Infrared thermal imaging temperature measurement method for dynamic water mist interference scene

Abstract

The invention discloses an infrared thermal imaging temperature measurement method for a dynamic water mist interference scene, and the method comprises the steps: synchronously obtaining an infrared thermal image and a visible light image of a measured object through a thermal infrared imager and a visible light camera, and extracting a region of interest where the measured object is located based on the temperature difference between the region of interest and a background region, and the temperature value of the measured object is obtained. In consideration of an error caused by dynamic water mist to infrared temperature measurement, the interference of channel difference contrast characteristics to the dynamic water mist is quantified from the color and texture characteristics of a visible light image of a measured object for the first time. The intelligent compensation model is used for estimating the water mist transmissivity of the visible light wave band and correcting the water mist transmissivity to the infrared wave band for the first time, and the temperature measurement error is directly predicted, so that the temperature measurement result is compensated. The problems that the thermal infrared imager is difficult in mechanism modeling analysis and difficult in direct and accurate temperature detection under dynamic water mist interference are solved, and real-time online high-accuracy temperature detection of a detected object under dynamic water mist interference is realized.

Description

Technical Field

[0001] The present invention belongs to the field of infrared thermal imaging temperature measurement, and specifically relates to an infrared thermal imaging temperature measurement method for dynamic water mist interference scenes. Background Art

[0002] Temperature is a parameter that must be measured in many industrial processes or scientific experiments. It is an important data support for temperature control, quality assessment, and safe production. Infrared thermal imagers are now widely used in industries such as steel, metallurgy, and power electronics because of their advantages such as non-contact temperature measurement, no harmful radiation, and real-time online estimation. However, due to the non-contact temperature measurement method of infrared thermal imagers, interference factors such as the emissivity of the surface of the object being measured, the temperature measurement distance, viewing angle, and transmittance will affect its temperature measurement accuracy.

[0003] At present, there are few studies on the influence of interfering media between infrared thermal imagers and measured objects. However, water mist is a common interfering medium between infrared thermal imagers and measured objects in industrial production processes. For example, engine blade temperature measurement, high-temperature steel billet surface temperature measurement, hot-rolled steel strip surface temperature measurement, and rotary kiln in sintering process will be affected by water mist interference factors, resulting in large deviations in temperature measurement results, limiting the application of infrared thermal imagers. Most of the existing infrared temperature measurement methods for medium interference are based on the radiation temperature measurement mechanism, and an infrared temperature measurement model for interfering media is established. In addition, the parameters in the model need to be estimated, which makes it difficult to directly apply to complex industrial temperature measurement scenarios.

[0004] Therefore, the present invention proposes an infrared thermal imaging temperature measurement method and system for dynamic water fog interference scenes, designs a temperature measurement system that combines infrared vision and visible light vision, extracts water fog-related features such as hue difference, multi-scale local maximum saturation, multi-scale local maximum brightness and channel difference contrast from the visible light image of the measured object, realizes the quantification of dynamic water fog transmittance, and then constructs a correction model between the water fog transmittance in the visible light band and the water fog transmittance in the infrared band. By establishing an infrared temperature measurement intelligent compensation model, the mapping relationship between water fog related features and infrared temperature measurement errors is directly constructed, and the compensation of erroneous infrared thermal imager temperature measurement results under water fog interference is realized, the temperature measurement accuracy of the infrared thermal imager is improved, and the problem of limited application of infrared thermal imagers in dynamic water fog interference scenes is overcome.

[0005] The present application is different from the prior art as follows:

[0006] Compared with application number 202211583199.6:

[0007] CN 116046176A Method and system for online compensation of infrared thermal imaging temperature measurement error against water mist interference

[0008] This invention proposes an infrared thermal imaging temperature measurement error compensation method for water mist interference. The water mist transmittance is estimated a priori through the dark channel, and the water mist transmittance in the infrared light band is determined in combination with the Lambert-Beer law. A temperature compensation model is constructed to correct the temperature measurement error.

[0009] However, in fact, there are calculation errors in the water mist transmittance estimated by the dark channel prior. The Lambert-Beer law cannot accurately determine the water mist transmittance in the infrared light band, resulting in error accumulation, and it is difficult to use it for temperature compensation in dynamic water mist interference scenes.

[0010] Compared with application number 202310820871.7:

[0011] CN 117073847 A Temperature field online detection method and system based on multi-source vision under water mist interference

[0012] The invention utilizes a two-stage water mist transmittance field estimation method and establishes a detection model based on the infrared temperature measurement principle to achieve temperature compensation.

[0013] This invention requires the prior measurement of the constant term in the relative water mist concentration in order to achieve online temperature compensation, and the compensation method based on the infrared temperature measurement principle can only be applied to the scene where the water mist temperature is equal to the ambient temperature, making it difficult to achieve intelligent online compensation of infrared temperature measurement. Summary of the invention

[0014] In order to solve the above technical problems, the present invention proposes an infrared thermal imaging temperature measurement method for dynamic water mist interference scenes. The temperature measurement method proposed in the present invention can compensate in real time online for the temperature measurement results obtained by the infrared thermal imager when there is complex dynamic water mist interference in the optical path. There is no need to model the infrared radiation temperature measurement mechanism of water mist interference, and the method has the advantages of being intelligent, fast, and easy to implement.

