Method for identifying inspection image of intelligent inspection robot of transformer substation

Abstract

The invention provides a transformer substation intelligent inspection robot inspection image identification method. The method comprises the following steps: acquiring a visible light image and an infrared light image of transformer substation equipment in a current transformer substation environment; matching based on the current humidity, the current illumination intensity and the current visibility to obtain a target image processing model; inputting the visible light image and the infrared light image into a target image processing model to obtain a first target image and a second target image; performing depth feature extraction on the first target image and the second target image to obtain a first depth image and a second depth image, and performing image fusion on the first depth image and the second depth image to obtain a fused image; and inputting the fused image into an image recognition model to obtain a recognition result, and determining the equipment state of the substation equipment based on the recognition result. According to the invention, the accuracy of image recognition in a complex and changeable transformer substation environment is improved, the equipment state of the transformer substation equipment is accurately recognized, and safe and stable operation of the transformer substation is guaranteed.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular, to a method for identifying inspection images of intelligent inspection robots in substations. Background Art

[0002] With the continuous expansion of the scale of the power system and the increasing requirement for power supply reliability, the safe and stable operation of substations has become increasingly important. In order to timely and accurately detect potential fault hazards and abnormal conditions that may exist in substation equipment, intelligent inspection robots are widely used in the daily inspection work of substations.

[0003] During the inspection process, intelligent inspection robots will collect a large number of device images. Accurately identifying and analyzing these images is a key link in achieving efficient fault diagnosis. At present, the existing methods for identifying inspection images of intelligent inspection robots in substations mainly include image recognition methods based on traditional machine learning, that is, for example, using artificial feature extraction combined with classifiers such as support vector machines (SVM) and AdaBoost for recognition. Usually, features such as texture, shape, and color are designed manually first, and then the classifier is used to determine whether the device status in the image is normal. However, the features designed manually by this method often rely on expert experience, have poor adaptability to the complex and changeable substation environment, and it is difficult to comprehensively and accurately capture the features in various device abnormal situations, resulting in low recognition accuracy. Especially when facing some subtle fault features or newly emerging fault types, misjudgment or missed judgment is likely to occur. Summary of the Invention

[0004] The present invention provides a method for identifying inspection images of intelligent inspection robots in substations, so as to improve the accuracy of image recognition in a complex and changeable substation environment, accurately identify the device status of substation equipment in inspection images, and ensure the safe and stable operation of substations.

[0005] In a first aspect, the present invention provides a method for identifying inspection images of intelligent inspection robots in substations, which is applied to intelligent inspection robots in substations. The identification method includes: Collect visible light images and infrared light images of substation equipment in the current substation environment; the current substation environment includes the current humidity, the current light intensity, and the current visibility; Based on the current humidity, the current light intensity, and the current visibility, perform matching in a preset model library to obtain a target image processing model in the current substation environment; Input the visible light image and the infrared light image into the target image processing model to obtain a first target image and a second target image respectively output by the target image processing model; Extract depth features from the first target image and the second target image to obtain a first depth image and a second depth image respectively, and fuse the first depth image and the second depth image to obtain a fused image; Input the fused image into a pre-trained image recognition model to obtain a recognition result input by the image recognition model, and determine the device state of the substation equipment based on the recognition result; Among them, the processing process of the target image processing model includes denoising processing, light uniformity processing, and defogging processing; the image recognition model is trained based on sample images and their corresponding state result labels.

[0006] In a second aspect, the present invention also provides a recognition system for inspection images of a substation intelligent inspection robot, which is applied to the substation intelligent inspection robot and is used to implement the recognition method for inspection images of the substation intelligent inspection robot as described in the first aspect; the recognition system includes: An acquisition module, configured to acquire visible light images and infrared light images of substation equipment in the current substation environment; the current substation environment includes the current humidity, the current light intensity, and the current visibility; A matching module, configured to perform matching in a preset model library based on the current humidity, the current light intensity, and the current visibility to obtain a target image processing model in the current substation environment; An image processing module, configured to input the visible light image and the infrared light image into the target image processing model to obtain a first target image and a second target image respectively output by the target image processing model; An extraction and fusion module, configured to extract depth features from the first target image and the second target image to obtain a first depth image and a second depth image respectively, and fuse the first depth image and the second depth image to obtain a fused image; A device state recognition module, configured to input the fused image into a pre-trained image recognition model to obtain a recognition result input by the image recognition model, and determine the device state of the substation equipment based on the recognition result; Among them, the processing process of the target image processing model includes denoising processing, light uniformity processing, and defogging processing; the image recognition model is trained based on sample images and their corresponding state result labels.

[0007] In a third aspect, the present invention also provides an electronic device, including: a memory, configured to store a computer software program; a processor, configured to read and execute the computer software program, thereby implementing the recognition method for inspection images of the substation intelligent inspection robot as described in any one of the above.

[0008] Fourthly, the present invention further provides a non-transitory computer-readable storage medium, in which a computer software program is stored. When the computer software program is executed by a processor, it implements the method for identifying inspection images of a substation intelligent inspection robot as described in any one of the above.

[0009] Fifthly, the present invention further provides a computer program product, including a computer program. When the computer program is executed by a processor, it implements the method for identifying inspection images of a substation intelligent inspection robot as described in any one of the above.

[0010] The method for identifying inspection images of a substation intelligent inspection robot provided by the embodiments of the present invention matches an accurate image processing model according to humidity, light intensity, and visibility, and performs denoising processing, light uniformity processing, and defogging processing on the collected images, so as to remove the influence of environmental factors in the substation environment on the image quality, improve the quality of the collected images, thereby improving the accuracy of image recognition in the complex and changeable substation environment. Further, deep feature extraction and image fusion are performed on the processed images, and the texture features and detail features of the images in different modes can be fused to obtain a more accurate fused image. Then, combined with the image recognition model, the equipment state of the substation equipment in the inspection image can be accurately identified, ensuring the safe and stable operation of the substation. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 is a schematic flowchart of the method for identifying inspection images of a substation intelligent inspection robot provided by the embodiments of the present invention; Figure 2 is a schematic structural diagram of the system for identifying inspection images of a substation intelligent inspection robot provided by the embodiments of the present invention; Figure 3 is an embodiment diagram of an electronic device provided by the embodiments of the present invention; Figure 4 is an embodiment diagram of a computer-readable storage medium provided by the embodiments of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0012] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative efforts shall fall within the protection scope of the present invention.

[0013] In the description of the present invention, the terms "first" and "second" are used for descriptive purposes only and cannot be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of the described features. In the description of the present invention, "a plurality of" means two or more unless otherwise specifically defined.

