Infrared thermal image high-temperature target monitoring method for periphery of fan cabin

Abstract

The invention discloses an infrared thermal image high-temperature target monitoring method for the periphery of a fan cabin, and the method comprises the steps: improving the contrast ratio of a target and a background in an image through the statistics of a gray histogram, the searching of a trough, the piecewise linear stretching, and the increase of the weight of a high-temperature segment; and the phenomenon that the target and part of the background are not clearly distinguished and the image whitening phenomenon caused by the sky effect are avoided. The method comprises the following steps: acquiring an original image; performing gray histogram statistics on the original image to obtain a histogram N1; performing mean filtering on the histogram N1 to obtain a filtered gray histogram N2; trough points in the gray level histogram N2 are detected, points smaller than trough gray levels in the gray level histogram N1 are removed, and a gray level histogram N3 is obtained; stretching the gray level histogram N3 after the points are removed in a piecewise linear mode, and increasing the weight proportion of a high-temperature section in the stretching process; outputting an image with a high-temperature target highlighting effect; and a convolutional neural network is adopted to improve the algorithm effect.

Description

Technical Field

[0001] The invention belongs to the technical field of infrared image processing, and in particular relates to a method for monitoring high-temperature targets by infrared thermal imaging of the perimeter of a fan cabin. Background Art

[0002] With the development of the production and manufacturing technology of infrared detectors, various types of corresponding infrared imaging devices have become more and more popular in people's lives. People are getting used to using infrared imaging devices, such as detecting the surrounding environment and looking for small animals during outdoor adventures at night. In these activities, infrared devices directly present the ambient temperature information to users in the form of thermal imaging. Compared with visible light devices, infrared instruments are less affected by the dim environment at night.

[0003] The main purpose of an infrared thermal imager is to detect high-temperature targets in the environment. In actual scenes, infrared instruments not only receive infrared radiation from the target, but also background radiation. If no processing is performed, the distinction between the target and the background is often not obvious enough, which is not conducive to rapid target detection. Therefore, it is often necessary to add a target highlighting function through image contrast stretching to meet the needs of different scenes.

[0004] The existing methods for improving the contrast of infrared images and highlighting high-temperature targets usually include: using linear stretching processing to increase the gain to achieve improved contrast; using local histogram equalization to highlight high-temperature targets. Increasing the gain in the linear stretching process is to increase the distinction between the target and the background by assigning a higher 8-bit grayscale to the high-temperature target and a lower 8-bit grayscale to the low-temperature scene; using local histogram equalization is to exclude background factors through block local processing to display the high-temperature target. However, there are the following disadvantages: 1) When using the method of increasing the gain in the linear stretching process, due to the large gray difference between the sky area and the non-sky area, after linear stretching, the grayscale of the non-sky area is mainly concentrated in the upper 8 bits, so the image will look whitish as a whole; using local histogram equalization, the grayscale is assigned according to the number of points. Although it can suppress the influence of the sky background, when there are a large number of trees or clouds in the scene, the distinction between the high-temperature target and the tree will not be obvious; for example, when there are a large number of trees in the scene, the grayscale of the tree trunk and the bird is close, the number of birds accounts for a small proportion, and the grayscale assigned is small, so the brightness of the bird in the image looks close to the brightness of the tree trunk. Use local histogram equalization. Summary of the invention

[0005] In order to make up for the shortcomings of the prior art, the present invention provides a method for monitoring high-temperature targets in infrared thermal images around a wind turbine cabin, which aims to highlight high-temperature targets in infrared images around a wind turbine cabin by image processing, and improve the contrast between the target and the background in the image by statistical grayscale histogram, finding troughs, piecewise linear stretching and increasing the weight of the high-temperature segment, effectively avoiding the phenomenon of unclear distinction between the target and part of the background and the whitening of the image caused by the sky effect.