[0015] To achieve the above object, the technical solution adopted by the present invention is:

[0016] An infrared thermal imaging temperature measurement method for dynamic water fog interference scenes includes the following steps:

[0017] (1) Using an infrared thermal imager and a visible light camera to synchronously acquire an infrared thermal image and a visible light image of the object under test, and performing multimodal image registration on the obtained infrared thermal image and visible light image;

[0018] (2) Based on the color and texture characteristics of the visible light image, the water mist-related features such as hue difference, multi-scale local maximum saturation, multi-scale local maximum brightness and channel difference contrast are calculated to identify and quantify the dynamic water mist interference between the object under test and the infrared thermal imager optical path;

[0019] (3) A two-stage infrared band water mist transmittance estimation method is proposed. First, the water mist transmittance in the visible light band is roughly estimated based on the quantitative characteristics of dynamic water mist interference. Then, a prediction model is established with the water mist transmittance, saturation, brightness and channel difference contrast of the object in the visible light band as input and the infrared band transmittance as output. The water mist transmittance in the visible light band is corrected to the infrared band.

[0020] (4) An intelligent compensation model for infrared temperature measurement is constructed, which can automatically identify whether the temperature measurement results of the infrared thermal imager are interfered by water mist and perform temperature compensation for the erroneous measurement results.

[0021] As a further improvement of the present invention, the specific process of step (1) is as follows:

[0022] a) Obtain infrared thermal images and visible light images of the object being measured, and calculate the temperature value of the area of ​​interest;

[0023] Install the infrared thermal imager and visible light camera together in front of the object to be measured, maintain a suitable measuring distance and focal length so that the object to be measured is in the middle of the infrared thermal image and visible light image as much as possible and occupies a large field of view;

[0024] Use an infrared thermal imager and a visible light camera to synchronously obtain infrared thermal images and visible light images of the object being measured for subsequent infrared temperature measurement;

[0025] Based on the temperature difference between the region of interest and the background area, the following algorithm is proposed to locate the region of interest where the object is located and extract the temperature value;

[0026] Step 1: Read the original infrared thermal image of the object to be measured, calculate the temperature histogram of the original infrared thermal image, and sort it in descending order according to the temperature value. Use the Otsu method to calculate the segmentation threshold T of the infrared thermal image. threshold , traverse the original infrared thermal image and retain the temperature value greater than T threshold The pixel points are considered to be the area of ​​interest of the object under test.

[0027] Step 2: After locating the area of ​​interest of the object being measured, sort the temperature values ​​in the area of ​​interest in descending order and take the average temperature T in the upper half of the temperature distribution. avg As the raw temperature measurement value of the object being measured;

[0028] b) Multimodal image registration;

[0029] According to the infrared thermal image and visible light image pair obtained in a), a feature-based method is used for image registration. After key point detection, the points with significance and stability in the image are identified to provide a basis for subsequent feature extraction. The areas around the detected key points are encoded in combination with feature description to generate representative feature vectors so that the infrared thermal image and the visible light image have comparability and feature matching. After that, the best matching feature point pair is found by comparing the feature descriptors in different images, and the homography matrix is ​​calculated to obtain the transformed infrared thermal image and visible light image pair.

[0030] As a further improvement of the present invention, the specific process of step (2) is as follows:

[0031] a) Water mist interference identification;

[0032] Use a visible light camera to identify whether there is water fog interference in the light path. Let the visible light image interfered by water fog be I. According to the characteristics of hue difference and channel difference mean, a reasonable threshold can be set to realize the identification of water fog interference.

[0033] 1) Hue difference;

[0034]

[0035] Where c∈{r,g,b} is the RGB channel of the visible light image; h is the hue channel; H(I) is the hue difference;

[0036] The hue difference indicates whether there is water mist interference in a certain area of ​​the image. It can be obtained by calculating the hue channel difference between the original image and its semi-inverse image. When the pixels in the image are not interfered by water mist, the hue difference changes very little, while when the pixels in the image are interfered by water mist, the hue difference changes greatly, which can be used to detect whether there is water mist interference.

[0037] 2) Channel difference mean;

[0038]

[0039] In the formula, max is the maximum value of each pixel position of the visible light image I on the RGB channel; min is the minimum value of each pixel position of the visible light image I on the RGB channel; mean is the mean operation;

[0040] The channel difference mean is used to characterize whether there is water fog interference in the current visible light image. It can be obtained by calculating the mean of the difference between the maximum channel and the minimum channel of the original image. When there is no water fog interference in the image, the channel difference mean is large, and when there is water fog interference in the image, the channel difference mean will become smaller and close to 0;

[0041] b) Quantification of water mist interference:

[0042] When there is no water mist on the optical path between the visible light camera and the object to be measured, it is not affected by the water mist, and the brightness and texture features of the object to be measured are not destroyed; when there is water mist on the optical path between the visible light camera and the object to be measured, the visible light imaging is affected by the water mist, and the reflected light of the object to be measured is affected by the suspended particles of the water mist. The proportion of global atmospheric light participating in the imaging increases, resulting in higher brightness, lower saturation and more blurred texture in the area with denser water mist in the visible light image;

[0043] 1) Multi-scale channel differences;

[0044]

[0045] In the formula, Ω r (x) is a local block with x as the center and radius r; f r (x) is the multi-scale channel difference;

[0046] Multi-scale channel difference characterizes the concentration of water mist in a certain area of ​​the image. It is a metric parameter used to describe the difference between the maximum channel and the minimum channel in the image. For the light fog area, there is a significant difference between the maximum channel and the minimum channel, and the corresponding channel difference is large. For the dense fog area, due to the serious decrease in contrast, the channel difference tends to 0, and different water mist concentration areas can be adaptively distinguished.