[0014] In the description of the present invention, the term "for example" is used to mean "serving as an example, illustration, or explanation". Any embodiment described as "for example" in the present invention is not necessarily construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the present invention. In the following description, details are set forth for purposes of explanation. It should be understood that those of ordinary skill in the art can recognize that the present invention can be implemented without these specific details. In other instances, well-known structures and processes are not elaborated in detail to avoid obscuring the description of the present invention with unnecessary details. Therefore, the present invention is not intended to be limited to the embodiments shown, but is to be accorded the widest scope consistent with the principles and features disclosed herein.

[0015] Optionally, refer to Figure 1 , Figure 1 which is a schematic flow chart of the method for identifying inspection images of a substation intelligent inspection robot provided by the present invention. In the embodiments of the present invention, the execution subject of the method for identifying inspection images of a substation intelligent inspection robot is the substation intelligent inspection robot. Therefore, the method for identifying inspection images of a substation intelligent inspection robot includes: Step 10: Acquire visible light images and infrared light images of substation equipment in the current substation environment.

[0016] Optionally, the substation intelligent inspection robot in the embodiments of the present invention can be deployed at a preset position on the substation or in the substation area to collect image information of substation equipment in the substation in real time.

[0017] Furthermore, a visible light camera and an infrared thermal imaging camera are installed on the substation intelligent inspection robot, and the times of the visible light camera and the infrared thermal imaging camera are the same to ensure that the image information collected by the visible light camera and the infrared thermal imaging camera is aligned in time. Therefore, after receiving the instruction for equipment status identification, the substation intelligent inspection robot calls the visible light camera and the infrared thermal imaging camera to acquire visible light images and infrared light images of substation equipment in the current substation environment, where the current substation environment includes the current humidity, the current light intensity, and the current visibility.

[0018] Step 20: Match based on the current humidity, current light intensity, and current visibility in the preset model library to obtain the target image processing model for the current substation environment.

[0019] Optionally, a preset model library is stored in the substation intelligent inspection robot. The preset model library includes multiple pre-trained image processing models. The processing capabilities of each image processing model for images in different environments are different, that is, each image processing model corresponds to different humidity, light intensity, and visibility. Among them, the image recognition model is trained based on the sample image and its corresponding status result label, and the training process is as described in Steps A1 to A6.

[0020] Since the image quality of the visible light image and the infrared light image collected under different humidity, light intensity, and visibility is different, in order to obtain the best-quality image, it is necessary to match the image processing model adapted to the current substation environment in the preset model library. That is, the substation intelligent inspection robot matches in the preset model library according to the current humidity, current light intensity, and current visibility to obtain the target image processing model adapted to the current substation environment, as specifically described in Steps 201 to 204.

[0021] Step 30: Input the visible light image and the infrared light image into the target image processing model to obtain the first target image and the second target image respectively output by the target image processing model.

[0022] Furthermore, the substation intelligent inspection robot inputs the visible light image and the infrared light image into the target image processing model. The target image processing model processes the visible light image and the infrared light image to obtain the first target image and the second target image respectively output by the target image processing model. Among them, the processing process of the target image processing model includes denoising processing, light uniformity processing, and defogging processing. Therefore, it can be understood that the target image processing model performs denoising processing, light uniformity processing, and defogging processing on the visible light image and the infrared light image. The denoising processing process is as described in Steps 301 to 303, the light uniformity processing is as described in Steps 304 to 308, and the defogging processing is as described in Steps 309 to 313.

[0023] Step 40: Perform deep feature extraction on the first target image and the second target image to obtain the first depth image and the second depth image respectively, and fuse the first depth image and the second depth image to obtain the fused image.

[0024] Furthermore, the substation intelligent inspection robot performs normalization processing on the first target image and the second target image, and maps the pixel value range of the image to the interval.

[0025] Therefore, the intelligent substation inspection robot normalizes each pixel point in the first target image where and represent the horizontal and vertical coordinates of the image respectively. The specific normalization formula is as follows:

[0026] where represents the number of pixel points in the first target image, represents the pixel value of the pixel point of the first target image after normalization at represents the pixel value of the pixel point of the first target image at represents the minimum pixel value in the first target image, represents the maximum pixel value in the first target image.

[0027] The intelligent substation inspection robot normalizes each pixel point in the second target image where The specific normalization formula is as follows:

[0028] where represents the number of pixel points in the second target image, represents the pixel value of the pixel point of the second target image after normalization at represents the pixel value of the pixel point of the second target image at represents the minimum pixel value in the second target image, represents the maximum pixel value in the second target image.

[0029] Furthermore, the embodiment of the present invention performs deep feature extraction based on a feature extraction algorithm with multi-scale convolution kernels and adaptive weighting. In one embodiment, the multi-scale convolution kernel includes a set of convolution kernels with different scales where represents the scale identifier of the th convolution kernel and the size of each convolution kernel is where ​​​​​​is odd, ensuring that the convolutional kernel has a central element, which is convenient for performing symmetric convolution operations on the image. And the element values in the convolutional kernel are determined by the following method of random initialization and adaptive adjustment: .

[0030] Among them, is a random number within the range of , is a weight adaptively determined according to the local texture complexity of the image, The calculation formula of .

[0031] Among them, and respectively represent the gradient amplitudes of the pixel points at the point of the first target image and the second target image after normalization, represents the upper left corner coordinates of the local area of the image where the current convolutional kernel acts.

[0032] Furthermore, through the multi-scale convolutional kernel with different scales in perform convolution operations on the normalized image respectively to obtain feature maps at different scales.

[0033] For the first target image after normalization processing, The calculation formula of its feature map at the th scale is as follows:

[0034] Among them, represents the first feature map obtained by performing convolution operations on the first target image at the th scale after normalization processing at the point of the pixel value of the pixel point.

[0035] Similarly, for the second target image after normalization processing, The calculation formula of its feature map at the th scale is as follows::

[0036] Among them, represents the second feature map obtained by performing convolution operations on the second target image at the th scale after normalization processing at the point of the pixel value of the pixel point.

[0037] Furthermore, non-linear activation is performed on the feature maps obtained at each scale. The first feature map The first target image feature map after activation is denoted as , and the second feature map The second target image feature map after activation is denoted as . The specific formula of the activation function is as follows: .

[0038] Among them, represents the pixel value of the feature map obtained by convolution, represents adjustable parameters.

[0039] Therefore, for each pixel point in the first target image feature map , the pixel value can be expressed as . For each pixel point in the second target image feature map , the pixel value can be expressed as .

[0040] Furthermore, the feature maps after activation at different scales are fused to obtain the first depth image and the second depth image . Among them, the fusion method in the embodiments of the present invention adopts the method of weighted summation, and the weights are determined according to the contribution degrees of the feature maps at each scale to the overall feature expression.