[0006] The technical problem solved by the present invention can be achieved through the following technical solutions:

[0007] The method for monitoring high-temperature targets by infrared thermal imaging around the perimeter of a fan cabin comprises the following specific steps:

[0008] S1. Obtain an image output by the infrared detector and subjected to non-uniformity correction and filtering noise reduction, and use it as the original image;

[0009] S2. Perform grayscale histogram statistics on the original image to obtain a histogram N1;

[0010] S3. Perform mean filtering on the histogram N1 to obtain a filtered grayscale histogram N2;

[0011] S4. According to the characteristics of the scene image, the trough points in the grayscale histogram N2 are detected, and the points whose grayscale is less than the trough grayscale in the grayscale histogram N1 are removed to obtain the grayscale histogram N3;

[0012] S5. The grayscale histogram N3 after the points are removed is stretched in a piecewise linear manner, and the weight ratio of the high temperature segment is increased during the stretching process;

[0013] S6. outputting an image with a high temperature target highlighting effect;

[0014] S7. Use convolutional neural network to improve the algorithm effect.

[0015] Furthermore, in step S1, the image output by the detector is I(x, y), where (x, y) represents the pixel coordinates in the image. The non-uniformity correction aims to eliminate the problem of inconsistent responses of each pixel of the detector. The filtering noise reduction uses a suitable filtering algorithm to remove noise interference in the image. A filter with a filter window width of K is used, where K is 1 or 2, to obtain a relatively pure original image I0(x, y):

[0016]

[0017] Where G(i,j,σ) is the filter kernel; σ is the width of the Gaussian filter (determines the degree of smoothing). The larger the σ is, the wider the frequency band of the Gaussian filter is and the better the smoothing is. The value range of i and j is between the filter module width (-K,K).

[0018] Further, in step S2, the gray-scale of the original image I0(x, y) is statistically analyzed point by point. Assuming the gray-scale level of the image is (0 - 16384), the number of pixels with gray value k, where 0 < k < L, in the image is counted as n k , thereby obtaining the gray-scale histogram N1(k) = n k .

[0019] Further, in step S3, assuming the size of the filtering window is w, for each N1 gray value in the histogram, the statistically counted number N2(k) of its filtered gray value is calculated as follows:

[0020]

[0021] where, when i < 0 or i > L, N1(i) = 0.

[0022] Further, in step S4, let ΔN2(k) = N2(k + 1) - N2(k). When ΔN2(k - 1) > 0 and ΔN2(k) < 0, the k point may be the valley point k min ; after finding the valley point, all points in the histogram N1 with gray value less than the valley gray value k min are set to zero, obtaining the histogram N3(k):

[0023]

[0024] Further, in step S5, the gray-scale histogram N3 is divided into multiple segments: the low gray-scale segment [0, k1], the medium-low gray-scale segment [k1, k2], and the high gray-scale segment [k2, L]. Assuming the stretching slopes of the low gray-scale segment, the medium gray-scale segment, and the high gray-scale segment are m1, m2, and m3 respectively, where m3 > m2 > m1, and each segment is linearly stretched with a different slope. By increasing the stretching slope of the high-temperature segment, the contrast between high and low temperatures is increased, that is:

[0025]

[0026] Further, the processing procedure of step S7 is as follows:

[0027] S71. Collect a large number of infrared pictures of the perimeter of the fan nacelle after the above processing with different numbers of segments and different stretching slopes as the data set, and perform unified preprocessing on the collected image data set;

[0028] S72. Establish a model and conduct training;

[0029] S73. Model evaluation and optimization.

[0030] Furthermore, the model adopts a convolutional neural network architecture based on the ResNet architecture, which includes multiple convolutional layers, pooling layers, and fully connected layers. The convolutional layer extracts features by sliding the convolution kernel on the image; the pooling layer is used to downsample the feature map; the fully connected layer is used to integrate the extracted features and finally output results related to the target.