[0047] 2) Multi-scale dark channel;

[0048]

[0049] In the formula, Ω r (x) is a local block with x as the center and radius r; D r (x) is a multi-scale dark channel;

[0050] The multi-scale dark channel represents the minimum value of all pixel colors in a local block of an image. It is a metric parameter used to describe the water mist content in a local area of ​​an image. The size of the local block affects the characterization performance of the dark channel. If the local block is too large, it will lead to excessive defogging, and if the local block is too small, it will not be able to play the role of feature description.

[0051] 3) Multi-scale local gradients;

[0052]

[0053] In the formula, Ω r (x) is a local block with x as the center and radius r; is the gradient map of the visible light image; T r (x) is the multi-scale local gradient;

[0054] The gradient represents the intensity change of a certain point in the image. It is a metric parameter used to describe the rate and direction of change of pixel intensity in the image. It can be obtained by taking the derivative of the image and expressed in vector form. The larger the gradient amplitude, the greater the intensity change of a certain pixel point in the image along the gradient direction. When the image is affected by water mist, the texture features in the visible light image change from complex uneven distribution to relatively uniform distribution, and the gradient decreases.

[0055] 4) Multi-scale local maximum contrast;

[0056]

[0057] In the formula, Ω r (x) is a local block with x as the center and radius r; Ω s (y) is a local block with y as the center and radius s; C r (x) is the multi-scale local maximum contrast;

[0058] The local contrast is the variance of the pixel intensity in the local block of the image and the central pixel. The local maximum is further calculated in this area to obtain the multi-scale local maximum contrast. When the image is affected by water fog, the visibility of the corresponding area decreases and the local maximum contrast also decreases.

[0059] 5) Multi-scale local maximum saturation;

[0060]

[0061] In the formula, Ω r (x) is a local block with x as the center and radius r; S r (x) is the multi-scale local maximum saturation;

[0062] Saturation represents the purity or intensity of color. It is a metric parameter that describes the vividness of color in the HSV color space. Its value range is 0 to 1. The greater the saturation, the greater the intensity of the color. When the image is affected by water mist, the color intensity of the visible light image will be weakened and the saturation will decrease due to the interference of suspended particles in the water mist.

[0063] 6) Multi-scale local maximum brightness;

[0064]

[0065] In the formula, Ω r (x) is a local block with x as the center and radius r; L I is the brightness map of the visible light image; L r (x) is the multi-scale local maximum brightness;

[0066] Brightness represents the brightness of the color and affects the brightness of the color visually. Its value range is 0 to 1. When the image is affected by water mist, the visible light image will change from normal brightness to very bright, and the local maximum brightness of the image will also increase accordingly.

[0067] 7) Channel difference contrast;

[0068] R c (x) = 255 - I c (x)(10)

[0069]

[0070] In the formula, R c (x) is the inverted image of the visible light image; Con(x) is the channel difference contrast;

[0071] Channel difference contrast is the ratio of the difference between the maximum channel and the minimum channel of the image to the maximum channel of the inverse image. Its value range is 0 to 1. When the image is affected by water mist, the brightness of the blurred area increases, and the corresponding channel difference will also increase. At the same time, the maximum channel of the inverse image obtained by exchanging the light and dark areas will become smaller, so the channel difference contrast will increase, and vice versa.

[0072] As a further improvement of the present invention, the specific process of step (3) is as follows:

[0073] a) Stacked asymmetric autoencoders;

[0074] The mapping function used by the encoder can be expressed as follows;

[0075]

[0076] In the formula, h is the hidden layer after encoding; s1 and s2 are sigmoid activation functions; W1, W2, b1 and b2 are weight coefficients and intercept terms respectively; is the input variable;

[0077] The decoder reconstructs the hidden layer h into the original input through a mapping function. The mapping function used by the decoder can be expressed as follows:

[0078]

[0079] In the formula, is the reconstructed original input after decoding; s3 is the sigmoid activation function; W3 and b3 are the weight coefficient and intercept term respectively;

[0080] The asymmetric autoencoder optimizes the model parameters through the loss function L in the following formula;

[0081]

[0082] Where m is the number of input variables.

[0083] The asymmetric autoencoder can be regarded as a three-layer neural network consisting of an encoder and a decoder. At the same time, the size of the output layer is the same as the size of the input layer. The asymmetric autoencoder performs unsupervised training on the encoder by making the reconstructed output as equal to the input as possible, thereby improving the feature extraction ability of the model.

[0084] A feature extraction layer is added during stacking connection so that the next asymmetric autoencoder can better learn the output of the previous asymmetric autoencoder and improve the prediction ability of the model. The feature extraction layer can be expressed as follows;

[0085] k=sigmoid(W4f in )(15)

[0086] f out =ReLU(k⊙(W5f in ))(16)

[0087] Where k is the weight vector; ⊙ is the element multiplication operation; W4 and W5 are weight coefficients; f in and f out are the input feature vector and the output feature vector respectively;

[0088] For the hidden layer features output by the asymmetric autoencoder, two learnable weight vectors of equal size are used to weight the input features respectively. One branch uses the sigmoid function as the activation function, performs element-wise multiplication with the other branch, and then uses the ReLU activation function to obtain the final output features.