[0041] For the first depth image , the pixel value of each pixel point in the first depth image can be expressed as: . Among them, represents the preset weight, and the calculation formula of the preset weight is: .

[0042] Among them, and respectively represent the width and height of the first target image , is the starting coordinate of the first target image (usually ).

[0043] For the second depth image , the pixel value of each pixel point in the second depth image can be expressed as: . Among them, Represents a preset weight.

[0044] Optionally, in order to better fuse the first depth image and the second depth image , a feature alignment operation is first performed. The embodiment of the present invention adopts a method based on feature point matching and local deformation, specifically: by detecting the significant feature points in the first depth image and the second depth image , the feature point coordinates of the first depth image are obtained as , and the feature point coordinates of the second depth image are obtained as , where represents the coordinates of the th feature point in the first depth image ( is the abscissa, is the ordinate), represents the coordinates of the th feature point in the second depth image ( is the abscissa, is the ordinate).

[0045] Furthermore, the matching relationship between the feature points is calculated, and the specific calculation formula is as follows: .

[0046] Where represents a parameter that controls the sensitivity of the similarity measure.

[0047] Furthermore, according to the matched feature points, the images are adjusted through the local deformation field to obtain the deformed first depth image and the deformed second depth image . The displacement vector of the local deformation field at the coordinate can be obtained by interpolation and other operations based on the matched feature points, and the specific formula is as follows: .

[0048] .

[0049] Where represents the dimension of the feature point coordinates of the first depth image , represents the dimension of the feature point coordinates of the second depth image ; represents an interpolation kernel function, such as a Gaussian kernel function.

[0050] Further, the embodiment of the present invention fuses the first depth image after deformation and the second depth image after deformation based on the algorithm of regional adaptive weighted fusion to obtain the fused image, specifically: the image is divided into a plurality of non-overlapping small regions, and for each region, the complexity metric index of each region is calculated, and the specific calculation formula is as follows: and the second depth image after deformation to obtain the fused image , specifically: the image is divided into a plurality of non-overlapping small regions, and for each region , calculate each region 's complexity metric index, and the specific calculation formula is as follows: .

[0051] Among them, represents the complexity metric index of each region , represents the gradient magnitude of the pixel point of the first depth image after deformation at the point, represents the gradient magnitude of the pixel point of the second depth image after deformation at the point, represents the number of pixel points in each region .

[0052] Further, according to the complexity metric index of each region , determine the fusion weights of the first depth image after deformation and the second depth image after deformation in each region , and the specific formula is as follows: ; ; .

[0053] Among them, represents the fusion weight of the first depth image after deformation in each region , represents the fusion weight of the second depth image after deformation in each region , and represents the preset balance parameter.

[0054] The pixel value of the pixel point of the fused image at the coordinate is calculated as follows: .

[0055] Among them, represents the number of regions, represents the first depth image after deformation at The pixel value of the pixel at the point represents the second depth image after deformation at the pixel value of the pixel at the point; represents the indicator function, when in the region inside, when not in the region inside, .

[0056] Step 50, input the fused image into a pre-trained image recognition model, obtain the recognition result input by the image recognition model, and determine the device state of the substation equipment based on the recognition result.

[0057] Optionally, input the fused image into a pre-trained image recognition model, and the image recognition model performs enhancement processing on the fused image. The image enhancement method based on multi-modal transformation in this embodiment of the present invention includes two transformation modes: spatial transformation mode and pixel value transformation mode. Specifically: For the spatial transformation mode, by randomly generating affine transformation parameters such as the translation amount and , the rotation angle , and the scaling factor to transform the image. The relationship between the new coordinates and the original coordinates is as follows: .

[0058] Among them, the transformation parameters need to meet the range constraints. For example, (in pixel units), , . When the image is enlarged, and when the image is reduced.

[0059] For the pixel value transformation mode, through the brightness-contrast adaptive adjustment algorithm. First, calculate the average brightness and the brightness standard deviation of the image: .

[0060] .

[0061] Among them, and are the width and height of the image respectively.

[0062] Furthermore, adjust the pixel values of the image. The specific formula is as follows: .

[0063] Among them, the enhanced fused image represents the pixel value of the pixel point at the coordinate , represents the gradient magnitude of the pixel point of the fused image at the coordinate .

[0064] After the above spatial transformation and pixel value transformation, the enhanced fused image is obtained.

[0065] Furthermore, the enhanced fused image is normalized again so that the pixel value range of the enhanced fused image adapts to the input requirements of the image recognition model, such as being normalized to the interval . The specific normalization formula is as follows: .

[0066] Among them, represents the minimum pixel value in the enhanced fused image , represents the maximum pixel value in the enhanced fused image .

[0067] At the same time, the enhanced fused image is cropped or padded according to the size requirements of the model input (if the sizes are inconsistent), and adjusted to a suitable format (for example, adjusted to the tensor format expected by the model, etc.) to obtain the processed input image , and is ready to be input into the pre-trained image recognition model.

[0068] Optionally, the image recognition model in the embodiments of the present invention includes multiple convolutional layers, attention layers, and fully connected layers. Therefore, after the processed input image is input into the image recognition model, at the convolutional layer stage, for the th convolutional layer and the th convolutional kernel , among them, represents the coordinates of the elements in the convolutional kernel, the size of the convolutional kernel is , and its parameter update is based on the current input feature map , that is, the output of the previous layer or the feature map of the input image after the previous processing and the prediction error feedback of the model are used for dynamic adjustment. The update formula is as follows: .

[0069] Among them, represents the learning rate parameter, and respectively represent the width and height of the feature map of the layer, represents the prediction error of the layer; represents the adaptive weight, and its calculation formula is as follows: .

[0070] After operations such as multiple convolutional layers and attention layers, image features are continuously extracted and fused, and finally a prediction result vector is output through a fully connected layer , where is the number of categories, such as the number of different substation equipment status categories, represents the prediction probability of belonging to the category status.

[0071] Furthermore, an uncertainty assessment is performed on the prediction probability vector output by the model. In the embodiments of the present invention, an uncertainty metric based on probability distribution entropy is used to measure uncertainty. The specific formula of the uncertainty metric is as follows: .

[0072] At the same time, calculate the category index corresponding to the maximum probability value of the probability vector, and the specific formula is as follows: .

[0073] Furthermore, according to the uncertainty metric and the maximum probability category index determine the final substation equipment status.

[0074] Optionally, in the embodiments of the present invention, an uncertainty threshold is set. If , then directly take the category corresponding to the maximum probability as the equipment status, that is, the equipment status is the category status.

[0075] If , it means that the model prediction uncertainty is relatively large. At this time, local feature information in the image needs to be further referred to. Therefore, the processed input image is divided into regions. For each region , calculate its relevance score with various equipment statuses, and the specific formula is as follows: .

