block is k feature maps, the input of the nth layer will have k0 + k × (n-1) feature maps, where k0 is the number of channels in the input layer. Each input channel is trained using Batch Normalization (BN), resulting in a large number of inputs to higher layers and a significant increase in computational complexity. To reduce computational difficulty, it is necessary to compress the number of channels and reduce the number of BN layers. Specifically, after six convolution operations, one BN training operation is performed. The output channel count of the first five operations is set to be 1 / 4 of the input channel count. Finally, a 1 × 1 convolution is used to fuse all the output feature maps together. Furthermore, to improve the feature extraction capability and enhance the dependency of the BN output image feature maps across multiple channels, this application... Figure 5 To achieve the above objectives, an additional SE layer (i.e., a squeeze-excited SE module) is added. The internal structure of the SE layer is as follows: Figure 6 As shown.

[0080] in, Figure 6 In this diagram, X represents the original image, U represents the image reconstructed from RGB to HSI space, 1X1XC represents a 1×1 convolution from U, Fsq(.) represents the convolution operation, Fex(.,W) represents the weighted operation on each channel of the image, and Fscale(.,.) represents the deconvolution operation. This represents the final denoised image.

[0081] It should be noted that the SE layer can demonstrate the interdependencies between multi-dimensional channels and adaptively determine the importance of each channel, assigning different weights and recalibrating the feature responses. This emphasizes important feature information while suppressing less important features. With the SE module added, the entire BN neural network model can improve feature extraction capabilities by combining channel and spatial features. The original BN neural network model architecture is shown in Figure 7, and the model architecture after adding the SE module is as follows: Figure 8 As shown, X represents the original image. This represents the image after denoising.

[0082] Example 2

[0083] Please see Figure 9 The local noise reduction device for edge-side transformer oil level images according to Embodiment 2 of this application includes:

[0084] Storage unit 100 stores program code for execution by multi-core processor unit 300, the program code including steps for performing the method as in any of the implementations in Embodiment 1;

[0085] For example, storage unit 100 includes 8GB DDR4 SDRAM and 32GB eMMC; it mainly stores the operating system, data acquisition program, data processing program and intermediate image data, etc.

[0086] The camera unit 200 is used to acquire raw images of the transformer oil level;

[0087] For example, the camera unit 200 is a high-definition digital CMOS wide-angle camera, which is aimed at the center of the transformer bushing at a certain height and the straight-line distance from the center of the transformer bushing is between 1.2 and 1.5 meters, with an allowable deviation of ±0.2 meters. The CMOS wide-angle is required to cover an angle of 0-80 degrees, which can cover the center bushing window of the transformer relatively well.

[0088] The multi-core processor unit 300 is used to execute the program code stored in the storage unit 100;

[0089] For example, the multi-core processor unit 300 consists of an integrated ARM quad-core Cortex-A72 and a NEOP neural network convolution operation core. Image acquisition, task scheduling, data caching, and communication are performed by the Cortex-A72 quad-core, while the NEOP neural network core is mainly responsible for image preprocessing, i.e., the BP neural network model proposed in this application, including computation, training, and feedback. The core also integrates a 2MB high-speed cache for direct instruction-level reading and storage of convolution operation data. The cache communicates with external DDR4 memory through a dual-channel DMA burst mode. The NEOP core can process up to 2 TOPS of computation speed, which basically meets the requirements of the improved single-image denoising algorithm of the neural network model under resource-constrained conditions proposed in this application.

[0090] The communication unit 400 is used to send the image feature values ​​preprocessed by the multi-core processor unit 300 to the background computer in real time.

[0091] For example, the communication unit 400 is a 5G wireless communication module, which mainly reports the image feature values ​​preprocessed by the multi-core processor to the background computer in real time.

[0092] The camera unit 200 is connected to the multi-core processor unit 300, the multi-core processor unit 300 is connected to the communication unit 400, and the storage unit 100 is connected to the multi-core processor unit 300.

[0093] In a further embodiment, it also includes: a power supply unit 500 for supplying power to the camera unit 200, the multi-core processor unit 300 and the communication unit 400, wherein the camera unit 200, the multi-core processor unit 300 and the communication unit 400 are all electrically connected to the power supply unit 500;

[0094] For example, the power supply unit 500 is a series of conversion circuits constructed from AC-CDC and DC-CDC converters. The AC-CDC conversion circuit can directly convert 110V-380V AC power into 12V DC power, and the DC-CDC conversion circuit can convert 12V DC power into five sets of power supplies, namely 5V, 3.3V, 1.8V, 1.2V, and 1V, for use by various modules inside the device.