[0089] b) Two-stage infrared band water mist transmittance estimation;

[0090] According to the dynamic water fog interference quantification method proposed in (2), due to the influence of water fog suspended particles in the air, when the water fog concentration is higher, the brightness in the corresponding area is higher and the saturation is lower. Therefore, the blurred area in the foggy image has the characteristics of high brightness and low saturation. In general, the water fog concentration will increase with the change of scene depth. The above phenomenon can be expressed by the following formula:

[0091] d(x)∝c(x)∝function(f r (x),D r (x),T r (x),C r (x),S r (x),L r (x),Con(x))(17)

[0092] Where d(x) is the scene depth; c(x) is the water mist concentration;

[0093] The relationship between transmittance and scene depth is as follows:

[0094] t(x)=e -βd(x) (18)

[0095] Where t(x) is the transmittance in the visible light band; d(x) is the scene depth; β is the atmospheric scattering coefficient;

[0096] Therefore, by quantifying the multi-scale channel differences, dark channels, gradients and other features, a rough estimate of the water mist transmittance in the visible light band can be made;

[0097] In addition, according to Lambert-Beer's law, the mass extinction coefficient of water mist is related to the wavelength;

[0098] t wm (λ)=exp[-β′d′C′k′(λ)](19)

[0099] Where, t wm (λ) is the water mist transmittance in the visible light band; β′ is the proportionality coefficient; d′ is the influence distance of the water mist; C′ is the relative water mist concentration; k′(λ) is the mass extinction coefficient of the water mist;

[0100] It can be seen from formula (17) that the saturation S and brightness V of the visible light image are related to most of the quantitative features of water mist concentration, and the channel difference contrast feature can quantify the water mist concentration. The water mist transmittance in the visible light band is corrected to the infrared band through the features such as saturation, brightness and channel difference contrast.

[0101] t inf (λ1)=function(t vis (λ2),S,V,Con)(20)

[0102] Where, t inf (λ1) is the water mist transmittance in the infrared band; t vis (λ2) is the water mist transmittance in the visible light band; λ1 and λ2 are the imaging bands of the infrared thermal imager and the visible light camera, respectively; S is the saturation; V is the brightness; Con is the channel difference contrast;

[0103] Saturation S and brightness V are defined as follows:

[0104]

[0105] V=max(R,G,B)(22)

[0106] Where R, G and B correspond to the red, green and blue color channels in the image respectively.

[0107] As a further improvement of the present invention, the specific process of step (4) is as follows:

[0108] Step 1: Collect infrared thermal images and visible light images of the object under test at different temperatures, different scenes, and different water mist concentrations. At the same time, collect the real temperature of the object under test and the measured temperature affected by water mist, and calculate the true value of the water mist transmittance label in the infrared band according to the following formula;

[0109]

[0110] In the formula, t label is the true value of the water mist transmittance label; T effect , T truth and T u are the measured temperature affected by water mist, the real temperature of the measured object and the ambient temperature respectively; n is the fitting coefficient related to the detector material of the infrared thermal imager;

[0111] Randomly extract two-thirds of the collected paired infrared thermal images and visible light image datasets as training sets, and the rest of the datasets as test sets;

[0112] Step 2: Establish a prediction model based on stacked asymmetric autoencoders to extract the quantitative features of water mist concentration in the visible light image of the object under test, convert the visible light image into the HSV color space, extract the saturation and brightness channels, and use the proposed two-stage infrared band water mist transmittance estimation method to obtain the infrared band water mist transmittance;

[0113] Step 3: Input the water mist transmittance, saturation, brightness, channel difference contrast and measured temperature in the infrared band into the prediction model for training, and test the prediction effect of the model on the temperature measurement error on the test set;

[0114] Step 4: Use the trained intelligent compensation model to compensate the measured temperature of the infrared thermal imager under the interference of water fog;

[0115] It is expressed by the following formula:

[0116]

[0117] ΔT=function(t inf ,S,V,Con,T origin )(25)

[0118] In the formula, T and T origin are the output temperature and the original temperature measurement result respectively; ΔT is the prediction error;

[0119] The loss function constraint based on physical information ensures that the intelligent model can accurately estimate the water mist transmittance in the infrared band while also estimating the measurement error. The loss function of the intelligent compensation model composed of physical loss and label loss can be expressed as follows:

[0120]

[0121] L total =L label +L physi (28)

[0122] Where, L total , L label and L physi They are the total loss function, label loss and physical loss respectively; m is the number of input variables; is the predicted water mist transmittance in the infrared band; T i ′、T i and T u are the affected measured temperature, true temperature and ambient temperature respectively; and are the predicted error and the true error respectively;

[0123] The collected infrared thermal images and visible light images are divided into data sets, and the established intelligent compensation model is trained using the training set. The temperature measurement error under water mist interference is predicted, thereby achieving accurate infrared temperature measurement under dynamic water mist interference.