[0076] Among them, represents the region and the The relevance score of the device status of the class Indicates the processed input image At The pixel value of the pixel point at the location Indicates the The feature template of the device status of the class at the coordinate At the location

[0077] Furthermore, calculate the comprehensive score of each category based on the regional relevance score , and the specific formula is as follows: . Among them, Indicates the number of divided regions Indicates the region The weight of, and the weight The calculation formula is as follows: .

[0078] Indicates the processed input image At the coordinate The gradient magnitude of the pixel point at the location

[0079] Finally, determine the device status according to the comprehensive score. The device status is the status corresponding to the class with the highest comprehensive score, that is, the device status is the Class status, where: .

[0080] In the embodiment of the present invention, an accurate image processing model is matched according to humidity, light intensity, and visibility to perform denoising processing, light uniformity processing, and defogging processing on the collected image, so as to remove the influence of environmental factors in the substation environment on the image quality, improve the quality of the collected image, thereby improving the accuracy of image recognition in the complex and changeable substation environment. Perform deep feature extraction and image fusion on the processed image, which can fuse the texture features and detail features of images in different modes to obtain a more accurate fused image. Then, combined with the image recognition model to recognize the fused image, the device status of the substation equipment in the inspection image can be accurately recognized, ensuring the safe and stable operation of the substation.

[0081] In one embodiment, the descriptions of steps 201 to 204 are as follows: Step 201, obtain the optimal humidity, optimal light intensity, and optimal visibility adapted to each image processing model in the preset model library.

[0082] Optionally, each image processing model in the preset model library of the embodiments of the present invention is adapted to its corresponding optimal humidity, optimal light intensity, and optimal visibility, that is, the quality of the image obtained by processing the image under the optimal humidity, optimal light intensity, and optimal visibility is the best.

[0083] Therefore, the substation intelligent inspection robot obtains the optimal humidity, optimal light intensity, and optimal visibility adapted to each image processing model.

[0084] Step 202: Based on the current humidity, current light intensity, and current visibility, and the optimal humidity, optimal light intensity, and optimal visibility of each image processing model, respectively determine the humidity matching value, light intensity matching value, and visibility matching value of each image processing model.

[0085] Further, the substation intelligent inspection robot calculates the humidity matching value of each image processing model according to the optimal humidity and current humidity of each image processing model. The specific formula is as follows: .

[0086] Wherein, represents the humidity matching value of the th image processing model, represents the current humidity, represents the optimal humidity of the th image processing model.

[0087] Further, the substation intelligent inspection robot calculates the light intensity matching value of each image processing model according to the optimal light intensity and current light intensity of each image processing model. The specific formula is as follows: .

[0088] Wherein, represents the light intensity matching value of the th image processing model, represents the current light intensity, represents the optimal light intensity of the th image processing model.

[0089] Further, the substation intelligent inspection robot calculates the visibility matching value of each image processing model according to the optimal visibility and current visibility of each image processing model. The specific formula is as follows: .

[0090] Wherein, represents the visibility matching value of the th image processing model, represents the current visibility, Indicates the optimal visibility of the th image processing model.

[0091] Step 203: Determine the comprehensive matching value for each image processing model based on the humidity matching value, light intensity matching value, and visibility matching value of each image processing model.

[0092] Furthermore, the intelligent substation inspection robot calculates the comprehensive matching value for each image processing model according to the humidity matching value, light intensity matching value, and visibility matching value of each image processing model. The specific calculation formula is as follows: .

[0093] Where, Indicates the comprehensive matching value of the th image processing model, , and represent preset weights, .

[0094] Step 204: Traverse all the comprehensive matching values, and determine the image processing model corresponding to the largest comprehensive matching value as the target image processing model.

[0095] Optionally, the intelligent substation inspection robot traverses all the comprehensive matching values, and determines the image processing model corresponding to the largest comprehensive matching value as the target image processing model.

[0096] In the embodiment of the present invention, an accurate image processing model is matched according to humidity, light intensity, and visibility to perform denoising processing, light uniformity processing, and defogging processing on the collected images, so as to remove the influence of environmental factors in the substation environment on the image quality, improve the quality of the collected images, and thus improve the accuracy of image recognition in the complex and changeable substation environment.

[0097] Optionally, the processing processes of the visible light image and the infrared light image in the target image processing model in the embodiment of the present invention are the same. Therefore, the processing process of the visible light image in the target image processing model is taken as an example in the embodiment of the present invention, and the processing process of the infrared light image is the same by analogy.

[0098] In one embodiment, the denoising process is as described in steps 301 to 303: Step 301: Obtain local pixel points in the visible light image based on a local sliding window of a preset size.

[0099] Optionally, the input visible light image in the embodiment of the present invention is , where, for the visible light image In The pixel value of the pixel at the position can be expressed as , where and respectively represent the abscissa and ordinate of the visible light image . The image size of the visible light image is , is the number of pixels of the visible light image in the height direction, is the number of pixels of the visible light image in the width direction.

[0100] Furthermore, in the embodiment of the present invention, the preset size of the local sliding window is , where is an odd number, such as 3, 5, etc. Therefore, through the local sliding window with the preset size, local pixel points in the visible light image are obtained.

[0101] Step 302: For each pixel point in the local pixel points, calculate the pixel mean and pixel standard deviation of each pixel point within the local sliding window.

[0102] Furthermore, for each pixel point in the local pixel points , calculate the pixel mean and pixel standard deviation of each pixel point within the local sliding window. The specific formulas are as follows:

[0103] where represents the pixel mean of each pixel point within the local sliding window.

[0104]

[0105] where represents the pixel standard deviation of each pixel point within the local sliding window.

[0106] Step 303: Based on the pixel mean and pixel standard deviation of each pixel point within the local sliding window, perform denoising processing on each pixel point until all pixel points in the visible light image are denoised, and a denoised image corresponding to the visible light image is obtained.

[0107] Furthermore, the preset noise threshold in the embodiment of the present invention can be determined according to the actual characteristics of the image and experience. Therefore, compare the pixel standard deviation of each pixel point within the local sliding window with the preset noise threshold. If , it is determined that the pixel may be affected by noise and is corrected. Specifically, each pixel is denoised by the pixel mean and pixel standard deviation of each pixel in the local sliding window. The formula for denoising each pixel is as follows: .

[0108] Among them, represents the pixel value of the denoised image at .

[0109] Furthermore, each pixel is denoised according to the above process until all pixels in the visible light image are denoised, and the denoised image corresponding to the visible light image is obtained .