[0031] Furthermore, the specific training process is:

[0032] The processed image data is taken as input, and the corresponding ideal target highlighting effect annotation is taken as output. A batch of images is input each time, and the output result of the network is calculated through forward propagation, and then compared with the real label to calculate the value of the loss function;

[0033] Using the back-propagation algorithm, according to the value of the loss function, the error is back-propagated from the output layer to the input layer, and the weight parameters of each layer in the network are adjusted to gradually reduce the loss function and continuously optimize the performance of the network. After multiple iterative training, the model learns the intrinsic relationship between different segmentation and slope settings and the target highlighting effect.

[0034] Furthermore, in step S73, the performance of the model is monitored using the validation set to observe changes in the model on the validation set to avoid overfitting of the model on the training set. If overfitting occurs, the model is optimized using a regularization method to adjust the network structure or hyperparameters. After the training is completed, the final model is evaluated using the test set to view the actual performance of the model on data that has not participated in the training, to verify the generalization ability of the model, and to ensure that it can accurately recommend the appropriate number of segments and the stretching slope of each segment on the actual infrared image of the fan cabin perimeter.

[0035] Compared with the prior art, the present invention has the following advantages:

[0036] 1) The method of the present invention eliminates the influence of background factors by detecting the trough before the stretching step, and increases the slope of the high-temperature section during the stretching process, which is equivalent to reducing the slope of the background of the low-temperature section, thereby eliminating the influence of background factors and preventing the image from turning white when the sky is used as the background.

[0037] 2) The method of the present invention adopts piecewise linear stretching, which is different from the local histogram equalization which stretches according to the number of points. Each segment of the piecewise linear still uses linear stretching, and the slope of the high-temperature segment is increased during the stretching process, so that the highlighting effect of the high-temperature target is obvious enough, and the situation of insufficient distinction between high and low temperature targets and insufficient highlighting effect will not occur. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 is a flow chart of the method of the present invention;

[0039] Figure 2 The infrared original image after non-uniformity correction and filtering noise reduction in this embodiment of the invention corresponds to the image processed in step S1;

[0040] Figure 3 To invent the "burr phenomenon" of the original grayscale histogram in this example, the histogram of the image before processing in step 3 is corresponding;

[0041] Figure 4 is a smoothed grayscale histogram after mean filtering in the example of the present invention, corresponding to the histogram of the image processed in step 3;

[0042] Figure 5 is the trough point found in the example of the present invention, corresponding to the trough point of the histogram before the processing in step 4;

[0043] Figure 6 This is the piecewise linear stretching result without trough screening and high temperature segment weight increase in the example of the present invention, that is, the image without processing in steps 4 and 5;

[0044] Figure 7 The image is the result of the piecewise linear stretching after the trough screening in the example of the present invention, and is processed by step 4, but the image is processed by the piecewise linear stretching step;

[0045] Figure 8 This is the result of piecewise linear stretching in which the trough is first screened out and then the weight of the high temperature section is increased in the example of the present invention, and the image is processed through steps 4 and 5. DETAILED DESCRIPTION

[0046] In order to make the purpose, technical solution and advantages of the present invention clearer, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0047] The present invention eliminates the interference of background factors by detecting the grayscale value of the trough, improves the prominence of the high-temperature target by increasing the stretching slope of the piecewise linear high-temperature segment, and uses a deep learning algorithm (convolutional neural network) to model and learn the data set. Through multiple iterative training, the model learns the intrinsic relationship between different segmentation and slope settings and the target prominence effect, and finally obtains a data model to improve the effect of the infrared thermal imaging high-temperature target monitoring method of the fan cabin perimeter.