[0124] Beneficial effects: Compared with the prior art, the technical solution of the present invention has the following beneficial technical effects:

[0125] (1) A channel difference contrast feature is proposed, which can calculate the channel difference and the ratio of the maximum channel of the inverted image based on the visible light image of the object to be measured, thereby quantifying the dynamic water mist interference in the optical path;

[0126] (2) A deep learning module based on stacked asymmetric autoencoders is proposed, and a feature extraction module based on one-dimensional convolution is combined with the stacked asymmetric autoencoder network to improve the prediction performance of the model;

[0127] (3) A two-stage infrared band water mist transmittance correction module based on features such as brightness, saturation and channel difference contrast is proposed to correct the water mist transmittance in the visible light band and obtain the water mist transmittance in the infrared band online;

[0128] (4) A knowledge-guided infrared temperature measurement intelligent compensation model was constructed, and a loss function constraint based on physical information was proposed. The physical loss and label loss together constitute the loss function of the intelligent compensation model. While reducing the loss of physical information, the model's prediction performance for water mist transmittance in the infrared band was improved, thereby improving the model's prediction accuracy for temperature measurement errors and achieving accurate online temperature detection under dynamic water mist interference.

[0129] (5) An infrared thermal imaging temperature measurement method for dynamic water fog scenes is proposed. By combining the advantages of infrared visual temperature measurement and visible light visual recognition of water fog interference, the surface temperature of the object being measured under dynamic water fog interference can be accurately obtained. BRIEF DESCRIPTION OF THE DRAWINGS

[0130] Figure 1 It is a temperature measurement system diagram of the present invention;

[0131] Figure 2 It is a diagram of the temperature measurement method of the present invention;

[0132] Figure 3 It is a regression prediction result diagram of the present invention;

[0133] Figure 4 is the absolute error before and after compensation of the temperature measurement method of the present invention;

[0134] Figure 5 It is the temperature measurement result before and after compensation of the temperature measurement method of the present invention. DETAILED DESCRIPTION

[0135] The following is a detailed description of the technical solution of the application in conjunction with the accompanying drawings. The described embodiments are only part of the embodiments involved in this patent. All non-innovative embodiments of other researchers in this field based on this embodiment are within the scope of protection of this patent.

[0136] The present invention proposes an infrared thermal imaging temperature measurement method and system for dynamic water fog interference scenes. Figure 1 It is a schematic diagram of an online temperature detection system with intelligent collaboration of an infrared thermal imager and a visible light camera, including an infrared thermal imager and a visible light camera. Figure 2 The infrared thermal imaging temperature measurement method proposed in the present invention includes the following steps:

[0137] (1) An infrared thermal imager and a visible light camera are used to synchronously acquire the infrared thermal image and the visible light image of the object under test, and multimodal image registration is performed on the obtained infrared thermal image and the visible light image.

[0138] (2) Based on the color and texture characteristics of visible light images, the characteristics related to water mist, such as hue difference, multi-scale local maximum saturation, multi-scale local maximum brightness, and channel difference contrast, are calculated to identify and quantify the dynamic water mist interference between the object under test and the infrared thermal imager optical path.

[0139] (3) A two-stage infrared band water mist transmittance estimation method is proposed. First, the water mist transmittance in the visible light band is roughly estimated based on the quantitative characteristics of dynamic water mist interference. Then, a prediction model is established with the water mist transmittance, saturation, brightness and channel difference contrast of the object in the visible light band as input and the infrared band transmittance as output. The water mist transmittance in the visible light band is corrected to the infrared band.

[0140] (4) An intelligent compensation model for infrared temperature measurement is constructed, which can automatically identify whether the temperature measurement results of the infrared thermal imager are interfered by water mist and perform temperature compensation for the erroneous measurement results.

[0141] The specific implementation plan is as follows:

[0142] (1) Obtain infrared thermal images and visible light images of the object to be measured, and locate the region of interest where the object is located for multimodal image registration

[0143] a) Obtain infrared thermal images and visible light images of the object being measured, and calculate the temperature value of the area of ​​interest

[0144] Install the infrared thermal imager and visible light camera together in front of the object to be measured, maintain a suitable measuring distance and focal length so that the object to be measured is in the middle of the infrared thermal image and visible light image as much as possible and occupies a larger field of view.

[0145] Use an infrared thermal imager and a visible light camera to synchronously acquire infrared thermal images and visible light images of the object being measured for subsequent infrared temperature measurement.

[0146] Since the area of ​​interest accounts for a small proportion in the infrared thermal image, and because the resolution, installation position, and field of view of the visible light camera and the infrared thermal imager are different, the rich color texture information in the visible light image does not correspond one-to-one to the temperature value in the infrared thermal image. It is necessary to locate the area of ​​interest where the object to be measured is located. Based on the temperature difference between the area of ​​interest and the background area, the present invention proposes the following algorithm to locate the area of ​​interest where the object to be measured is located and extract the temperature value.

[0147] Step 1: Read the original infrared thermal image of the object to be measured, calculate the temperature histogram of the original infrared thermal image, and sort it in descending order according to the temperature value. Use the Otsu method to calculate the segmentation threshold T of the infrared thermal image. threshold , traverse the original infrared thermal image and retain the temperature value greater than T thresholdThe area composed of these pixels is considered to be the area of ​​interest of the object being measured.

[0148] Step 2: After locating the area of ​​interest of the object being measured, sort the temperature values ​​in the area of ​​interest in descending order and take the average temperature T in the upper half of the temperature distribution. avg As the original temperature measurement value of the object being measured.

[0149] b) Multimodal image registration

[0150] According to the infrared thermal image and visible light image pair obtained in a), a feature-based method is used for image registration. After key point detection, the points with significance and stability in the image are identified to provide a basis for subsequent feature extraction. The areas around the detected key points are encoded in combination with feature description to generate representative feature vectors so that the infrared thermal image and the visible light image have comparability and feature matching. After that, the best matching feature point pair is found by comparing the feature descriptors in different images, and the homography matrix is ​​calculated to obtain the transformed infrared thermal image and visible light image pair.