[0110] In the embodiment of the present invention, the acquired image is denoised by the target image processing model, so that the influence of environmental factors in the substation environment on the image quality can be removed, the quality of the acquired image can be improved, and thus the accuracy of image recognition in the complex and changeable substation environment can be improved.

[0111] In one embodiment, the process of uniform illumination processing is as described in steps 304 to 308: Step 304, divide the denoised image into image blocks of a preset size.

[0112] Optionally, the preset size in the embodiment of the present invention is , is a suitable positive integer, for example = 8, 16, etc. Therefore, the denoised image is divided into non-overlapping image blocks of size . Therefore, the denoised image can be divided into image blocks.

[0113] Step 305, calculate the average brightness value of each image block based on the pixel values obtained after denoising each pixel in each image block.

[0114] Furthermore, for each image block , obtain the pixel values obtained after denoising each pixel in each image block , and calculate the average brightness value of each image block according to the pixel values obtained after denoising each pixel in each image block . The specific formula is as follows: .

[0115] Among them, it represents Indicates the th image block, Indicates the average brightness value of the th image block.

[0116] Step 306: Calculate the image brightness value of the denoised image based on the average brightness value of each image block.

[0117] Furthermore, according to the average brightness value of each image block calculate the image brightness value of the denoised image , and the specific calculation formula is as follows: . .

[0118] Step 307: Calculate the brightness adjustment coefficient of each image block based on the average brightness value of each image block and the image brightness value.

[0119] Furthermore, calculate the brightness adjustment coefficient of each image block according to the average brightness value of each image block and the image brightness value, and the specific formula is as follows: .

[0120] Wherein, indicates the brightness adjustment coefficient of the th image block.

[0121] Step 308: Update the pixel values obtained after denoising each pixel point in each image block based on the brightness adjustment coefficient of each image block to obtain the image with uniform illumination processing.

[0122] Furthermore, according to the brightness adjustment coefficient of each image block, each pixel point in each image block the pixel values obtained after denoising are updated to obtain the pixel value of the pixel point at of the image after uniform illumination processing , and all updated pixel points obtain the image with uniform illumination processing , .

[0123] In the embodiment of the present invention, the image collected is subjected to uniform illumination processing through the target image processing model, so that the influence of environmental factors in the substation environment on the image quality can be removed, the quality of the collected image can be improved, and thus the accuracy of image recognition in the complex and changeable substation environment can be improved.

[0124] In one embodiment, the defogging process is as described in steps 309 to 313: Step 309: Obtain the pixel values of each pixel in each color channel of the image after illumination uniformity processing.

[0125] Optionally, obtain the pixel values of each pixel in each color channel of the image after illumination uniformity processing. For the pixels in the image after illumination uniformity processing, their pixel values in each color channel can be expressed as , where

[0126] represents the color channel.

[0127] Optionally, for the dark channel image in an embodiment of the present invention, the dark channel of the pixel at is the minimum value of the pixel values in the three color channels (such as the RGB color space) within a neighborhood centered at (the size is set to , being an appropriate value). Therefore, according to the pixel values of each pixel in each color channel, determine the minimum value of the pixel values in the color channels within a preset neighborhood centered at each pixel to obtain the dark channel of each pixel. Among them, the formula for the dark channel of the pixel at in the dark channel image is as follows: .

[0128] Among them, represents the RGB color space, represents the side length size of the preset neighborhood.

[0129] Step 311: Perform guided filtering calculation based on the dark channel of each pixel to obtain the estimated transmittance value.

[0130] Furthermore, perform guided filtering calculation according to the dark channel of each pixel to obtain the estimated transmittance value , and the specific formula for the estimated transmittance value is as follows: .

[0131] Among them, represents the guided filtering function of the image after illumination uniformity processing.

[0132] ​For the guided filter calculation, the specific analysis process is as follows: The core idea of the guided filter is to perform a filtering operation on the image to be processed (i.e., the dark channel image in the embodiment of the present invention) based on the guidance image (i.e., the image after uniform illumination processing in the embodiment of the present invention) ), so that the filtered result can not only retain important detail information such as the edges of the image to be processed, but also be affected by the structural characteristics of the guidance image, achieving a better smoothing and estimation effect. ).

[0133] Optionally, assume that the estimated value of the transmittance after filtering in a local window centered on the pixel point (the window size is set to , is an odd number) and the guidance image and the dark channel image to be filtered satisfy a linear relationship, that is: .

[0134] Among them, represents the th local window divided by the image. In the embodiment of the present invention, the image is divided into multiple such local windows for filtering processing respectively; represents the set of pixel points covered by the th local window, and are the linear coefficients to be determined in each local window.

[0135] Furthermore, in order to determine the coefficients and in each local window, the embodiment of the present invention defines an objective function, and the goal is to make the filtered result as close as possible to the true transmittance, and at the same time require that the coefficients and in the local window change as gently as possible to ensure the smoothness of filtering. Among them, the objective function can be expressed as: .

[0136] Among them, measures the error between the estimated transmittance and the dark channel image, represents the regularization term, represents the regularization parameter.

[0137] Optionally, in order to find the coefficients that minimize the objective function and , the embodiment of the present invention respectively performs on and Take the partial derivatives and set them equal to zero.

[0138] For Take the partial derivative and set it to zero: .

[0139] After rearrangement, we get: .

[0140] Among them, represents the number of pixel points in the local window , that is .

[0141] For Take the partial derivative and set it to zero: .

[0142] Substitute the expression of obtained above into this formula, and after simplification, we get: .

[0143] Through such calculations, the expressions of the coefficients and in each local window are obtained, so that the estimated transmittance value after filtering for each pixel point in the corresponding window can be calculated according to the linear filtering model defined above.

[0144] Furthermore, traverse all pixel points of the entire image. For each pixel point , determine the local window it belongs to, and then according to the coefficients and corresponding to this window that have been calculated, calculate the estimated transmittance value after filtering according to the linear filtering model.

[0145] Step 312: Determine the estimated value of atmospheric light intensity as the maximum value of the pixel values corresponding to the pixel points with a preset percentage of the pixel values in the dark channel in the image after uniform illumination processing.

[0146] Furthermore, obtain the pixel values corresponding to the pixel points with a preset percentage (such as ) of the pixel values in the dark channel in the image after uniform illumination processing, and determine the estimated value of atmospheric light intensity as the maximum value of the pixel values corresponding to the pixel points with a preset percentage of the pixel values in the dark channel in the image after uniform illumination processing . Therefore, the specific formula for the estimated value of atmospheric light intensity is: Among them, represents the pixel value of the pixel point of the image after the illumination uniformity processing at , represents the estimated value of the atmospheric light intensity, represents the estimated value of the transmittance, represents the previous preset percentage, represents the dark channel image.