[0048] like Figure 1 As shown, a method for monitoring high-temperature targets by infrared thermal imaging around a fan cabin comprises the following specific steps:

[0049] Step 1: Obtain the image output by the infrared detector after non-uniformity correction and filtering noise reduction, and use it as the original image, such as Figure 2 as shown

[0050] An output image is obtained from an infrared detector, and the image is subjected to non-uniformity correction and filtering for noise reduction. The image output by the detector is I(x, y), where (x, y) represents the pixel coordinates in the image. Non-uniformity correction aims to eliminate the problem of inconsistent responses of each pixel of the detector, and filtering for noise reduction uses a suitable filtering algorithm (such as Gaussian filtering, etc.) to remove noise interference in the image. In this embodiment, a filter with a filter window width of K is used, and K usually takes a value of 1 or 2 to obtain a relatively pure original image I0(x, y), that is:

[0051]

[0052] In the formula, G(i, j, σ) is the filter kernel; σ is the Gaussian filter width (determining the smoothness), the larger σ is, the wider the frequency band of the Gaussian filter is, and the better the smoothness is; the value ranges of i and j are between the filter module widths (-K, K), for example (-2, 2).

[0053] Step 2: Perform a gray-level histogram statistics on the original image to obtain a histogram N1.

[0054] The histogram shows the number of corresponding points for each gray level, and the gray levels of each point of the original image I0(x, y) are statistically counted. Let the gray level of the image be (0 - 16384), and count the number n of pixel points with a gray value of k (0 < k < L) in the image k , so as to obtain the gray-level histogram N1(k) = n k . For example, for an image of 640 * 480, traverse each pixel point. If the gray value of a certain pixel point is 100, then the value of n 100 is incremented by 1; through such a statistical process, a complete gray-level histogram is obtained, where the abscissa is the gray value and the ordinate is the number of pixel points corresponding to the gray value.

[0055] Step 3: Perform mean filtering on the histogram N1 to obtain a filtered gray-level histogram N2.

[0056] In an actual scenario, the gray-level histogram of the original image often has a "spike phenomenon", as Figure 3 shown, that is, the values of adjacent gray levels in the image do not change smoothly, but there are drastic jumps in gray levels. In order to facilitate subsequent detection of valleys in the histogram, in this method, by performing mean filtering on the gray-level histogram N1, a relatively smooth gray-level histogram N2 is obtained, as Figure 4 shown.

[0057] Let the filter window size be w (w is usually an odd number, such as 5 or 3). For each N1 gray value in the histogram, the statistical quantity N2(k) of its filtered gray value is calculated as follows:

[0058]

[0059] Among them, when i<0 or i>L, N1(i)=0.

[0060] Step 4: According to the characteristics of the scene image, the trough points in the grayscale histogram N2 are detected, and the points in the grayscale histogram N1 that are less than the trough grayscale are removed to obtain the grayscale histogram N3.

[0061] The histogram of most images has multiple peaks and valleys, and each "peak-valley" structure often represents the grayscale of a certain object or scene. For example, two "peak-valley" structures usually correspond to the grayscale distribution of the target and the background respectively. The ultimate goal is to highlight the high-temperature target, which belongs to the high-grayscale "peak-valley" structure. By detecting the valley points in the grayscale histogram N2, the "peak-valley" structure of the low grayscale range is found, and the points in the histogram N1 that belong to the "peak-valley" structure of the low grayscale range are set to zero, that is, the interference of the background is eliminated, and the grayscale histogram N3 after the points are removed is obtained.

[0062] The peak and trough structure in the image histogram is related to the object and the scene. In the infrared image of the perimeter of the wind turbine nacelle, the high-temperature target is in a high-grayscale "peak-trough" structure, and the grayscale range of the target and the background can be determined by detecting the trough points.

[0063] The method to detect the valley point in histogram N2 is to calculate the difference between adjacent grayscale values. When the difference changes from positive to negative, there may be a valley point. Let ΔN2(k)=N2(k+1)-N2(k). When ΔN2(k-1)>0 and ΔN2(k)<0, point k may be the valley point k. min ,like Figure 5 After finding the trough point, the grayscale value in the histogram N1 that is less than the trough grayscale k min All points are set to zero to obtain the histogram N3(k). This eliminates the interference of the low gray background and allows subsequent processing to focus on the gray area related to the high temperature target, that is:

[0064]

[0065] Step 5: The grayscale histogram N3 after the points are removed is stretched in a piecewise linear manner, and the weight ratio of the high temperature segment is increased during the stretching process.