[0095] Example 3

[0096] The present application relates to a computer-readable storage medium in Embodiment 3, wherein the computer-readable medium stores program code for execution by a device, the program code including steps for performing the method as in any implementation of Embodiment 1 of the present application;

[0097] The computer-readable storage medium may be a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM); the computer-readable storage medium may store program code, and when the program stored in the computer-readable storage medium is executed by a processor, the processor is used to perform the steps of the method in any of the implementations of Embodiment 1 of this application.

[0098] Example 4

[0099] An electronic device according to Embodiment 4 of this application includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor. When the program or instructions are executed by the processor, they implement the method in any of the implementations in Embodiment 1 of this application.

[0100] The processor can be a general-purpose central processing unit (CPU), microprocessor, application-specific integrated circuit (ASIC), graphics processing unit (GPU), or one or more integrated circuits, used to execute relevant programs to implement the method in any of the implementations of Embodiment 1 of this application.

[0101] The processor can also be an integrated circuit electronic device with signal processing capabilities. In implementation, each step of the method in any of the implementations of Embodiment 1 of this application can be completed by the integrated logic circuitry in the processor's hardware or by software instructions.

[0102] The aforementioned processor can also be a general-purpose processor, a digital signal processor, an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules can be located in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the functions required by the units included in the data processing apparatus of the embodiments of this application, or executes the methods in any implementation of Embodiment 1 of this application.

[0103] The above are merely preferred embodiments of this application; however, the scope of protection of this application is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in this application, based on the technical solution and its improved concept, should be covered within the scope of protection of this application.

Claims

1. A local denoising method for edge-side transformer oil level images, characterized in that, include: Obtain the raw image of the transformer oil level; Construct a BN neural network model; The luminance component I is extracted separately by performing HSI spatial transformation on the original image. x ; The luminance component I is analyzed using the BN neural network model. x Convolutional processing and nonlinear mapping are performed, followed by feature learning to obtain the residual brightness F(I). X ); Brightness component I X and residual brightness F(I) X The images are superimposed to obtain the reconstructed image; The reconstructed image is then subjected to an inverse HSI transform and mapped to the RGB space to obtain a denoised image.

2. The local denoising method for edge-side transformer oil level images according to claim 1, characterized in that, The BN neural network model comprises N layers, each layer containing a nonlinear transformation operation: H n =(*), and a BN layer is set after each nonlinear transformation operation, where n represents the nth layer in the network, H n = (*) represents a network action, which includes normalization, activation, convolution, or pooling. The output X of the nth layer is... n Represented as: X n =H n ([X0, X1, ..., X n-1 ]), where [X0, X1, ..., X n-1 [] represents the feature map output of the 0th, 1st, ..., n-1st layer.

3. The local denoising method for edge-side transformer oil level images according to claim 2, characterized in that, Also includes: The BN neural network model is optimized.

4. The local denoising method for edge-side transformer oil level images according to claim 3, characterized in that, Optimizing the BN neural network model includes: Cancel the BN layer after each nonlinear transformation operation; Add a BN layer after all nonlinear transformation operations; An SE layer is added after the BN layer to enhance the dependency of the feature map in the multi-dimensional channels of the BN layer output image.

5. The local denoising method for edge-side transformer oil level images according to claim 1, characterized in that, The residual luminance F(I) X )=(I Y -I X ), where I Y This represents the luminance component of the original image without background interference.

6. The local denoising method for edge-side transformer oil level images according to claim 5, characterized in that, The luminance component I' of the reconstructed image Y =F(I X )+I X .

7. A local noise reduction device for edge-side transformer oil level images, characterized in that, include: A storage unit stores program code for execution by a multi-core processor unit, the program code including steps for performing the method as described in any one of claims 1-6; The camera unit is used to acquire raw images of the transformer oil level; A multi-core processor unit is used to execute the program code stored in the storage unit; The communication unit is used to send the image feature values ​​preprocessed by the multi-core processor unit to the background computer in real time. The camera unit is connected to the multi-core processor unit, the multi-core processor unit is connected to the communication unit, and the storage unit is connected to the multi-core processor unit.

8. The local noise reduction device for edge-side transformer oil level images according to claim 7, characterized in that, Also includes: A power supply unit is used to supply power to the camera unit, the multi-core processor unit, and the communication unit, wherein the camera unit, the multi-core processor unit, and the communication unit are all electrically connected to the power supply unit.

9. A computer-readable storage medium, characterized in that, The computer-readable medium stores program code for execution by the device, the program code including steps for performing the method as described in any one of claims 1-6.

10. An electronic device, characterized in that, The electronic device includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the method as described in any one of claims 1-6.

Images