[0151] (2) Dynamic water mist interference identification and quantification

[0152] a) Water mist interference identification

[0153] Since there is little texture detail information in the infrared thermal image, it is difficult to identify whether the temperature measurement result of the infrared thermal imager is interfered by water mist. Therefore, the present invention proposes to use a visible light camera to identify whether there is water mist interference in the optical path. Let the visible light image interfered by water mist be I. According to the hue difference and channel difference mean characteristics, a reasonable threshold can be set to realize the identification of water mist interference.

[0154] 1) Hue Difference

[0155]

[0156] Where c∈{r,g,b} is the RGB channel of the visible light image; h is the hue channel; and H(I) is the hue difference.

[0157] The hue difference indicates whether there is water mist interference in a certain area of ​​the image, which can be obtained by calculating the hue channel difference between the original image and its semi-inverse image. When the pixels in the image are not disturbed by water mist, the hue difference changes little, while when the pixels in the image are disturbed by water mist, the hue difference changes greatly, which can be used to detect whether there is water mist interference.

[0158] 2) Channel Difference Mean

[0159]

[0160] Wherein, max is the maximum value of each pixel position of the visible light image I on the RGB channel; min is the minimum value of each pixel position of the visible light image I on the RGB channel; mean is the mean operation.

[0161] The channel difference mean is used to characterize whether the current visible light image has water fog interference, which can be obtained by calculating the mean of the difference between the maximum channel and the minimum channel of the original image. When there is no water fog interference in the image, the channel difference mean is larger, and when there is water fog interference in the image, the channel difference mean will become smaller and close to 0.

[0162] b) Quantification of water mist interference

[0163] When there is no water mist on the optical path between the visible light camera and the object to be measured, it is not affected by the water mist, and the brightness and texture characteristics of the object to be measured are not destroyed; when there is water mist on the optical path between the visible light camera and the object to be measured, the visible light imaging is affected by the water mist, and the reflected light of the object to be measured is affected by the suspended particles in the water mist. The proportion of global atmospheric light participating in the imaging increases, resulting in higher brightness, lower saturation and more blurred texture in areas with denser water mist in the visible light image.

[0164] 1) Multi-scale channel differences

[0165]

[0166] In the formula, Ω r (x) is a local block with x as the center and radius r; f r (x) is the multi-scale channel difference.

[0167] The multi-scale channel difference characterizes the concentration of water mist in a certain area of ​​the image, and is a metric parameter used to describe the difference between the maximum channel and the minimum channel in the image. For light fog areas, there is a significant difference between the maximum channel and the minimum channel, and the corresponding channel difference is large. For dense fog areas, due to the severe decrease in contrast, the channel difference tends to 0, and different water mist concentration areas can be adaptively distinguished.

[0168] 2) Multi-scale dark channel

[0169]

[0170] In the formula, Ω r (x) is a local block with x as the center and radius r; D r (x) is the multi-scale dark channel.

[0171] The multi-scale dark channel represents the minimum value of all pixel colors in a local block of an image, and is a metric parameter used to describe the water mist content in a local area of ​​the image. The size of the local block will affect the characterization performance of the dark channel. If the local block is too large, it will lead to excessive defogging, and if the local block is too small, it will not be able to play the role of feature description.

[0172] 3) Multi-scale local gradients

[0173]

[0174] In the formula, Ω r (x) is a local block with x as the center and radius r; is the gradient map of the visible light image; T r (x) is the multi-scale local gradient.

[0175] The gradient represents the intensity change of a certain point in the image. It is a metric parameter used to describe the rate and direction of change of pixel intensity in the image. It can be obtained by taking the derivative of the image and expressed as a vector. The larger the gradient amplitude, the greater the intensity change of a certain pixel point in the image along the gradient direction. When the image is affected by water mist, the texture features in the visible light image change from a complex uneven distribution to a more uniform distribution, and the gradient decreases.

[0176] 4) Multi-scale local maximum contrast

[0177]

[0178] In the formula, Ω r (x) is a local block with x as the center and radius r; Ω s (y) is a local block with y as the center and radius s; C r (x) is the multi-scale local maximum contrast.

[0179] Local contrast is the variance of the pixel intensity in the local block of the image and the central pixel. The local maximum is further calculated in this area to obtain the multi-scale local maximum contrast, which is a metric parameter used to describe the visibility of the image in the local area. When the image is affected by water fog, the visibility of the corresponding area decreases and the local maximum contrast also decreases.

[0180] 5) Multi-scale local maximum saturation

[0181]

[0182] In the formula, Ω r (x) is a local block with x as the center and radius r; S r (x) is the multi-scale local maximum saturation.

[0183] Saturation represents the purity or intensity of color. It is a metric parameter that describes the vividness of color in the HSV color space. Its value range is 0 to 1. The greater the saturation, the greater the intensity of the color. When the image is affected by water mist, the color intensity of the visible light image will be weakened and the saturation will decrease due to the interference of suspended particles in the water mist.

[0184] 6) Multi-scale local maximum brightness

[0185]

[0186] In the formula, Ω r (x) is a local block with x as the center and radius r; L I is the brightness map of the visible light image; L r (x) is the multi-scale local maximum brightness.

[0187] Brightness represents the brightness of the color and affects the brightness of the color visually, and its value range is 0 to 1. When the image is affected by water mist, the visible light image will change from normal brightness to very bright, and the local maximum brightness of the image will also increase accordingly.