[0147] Step 313: Update the pixel value of each pixel point in the image after the illumination uniformity processing based on the estimated value of the transmittance and the estimated value of the atmospheric light intensity to obtain the first target image.

[0148] Furthermore, update the pixel value of each pixel point in the image after the illumination uniformity processing according to the estimated value of the transmittance and the estimated value of the atmospheric light intensity to obtain the first target image. The specific formula is as follows: .

[0149] Among them, That is, represents the image after the defogging processing, represents the pixel value of the pixel point of the image after the defogging processing at , represents the preset coefficient.

[0150] In the embodiment of the present invention, the acquired image is defogged through the target image processing model, so that the influence of the environmental factors of the substation environment on the image quality can be removed, the quality of the acquired image can be improved, and thus the accuracy of image recognition in the complex and changeable substation environment can be improved.

[0151] In one embodiment, the descriptions of steps A1 to A6 are as follows: Step A1: Input each sample image into the initial model to obtain the first state prediction result of each sample image output by the initial model.

[0152] Optionally, the training constraint condition for model training in the embodiment of the present invention is that the value of the loss function is less than or equal to the first preset threshold, and the output result of the target optimization function is less than or equal to the second preset threshold, where the first preset threshold and the second preset threshold are set according to the actual situation. The target optimization function is constructed based on the model parameters of the initial model, and the model parameters include the weight matrix and bias term of the hidden layer and the output layer. The initial model is trained based on the neural network.

[0153] Therefore, input each sample image into the initial model to obtain the first state prediction result of each sample image output by the initial model.

[0154] Step A2: Input the first state prediction result and the state result label of each sample image into the loss function and the target optimization function of the initial model to obtain the first loss function value and the first output result.

[0155] Further, input the first state prediction result and the state result label of each sample image into the loss function and the target optimization function of the initial model to obtain the first loss function value and the first output result. The formula of the loss function of the initial model can be expressed as: 。

[0156] The formula of the standard optimization function of the initial model can be expressed as: 。

[0157] 。

[0158] 。

[0159] 。

[0160] Among them, represents the loss function, represents the number of sample images, represents the th state prediction result of the sample image, represents the th state result label of the sample image, represents the exponential function, represents the value of the model parameters at the th step of training, represents the value of the model parameters at the th step of training, represents the learning rate, represents the loss function with respect to the model parameters gradient, represents the first gradient contribution value of a single sample image to the loss function with respect to the model parameters , represents the second gradient contribution value of a single sample image to the loss function with respect to the model parameters ; represents the sign function. When , ; when , ; when , ; represents the State prediction results of a sample image Regarding the model parameters Partial derivatives of

[0161] Step A3, if the first loss function value is greater than the first preset threshold, or / and, the first output result is greater than the second preset threshold, then update the parameters in the target optimization function based on the first state prediction results and state result labels of each sample image.

[0162] Furthermore, if it is determined that the first loss function value is less than or equal to the first preset threshold, and the first output result of the target optimization function is less than or equal to the second preset threshold, it indicates that the model parameters of the initial model and the parameters in the target optimization function are the optimal parameters, and the initial model at this time is determined as the image recognition model.

[0163] Furthermore, if it is determined that the first loss function value is greater than the first preset threshold, or / and, the first output result is greater than the second preset threshold, it indicates that the model parameters of the initial model or / and the parameters in the target optimization function are not yet the optimal parameters. At this time, it is necessary to update the parameters in the target optimization function and , adjust the parameters such that the weight matrices and bias terms of the hidden layer and the output layer are trained to be optimal parameters, and the learning rate is trained to be an optimal parameter, because if the learning rate is too large, it may cause the model to "skip over" the optimal solution during training, making it difficult to converge or even increasing the loss function value; if the learning rate is too small, the update of the model parameters will be very slow, which will increase the time and computing resources required for training and it will take a long time to achieve a good training effect.

[0164] Step A4, input each sample image into the initial model with optimized parameters, and obtain the second state prediction results of each sample image output by the optimized initial model.

[0165] Step A5, input the second state prediction results and state result labels of each sample image into the loss function and the target optimization function with optimized parameters, and obtain the second loss function value and the second output result.

[0166] Furthermore, input each sample image into the initial model with optimized parameters, and obtain the second state prediction results of each sample image output by the optimized initial model.

[0167] Furthermore, input the second state prediction results and state result labels of each sample image into the loss function and the target optimization function with optimized parameters, and obtain the second loss function value and the second output result.

[0168] Step A6: Based on the training constraint conditions, make conditional judgments on the second loss function value and the second output result until the training constraint conditions are met, and obtain the image recognition model.

[0169] Further, if it is determined that the second loss function value is less than or equal to the first preset threshold, and the second output result of the target optimization function is less than or equal to the second preset threshold, it indicates that the model parameters of the initial model and the parameters in the target optimization function are the optimal parameters, and the initial model at this time is determined as the image recognition model.

[0170] Further, if it is determined that the second loss function value is greater than the first preset threshold, or / and the second output result is greater than the second preset threshold, it indicates that the model parameters of the initial model or / and the parameters in the target optimization function are not yet the optimal parameters. At this time, it is necessary to update the parameters in the target optimization function according to the second state prediction result and the state result label of each sample image and , until the second loss function value and the second output result meet the training constraint conditions, and obtain the image recognition model.

[0171] In the embodiment of the present invention, an image processing model is trained. Therefore, the collected images are denoised, the illumination is made uniform, and the fog is removed through the image processing model, so as to be able to remove the influence of environmental factors in the substation environment on the image quality, improve the quality of the collected images, and thus improve the accuracy of image recognition in the complex and changeable substation environment.

[0172] Further, the recognition system for inspection images of a substation intelligent inspection robot provided by the present invention is described below. The recognition system for inspection images of the substation intelligent inspection robot described below can be mutually corresponding and referred to the recognition method for inspection images of the substation intelligent inspection robot described above.

[0173] Optionally, referring to Figure 2 , Figure 2 is a schematic structural diagram of the recognition system for inspection images of a substation intelligent inspection robot provided by the present invention, which is applied to a substation intelligent inspection robot. The recognition system for inspection images of the substation intelligent inspection robot includes: An acquisition module 210, configured to acquire visible light images and infrared light images of substation equipment in the current substation environment; the current substation environment includes the current humidity, the current illumination intensity, and the current visibility; A matching module 220, configured to perform matching in a preset model library based on the current humidity, the current illumination intensity, and the current visibility, and obtain a target image processing model in the current substation environment; The image processing module 230 is configured to input the visible light image and the infrared light image into the target image processing model, and obtain a first target image and a second target image respectively output by the target image processing model; The extraction and fusion module 240 is configured to perform deep feature extraction on the first target image and the second target image to obtain a first depth image and a second depth image respectively, and perform image fusion on the first depth image and the second depth image to obtain a fused image; The device status recognition module 250 is configured to input the fused image into a pre-trained image recognition model, obtain a recognition result input by the image recognition model, and determine the device status of the substation device based on the recognition result.