[0066] The grayscale histogram N3 is divided into multiple segments, such as low grayscale segment [0, k1], medium and low grayscale segment [k1, k2], and high grayscale segment [k2, L]. The stretching slopes of the low grayscale segment, medium grayscale segment, and high grayscale segment are set to m1, m2, and m3, respectively (m3>m2>m1). Each segment is linearly stretched with a different slope. In order to highlight the high temperature segment target, the contrast between high and low temperatures is improved by increasing the weight ratio of the high temperature segment, that is, increasing the stretching slope of the high temperature segment, that is:

[0067]

[0068] like Figure 6 As shown in the figure, the piecewise linear stretching result without trough screening and increasing the weight of the high temperature section, and the image without processing in steps 4 and 5; Figure 7 As shown in the figure, the result of the piecewise linear stretching after the trough screening is first performed, and the image of the piecewise linear stretching processing step is processed by step 4; Figure 8 As shown in the figure, the result of piecewise linear stretching after first filtering out the trough and then increasing the weight of the high temperature segment is shown, and the image is processed by steps 4 and 5. By comparison, it can be seen that the present invention uses a piecewise linear stretching method and increases the stretching slope of the high temperature segment at the same time, and finally outputs an image with a highlighting effect.

[0069] Step 6: Output an image with a high temperature target highlighting effect.

[0070] After the above steps 1-5, the interference of background factors is eliminated, the weight of the high-temperature target is increased, and the final stretching result of the image will highlight the high-temperature target.

[0071] Step 7: Use convolutional neural network to improve the algorithm effect.

[0072] A large number of infrared images of the perimeter of the wind turbine nacelle after the above processing with different numbers of segments (such as divided into 2 segments, 3 segments, 4 segments, etc.) and different stretching slopes (such as setting different specific slope values ​​for each segment) are collected as a data set.

[0073] The classic convolutional neural network architecture is selected, and the ResNet architecture is used as the basis. ResNet has a residual connection structure, which can effectively solve the gradient vanishing problem during deep network training, and facilitate the model to learn the mapping relationship between more complex image features and target highlighting effects. It contains multiple convolutional layers, pooling layers, and fully connected layers, and different layers have different functions. The convolution layer extracts features by sliding the convolution kernel on the image. For example, a convolution kernel slides on the image, and performs weighted summation and other operations on each local area to extract different feature maps; the pooling layer (such as the maximum pooling layer) is used to downsample the feature map, reducing the amount of data while retaining key features. For example, the maximum pooling will select the maximum value in the local area as the representative value of the area, making the feature more robust; the fully connected layer is used to integrate the extracted features and finally output the results related to the target. The processing process is as follows:

[0074] (1) Data preprocessing and input

[0075] For the collected image data sets, unified preprocessing is required. First, the image size is adjusted to the fixed size required by the network input, such as adjusting all images to 32*32 pixels (to adapt to the default input size requirements of some common network architectures). Then the image is normalized and the pixel value range is normalized to a specific interval (such as [0,1] or [-1,1], etc.), which helps to speed up the training convergence of the network. The preprocessed image data is input into the convolutional neural network in the form of batches. A batch usually contains several images (such as 32, 64, etc.), which is convenient for parallel computing and improves training efficiency.