[0188] 7) Channel Difference Contrast

[0189] R c (x) = 255 - I c (x)(10)

[0190]

[0191] In the formula, R c (x) is the inverted image of the visible light image; Con(x) is the channel difference contrast.

[0192] Channel difference contrast is the ratio of the difference between the maximum channel and the minimum channel of the image to the maximum channel of the inverse image. Its value range is 0 to 1. When the image is affected by water mist, the brightness of the blurred area increases, and the corresponding channel difference will also increase. At the same time, the maximum channel of the inverse image obtained by exchanging the light and dark areas will become smaller, so the channel difference contrast will increase, and vice versa.

[0193] (3) Two-stage infrared band water mist transmittance estimation based on stacked asymmetric autoencoders

[0194] a) Stacked Asymmetric Autoencoders

[0195] The present invention proposes a new autoencoder model structure, which is different from the previous symmetrical encoder-decoder design method, in which the mapping function used by the encoder can be expressed by the following formula.

[0196]

[0197] In the formula, h is the hidden layer after encoding; s1 and s2 are sigmoid activation functions; W1, W2, b1 and b2 are weight coefficients and intercept terms respectively; is the input variable.

[0198] The decoder reconstructs the hidden layer h into the original input through a mapping function. The mapping function used by the decoder can be expressed as follows.

[0199]

[0200] In the formula, is the reconstructed original input after decoding; s3 is the sigmoid activation function; W3 and b3 are the weight coefficient and intercept term respectively.

[0201] The asymmetric autoencoder optimizes the model parameters through the loss function L in the following formula.

[0202]

[0203] Where m is the number of input variables.

[0204] The asymmetric autoencoder can be regarded as a three-layer neural network consisting of an encoder and a decoder. At the same time, the size of the output layer is the same as the size of the input layer. The asymmetric autoencoder performs unsupervised training on the encoder by making the reconstructed output as equal to the input as possible, thereby improving the feature extraction ability of the model.

[0205] The stacked asymmetric autoencoder is a deep network composed of multiple trained asymmetric autoencoders stacked in layers. It can learn more complex relationships between input and output layers and obtain higher-order and more abstract features of the original input. After the first asymmetric autoencoder is trained, the output of its hidden layer is used as the input of the second asymmetric autoencoder, and the error back propagation algorithm is used to train it layer by layer in the above manner. A fully connected layer is added to the trained stacked asymmetric autoencoder to establish a deep neural network that can be used for data prediction.

[0206] In addition, unlike the simple method of directly stacking and connecting asymmetric autoencoders, the present invention proposes adding a feature extraction layer during the stacking connection so that the next asymmetric autoencoder can better learn the output of the previous asymmetric autoencoder and improve the prediction ability of the model. The feature extraction layer can be represented by the following formula.

[0207] k=sigmoid(W4f in )(15)

[0208] f out =ReLU(k⊙(W5f in ))(16)

[0209] Where k is the weight vector; ⊙ is the element multiplication operation; W4 and W5 are weight coefficients; f in and f out are the input feature vector and the output feature vector respectively.

[0210] For the hidden layer features output by the asymmetric autoencoder, two learnable weight vectors of equal size are used to weight the input features respectively. One branch uses the sigmoid function as the activation function, performs element-wise multiplication with the other branch, and then uses the ReLU activation function to obtain the final output features.

[0211] b) Two-stage infrared band water mist transmittance estimation

[0212] According to the dynamic water fog interference quantification method proposed in (2), due to the influence of water fog suspended particles in the air, when the water fog concentration is higher, the brightness in the corresponding area is higher and the saturation is lower. Therefore, the blurred area in the foggy image has the characteristics of high brightness and low saturation. In addition, in general, the water fog concentration will increase with the change of scene depth. The above phenomenon can be expressed by the following formula:

[0213] d(x)∝c(x)∝function(f r (x),D r (x),T r (x),C r (x),S r (x),L r (x),Con(x))(17)

[0214] Where d(x) is the scene depth and c(x) is the water mist concentration.

[0215] The relationship between transmittance and scene depth is as follows:

[0216] t(x)=e -βd(x) (18)

[0217] Where t(x) is the transmittance in the visible light band; d(x) is the scene depth; β is the atmospheric scattering coefficient.

[0218] Therefore, by quantifying features such as multi-scale channel differences, dark channels, and gradients, a rough estimate of the water mist transmittance in the visible light band can be made.

[0219] In addition, according to the Beer-Lambert law, the mass extinction coefficient of water mist is related to the wavelength.

[0220] t wm (λ)=exp[-β′d′C′k′(λ)](19)

[0221] Where, t wm (λ) is the water mist transmittance in the visible light band; β′ is the proportionality coefficient; d′ is the influence distance of the water mist; C′ is the relative water mist concentration; k′(λ) is the mass extinction coefficient of the water mist.

[0222] The imaging bands of visible light cameras and infrared thermal imagers are different. The visible light band water mist transmittance estimated from the visible light image cannot be directly used for temperature compensation and needs to be corrected to the infrared band. It can be seen from formula (17) that the saturation S and brightness V of the visible light image are related to most of the quantitative features of water mist concentration, and the channel difference contrast feature can quantify the water mist concentration. The present invention proposes to correct the water mist transmittance of the visible light band to the infrared band through features such as saturation, brightness and channel difference contrast.