[0174] In the embodiment of the present invention, an accurate image processing model is matched according to humidity, light intensity, and visibility to perform denoising processing, light uniformity processing, and defogging processing on the collected images, so as to remove the influence of environmental factors in the substation environment on the image quality, improve the quality of the collected images, thereby improving the accuracy of image recognition in the complex and changeable substation environment. Performing deep feature extraction and image fusion on the processed images can fuse the texture features and detail features of images in different modes to obtain a more accurate fused image. Then, combining the image recognition model to recognize the fused image can accurately recognize the device status of the substation device in the inspection image, ensuring the safe and stable operation of the substation.

[0175] Please refer to Figure 3 , Figure 3 which is an embodiment diagram of the electronic device provided by the embodiment of the present invention. As Figure 3 shown, the embodiment of the present invention provides an electronic device 300, including a memory 310, a processor 320, and a computer program 311 stored on the memory 310 and executable on the processor 320. When the processor 320 executes the computer program 311, the following steps are implemented: Collect visible light images and infrared light images of substation devices in the current substation environment; the current substation environment includes the current humidity, the current light intensity, and the current visibility; Match based on the current humidity, the current light intensity, and the current visibility in a preset model library to obtain a target image processing model in the current substation environment; Input the visible light image and the infrared light image into the target image processing model, and obtain a first target image and a second target image respectively output by the target image processing model; Perform deep feature extraction on the first target image and the second target image to obtain a first depth image and a second depth image respectively, and perform image fusion on the first depth image and the second depth image to obtain a fused image; Input the fused image into a pre-trained image recognition model to obtain the recognition result input by the image recognition model, and determine the device status of the substation equipment based on the recognition result.

[0176] Please refer to Figure 4 , Figure 4 which is the embodiment diagram of the computer-readable storage medium provided by the embodiment of the present invention. As Figure 4 shown, this embodiment provides a computer-readable storage medium 400, on which a computer program 311 is stored. When the computer program 311 is executed by a processor, the following steps are implemented: Collect visible light images and infrared light images of substation equipment in the current substation environment; the current substation environment includes the current humidity, the current light intensity, and the current visibility; Match based on the current humidity, the current light intensity, and the current visibility in a preset model library to obtain the target image processing model in the current substation environment; Input the visible light image and the infrared light image into the target image processing model to obtain the first target image and the second target image respectively output by the target image processing model; Perform deep feature extraction on the first target image and the second target image to obtain the first depth image and the second depth image respectively, and fuse the first depth image and the second depth image to obtain a fused image; Input the fused image into a pre-trained image recognition model to obtain the recognition result input by the image recognition model, and determine the device status of the substation equipment based on the recognition result.

[0177] On the other hand, the present invention also provides a computer program product. The computer program product includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the recognition method of the inspection image of the substation intelligent inspection robot provided by the above-mentioned various methods. The method includes: Collect visible light images and infrared light images of substation equipment in the current substation environment; the current substation environment includes the current humidity, the current light intensity, and the current visibility; Match based on the current humidity, the current light intensity, and the current visibility in a preset model library to obtain the target image processing model in the current substation environment; Input the visible light image and the infrared light image into the target image processing model to obtain the first target image and the second target image respectively output by the target image processing model; Perform deep feature extraction on the first target image and the second target image to obtain the first depth image and the second depth image respectively, and fuse the first depth image and the second depth image to obtain a fused image; Input the fused image into a pre-trained image recognition model to obtain the recognition result input by the image recognition model, and determine the device status of the substation equipment based on the recognition result.

[0178] The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art can understand and implement it without creative efforts.

[0179] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the essence of the above technical solution, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0180] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments or perform equivalent replacements for some of the technical features. These modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for recognizing inspection images of a substation intelligent inspection robot, characterized in that: Applied to the intelligent inspection robot of a substation, the identification method includes: Collecting visible light images and infrared light images of substation equipment in the current substation environment; the current substation environment includes current humidity, current light intensity and current visibility; Based on the current humidity, the current light intensity and the current visibility, a target image processing model under the current substation environment is obtained by matching in a preset model library; Inputting the visible light image and the infrared light image into the target image processing model to obtain a first target image and a second target image outputted by the target image processing model respectively; Performing depth feature extraction on the first target image and the second target image to obtain a first depth image and a second depth image respectively, and performing image fusion on the first depth image and the second depth image to obtain a fused image; Inputting the fused image into a pre-trained image recognition model to obtain a recognition result of the image recognition model input, and determining the equipment status of the substation equipment based on the recognition result; Among them, the processing process of the target image processing model includes denoising processing, illumination uniformity processing and defogging processing; the image recognition model is trained based on sample images and their corresponding state result labels.

2. The method for recognizing inspection images of a substation intelligent inspection robot according to claim 1, characterized in that: The matching in a preset model library based on the current humidity, the current light intensity and the current visibility to obtain a target image processing model under the current substation environment includes: Obtaining the optimal humidity, optimal light intensity and optimal visibility adapted by each image processing model in the preset model library; Based on the current humidity, the current light intensity and the current visibility, and the optimal humidity, optimal light intensity and optimal visibility of each image processing model, respectively determine a humidity matching value, a light intensity matching value and a visibility matching value for each image processing model; Determine a comprehensive matching value with each image processing model based on the humidity matching value, the light intensity matching value and the visibility matching value with each image processing model; All comprehensive matching values ​​are traversed, and the image processing model corresponding to the largest comprehensive matching value is determined as the target image processing model.

3. The method for recognizing inspection images of a substation intelligent inspection robot according to claim 2, characterized in that: The calculation formula of the comprehensive matching value is as follows: ； ； ； ； in, Indicates The comprehensive matching value of the image processing model, Indicates Humidity matching value of the image processing model, Indicates The illumination intensity matching value of the image processing model, Indicates The visibility matching value of the image processing model, Indicates the current humidity. Indicates The optimal humidity of the image processing model, Indicates the current light intensity. Indicates The optimal illumination intensity for an image processing model, Indicates the current visibility. Indicates The optimal visibility of an image processing model, , and Indicates the preset weight.