[0076] (2) Training process

[0077] The dataset is divided into a training set, a validation set, and a test set, generally in a common ratio such as 7:2:1 or 8:1:1. In the training phase, the processed image data is used as input, and the corresponding ideal target highlighting effect annotation is used as output (that is, the label, for example, which areas of the image are manually annotated as high-temperature targets and to what extent the highlighting effect should be achieved). Each time a batch of images is input, the output of the network is calculated through forward propagation, and then compared with the true label, and the value of the loss function (such as mean square error loss function, cross entropy loss function, etc., select the appropriate loss function according to the specific task) is calculated to measure the difference between the network output and the true result. Then, the back propagation algorithm is used to back propagate the error from the output layer to the input layer according to the value of the loss function, and the weight parameters of each layer in the network are adjusted so that the loss function gradually decreases and the performance of the network is continuously optimized. After multiple iterative training (for example, the number of training rounds is set to 100 or more, which is adjusted according to the actual situation), the model learns the intrinsic relationship between different segmentation and slope settings and the target highlighting effect, and finally obtains a stable and reliable data model.

[0078] (3) Model evaluation and optimization

[0079] During the training process, the performance of the model is monitored using the validation set, and the changes in the loss value, accuracy and other indicators of the model on the validation set are observed to avoid overfitting of the model on the training set (i.e., the model performs well on the training set, but performs poorly on new data). If overfitting is found, some regularization methods (such as L1 regularization, L2 regularization, Dropout, etc.) can be used to optimize the model, adjust the network structure or hyperparameters (such as learning rate, batch size, etc.) to continuously improve the performance of the model on the validation set. After the training is completed, the final model is evaluated using the test set to check the actual performance of the model on data that did not participate in the training, verify the generalization ability of the model, and ensure that it can accurately recommend the appropriate number of segments and the stretching slope of each segment on the actual infrared image of the perimeter of the fan cabin, thereby improving the effect of the infrared thermal imaging high-temperature target monitoring method of the perimeter of the fan cabin.

[0080] The deep learning algorithm (convolutional neural network, etc.) is used to model and learn the dataset. The processed image is used as input, and the corresponding ideal target highlighting effect annotation is used as output. Through multiple iterative training, the model learns the intrinsic relationship between different segmentation and slope settings and the target highlighting effect, and finally obtains the data model. The data model can intelligently recommend the appropriate number of segments and the stretching slope of each segment based on the input image features, thereby improving the effect of the infrared thermal imaging high-temperature target monitoring method around the wind turbine cabin.

[0081] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for monitoring high-temperature targets by infrared thermal imaging around the perimeter of a fan cabin, characterized in that: The specific steps include: S1. Obtain an image output by the infrared detector and subjected to non-uniformity correction and filtering noise reduction, and use it as the original image; S2. Perform grayscale histogram statistics on the original image to obtain a histogram N1; S3. Perform mean filtering on the histogram N1 to obtain a filtered grayscale histogram N2; S4. According to the characteristics of the scene image, the trough points in the grayscale histogram N2 are detected, and the points whose grayscale is less than the trough grayscale in the grayscale histogram N1 are removed to obtain the grayscale histogram N3; S5. The grayscale histogram N3 after the points are removed is stretched in a piecewise linear manner, and the weight ratio of the high temperature segment is increased during the stretching process; S6. outputting an image with a high temperature target highlighting effect; S7. Use convolutional neural network to improve the algorithm effect.

2. The method for monitoring high temperature targets by infrared thermal imaging around the perimeter of a fan cabin according to claim 1 is characterized in that: In step S1, the image output by the detector is I(x, y), where (x, y) represents the pixel coordinates in the image. The non-uniformity correction aims to eliminate the problem of inconsistent responses of each pixel of the detector. The filtering noise reduction uses a suitable filtering algorithm to remove noise interference in the image. A filter with a filter window width of K is used, and K takes a value of 1 or 2 to obtain a relatively pure original image I0(x, y): Where G(i,j,σ) is the filter kernel; σ is the Gaussian filter width. The larger the σ is, the wider the frequency band of the Gaussian filter is and the better the smoothness is. The value range of i and j is the filter module width (-K,K).