4. The method for recognizing inspection images of a substation intelligent inspection robot according to claim 1, characterized in that: The process of the target image processing model performing denoising on the visible light image includes: Acquire local pixel points in the visible light image based on a local sliding window of a preset size; For each pixel in the local pixel points, calculating the pixel mean and pixel standard deviation of each pixel in the local sliding window; Performing denoising processing on each pixel point based on a pixel mean and a pixel standard deviation of each pixel point within the local sliding window until all pixels in the visible light image are denoised to obtain a denoised image corresponding to the visible light image; The formula for denoising each pixel based on the pixel mean and pixel standard deviation of each pixel in the local sliding window is as follows: ； in, Indicates that the image after denoising is The pixel value at represents a visible light image, Indicates that the visible light image is The pixel value at represents the horizontal and vertical coordinates in the image, Represents the pixel mean of each pixel in the local sliding window, represents the pixel standard deviation of each pixel in the local sliding window, Indicates the preset noise threshold.

5. The method for recognizing inspection images of a substation intelligent inspection robot according to claim 4, characterized in that: The process of the target image processing model performing uniform illumination processing on the visible light image includes: Dividing the denoised image into image blocks of a preset size; each image block does not overlap with each other; Based on the pixel value obtained after denoising of each pixel point in each image block, the average brightness value of each image block is calculated; Calculating the image brightness value of the denoised image based on the average brightness value of each image block; Calculating a brightness adjustment coefficient for each image block based on an average brightness value of each image block and the image brightness value; Based on the brightness adjustment coefficient of each image block, the pixel value of each pixel point in each image block after denoising is updated to obtain an image after uniform illumination processing; Wherein, based on the average brightness value of each image block and the image brightness value, the calculation formula for calculating the brightness adjustment coefficient of each image block is as follows: ； in, Indicates The brightness adjustment coefficient of the image block, Indicates The average brightness value of the image blocks, Represents the image brightness value, Indicates the number of pixels of the visible light image in the height direction, Indicates the number of pixels of the visible light image in the width direction, Indicates the preset size.

6. The method for recognizing inspection images of a substation intelligent inspection robot according to claim 5, characterized in that: The process of the target image processing model performing defogging on the visible light image includes: Obtaining the pixel value of each pixel in each color channel in the image after the uniform illumination processing; Based on the pixel value of each pixel point in each color channel, the minimum value of the pixel value in the color channel in a preset neighborhood centered on each pixel point is determined to obtain a dark channel for each pixel point; Based on the dark channel of each pixel, guided filtering calculation is performed to obtain the transmittance estimation value; Determine the maximum value of the pixel values ​​corresponding to the pixel points with a preset percentage before the pixel values ​​in the dark channel in the image after the uniform illumination processing as the estimated value of the atmospheric light intensity; The pixel value of each pixel in the image after the uniform illumination processing is updated based on the transmittance estimation value and the atmospheric light intensity estimation value to obtain the first target image; The specific formula for updating the pixel value of each pixel in the image after the uniform illumination processing based on the transmittance estimation value and the atmospheric light intensity estimation value is as follows: ； ； ； in, Indicates that the image after dehazing is The pixel value at Indicates that the image after uniform illumination processing is The pixel value at represents the estimated value of atmospheric light intensity, represents the transmittance estimate, represents the preset coefficient, Indicates the previous preset percentage, represents the dark channel image, Indicates the image after uniform illumination processing The guided filter function is represents the color channel, represents the RGB color space, Indicates the side length of the preset neighborhood. Represents pixel In each The pixel value on the color channel.

7. The method for recognizing inspection images of a substation intelligent inspection robot according to any one of claims 1 to 6, characterized in that: The training constraint conditions are: the loss function value is less than or equal to a first preset threshold, and the output result of the target optimization function is less than or equal to a second preset threshold; The target optimization function is constructed based on model parameters of an initial model, wherein the model parameters include weight matrices and bias items of a hidden layer and an output layer, and the initial model is obtained based on neural network training; The specific process of training the image recognition model includes: Input each sample image into the initial model to obtain a first state prediction result of each sample image output by the initial model; Inputting the first state prediction result and the state result label of each sample image into the loss function and the target optimization function of the initial model to obtain a first loss function value and a first output result; If the first loss function value is greater than a first preset threshold, or / and the first output result is greater than a second preset threshold, updating the parameters in the target optimization function based on the first state prediction result and the state result label of each sample image; Input each sample image into the initial model after parameter optimization, and obtain the second state prediction result of each sample image output by the optimized initial model; Inputting the second state prediction result and the state result label of each sample image into the loss function and the target optimization function after parameter optimization to obtain a second loss function value and a second output result; Performing conditional judgment on the second loss function value and the second output result based on the training constraint condition until the training constraint condition is satisfied, thereby obtaining the image recognition model; Among them, the formula of the loss function of the initial model can be expressed as: ； The formula of the objective optimization function can be expressed as: ； ； ； ； in, represents the loss function, represents the number of sample images, Indicates The state prediction results of sample images are: Indicates The state result label of sample images, represents the exponential function, Indicates that in the training The values ​​of the model parameters at step time, Indicates that in the training The values ​​of the model parameters at step time, represents the learning rate, Represents the loss function For model parameters The gradient of Represents a single sample image The loss function is related to the model parameters The first gradient contribution value of Represents a single sample image The loss function is related to the model parameters The second gradient contribution value of represents a symbolic function, when hour, ;when hour, ;then , ; Indicates The state prediction results of sample images About Model Parameters The partial derivative of .

8. A system for recognizing inspection images of a substation intelligent inspection robot, characterized in that: Applied to a substation intelligent inspection robot, used to implement the inspection image recognition method of the substation intelligent inspection robot according to any one of claims 1 to 7; the recognition system comprises: An acquisition module, used to acquire visible light images and infrared light images of substation equipment in a current substation environment; the current substation environment includes current humidity, current light intensity and current visibility; A matching module, used for matching in a preset model library based on the current humidity, the current light intensity and the current visibility to obtain a target image processing model under the current substation environment; An image processing module, used for inputting the visible light image and the infrared light image into the target image processing model to obtain a first target image and a second target image respectively output by the target image processing model; An extraction and fusion module is used to extract depth features from the first target image and the second target image to obtain a first depth image and a second depth image respectively, and to fuse the first depth image and the second depth image to obtain a fused image; An equipment status recognition module, used for inputting the fused image into a pre-trained image recognition model, obtaining a recognition result of the image recognition model input, and determining the equipment status of the substation equipment based on the recognition result; Among them, the processing process of the target image processing model includes denoising processing, illumination uniformity processing and defogging processing; the image recognition model is trained based on sample images and their corresponding state result labels.

9. An electronic device, comprising: Memory for storing computer software programs; A processor, used to read and execute the computer software program, characterized in that when the processor executes the computer software program, it implements the method for recognizing the inspection image of the substation intelligent inspection robot as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer software program stored therein, characterized in that: When the computer software program is executed by the processor, the method for recognizing inspection images of the substation intelligent inspection robot as claimed in any one of claims 1 to 7 is implemented.

Images