3. The method for monitoring high temperature targets by infrared thermal imaging around the perimeter of a fan cabin according to claim 2 is characterized in that: In step S2, the gray level of each point of the original image I0(x, y) is statistically counted. Assuming that the gray level of the image is (0 - 16384), the number of pixels n with gray value k, where 0 < k < L, in the image is counted k , so as to obtain the gray histogram N1(k) = n k .

4. The method for monitoring high temperature targets by infrared thermal imaging around the perimeter of a fan cabin according to claim 3 is characterized in that: In step S3, assuming that the filter window size is w, for each N1 grayscale value in the histogram, the grayscale value statistics N2(k) after filtering are calculated as follows: Among them, when i<0 or i>L, N1(i)=0.

5. The method for monitoring high temperature targets by infrared thermal imaging around the perimeter of a fan cabin according to claim 4, characterized in that: In step S4, ΔN2(k)=N2(k+1)-N2(k). When ΔN2(k-1)>0 and ΔN2(k)<0, point k may be the trough point k. min ; After finding the trough point, the grayscale value in the histogram N1 is less than the grayscale k of the trough min Set all points to zero and get histogram N3(k):

6. A method for monitoring high temperature targets by infrared thermal imaging around the perimeter of a fan cabin according to claim 5, characterized in that: In step S5, the grayscale histogram N3 is divided into multiple segments, a low grayscale segment [0, k1], a medium-low grayscale segment [k1, k2], and a high grayscale segment [k2, L]. The stretching slopes of the low grayscale segment, the medium grayscale segment, and the high grayscale segment are respectively m1, m2, and m3, m3>m2>m1, and each segment is linearly stretched with a different slope. By increasing the stretching slope of the high temperature segment, the contrast between high and low temperatures is improved, that is:

7. A method for monitoring high temperature targets by infrared thermal imaging around the perimeter of a fan cabin according to claim 6, characterized in that: The processing process of step S7 is as follows: S71. Collect a large number of infrared images of the perimeter of the wind turbine nacelle processed with different numbers of segments and different stretching slopes as a data set, and perform unified preprocessing on the collected image data sets; S72. Establish a model and perform training; S73. Model evaluation and optimization.

8. The method for monitoring high temperature targets by infrared thermal imaging around the perimeter of a fan cabin according to claim 7, characterized in that: The model adopts a convolutional neural network architecture based on the ResNet architecture, which includes multiple convolutional layers, pooling layers, and fully connected layers. The convolutional layer extracts features by sliding the convolution kernel on the image; the pooling layer is used to downsample the feature map; the fully connected layer is used to integrate the extracted features and finally output results related to the target.

9. The method for monitoring high-temperature targets by infrared thermal imaging around the perimeter of a fan cabin according to claim 8, characterized in that: The specific training process is: The processed image data is taken as input, and the corresponding ideal target highlighting effect annotation is taken as output. A batch of images is input each time, and the output result of the network is calculated through forward propagation, and then compared with the real label to calculate the value of the loss function; Using the back-propagation algorithm, according to the value of the loss function, the error is back-propagated from the output layer to the input layer, and the weight parameters of each layer in the network are adjusted to gradually reduce the loss function and continuously optimize the performance of the network. After multiple iterative training, the model learns the intrinsic relationship between different segmentation and slope settings and the target highlighting effect.

10. The method for monitoring high temperature targets by infrared thermal imaging around the perimeter of a fan cabin according to claim 7, characterized in that: In step S73, the performance of the model is monitored using the validation set to observe the changes of the model on the validation set to avoid overfitting of the model on the training set. If overfitting occurs, the model is optimized using a regularization method to adjust the network structure or hyperparameters. After training is completed, the final model is evaluated using the test set to check the actual performance of the model on data that has not participated in training, verify the generalization ability of the model, and ensure that it can accurately recommend the appropriate number of segments and the stretching slope of each segment on the actual infrared image of the wind turbine cabin perimeter.

Images