[0223] t inf (λ1)=function(t vis (λ2),S,V,Con)(20)

[0224] Where, t inf (λ1) is the water mist transmittance in the infrared band; t vis (λ2) is the water mist transmittance in the visible light band; λ1 and λ2 are the imaging bands of the infrared thermal imager and the visible light camera, respectively; S is the saturation; V is the brightness; Con is the channel difference contrast.

[0225] Saturation S and brightness V are defined as follows:

[0226]

[0227] V=max(R,G,B)(22)

[0228] Where R, G and B correspond to the red, green and blue color channels in the image respectively.

[0229] (4) Constructing an intelligent compensation model for infrared temperature measurement

[0230] Since the visible light image and infrared thermal image of the object under test affected by water fog need to be taken manually, and the dynamic water fog interference is simulated in a laboratory environment, the amount of sample data available for learning is relatively small. The stacked asymmetric autoencoder is an unsupervised learning method that can be used for small sample data, which aims to extract complex features of high-dimensional data through layer-by-layer training. The infrared thermal imaging temperature measurement method for dynamic water fog interference scenarios proposed in the present invention can intelligently identify whether the current temperature measurement result is affected by water fog based on the visible light image of the object under test, and intelligently compensate for the erroneous temperature measurement result based on the color and texture characteristics of the area affected by water fog in the visible light image. The specific steps are as follows:

[0231] Step 1: Collect infrared thermal images and visible light images of the object under test at different temperatures, different scenes, and different water mist concentrations. At the same time, collect the real temperature of the object under test and the measured temperature affected by water mist, and calculate the true value of the water mist transmittance label in the infrared band according to the following formula.

[0232]

[0233] In the formula, t label is the true value of the water mist transmittance label; T effect , T truth and T u are the measured temperature affected by water mist, the real temperature of the measured object and the ambient temperature respectively; n is the fitting coefficient related to the detector material of the infrared thermal imager.

[0234] Two-thirds of the collected paired infrared thermal images and visible light images are randomly selected as training sets, and the rest of the datasets are used as test sets.

[0235] Step 2: Establish a prediction model based on stacked asymmetric autoencoders to extract the quantitative features of water mist concentration in the visible light image of the object under test, convert the visible light image into the HSV color space, extract the saturation and brightness channels, and use the two-stage infrared band water mist transmittance estimation method proposed in (3) to obtain the infrared band water mist transmittance.

[0236] Step 3: Input the water mist transmittance, saturation, brightness, channel difference contrast and measured temperature in the infrared band into the prediction model for training, and test the model's prediction effect on temperature measurement error on the test set.

[0237] Step 4: Use the trained intelligent compensation model to compensate for the measured temperature of the infrared thermal imager under the interference of water fog.

[0238] The working mode of the infrared thermal imaging temperature measurement method and system for dynamic water fog interference scenes proposed in the present invention can be expressed by the following formula:

[0239]

[0240] ΔT=function(t inf ,S,V,Con,T origin )(25)

[0241] In the formula, T and T origin are the output temperature and the original temperature measurement result respectively; ΔT is the prediction error.

[0242] The present invention proposes a loss function constraint based on physical information to ensure that the intelligent model can accurately estimate the water mist transmittance in the infrared band while also estimating the measurement error. The loss function of the intelligent compensation model composed of physical loss and label loss can be expressed as follows:

[0243]

[0244] L total =L label +L physi (28)

[0245] Where, L total , L label and L physi They are the total loss function, label loss and physical loss respectively; m is the number of input variables; is the predicted water mist transmittance in the infrared band; T i ′、T i and T u are the affected measured temperature, true temperature and ambient temperature respectively; and are the predicted error and the true error respectively.

[0246] The collected infrared thermal images and visible light images are divided into data sets, and the established intelligent compensation model is trained using the training set. The temperature measurement error under water mist interference is predicted, thereby achieving accurate infrared temperature measurement under dynamic water mist interference.

[0247] Example 1

[0248] This embodiment uses a high-temperature blackbody furnace to generate standard infrared radiation for model training. A metal cup filled with hot water is used as the experimental object, and the temperature measurement method of the invention is applied to the object. The temperature measurement device is installed in front of the metal cup at a distance of 0.5m. Figure 1 The temperature measurement system shown.

[0249] The final prediction result is Figure 3 As shown in the figure, the absolute error is Figure 4 As shown, attached Figure 5 The invented temperature measurement method is used to compensate the temperature of the metal cup measured before and after. At the same time, a patch thermocouple is used to detect the actual temperature of the metal cup to verify the effectiveness of the temperature measurement method.

[0250] The above description is only a preferred embodiment of the present invention and does not constitute any other form of limitation to the present invention. Any modification or equivalent change made based on the technical essence of the present invention still falls within the scope of protection required by the present invention.

Claims

1. An infrared thermal imaging temperature measurement method for dynamic water fog interference scenes, characterized in that: The steps include: (1) Using an infrared thermal imager and a visible light camera to synchronously acquire an infrared thermal image and a visible light image of the object under test, and performing multimodal image registration on the obtained infrared thermal image and visible light image; (2) Based on the color and texture characteristics of the visible light image, the water mist-related features such as hue difference, multi-scale local maximum saturation, multi-scale local maximum brightness and channel difference contrast are calculated to identify and quantify the dynamic water mist interference between the object under test and the infrared thermal imager optical path; (3) A two-stage infrared band water mist transmittance estimation method is proposed. First, the water mist transmittance in the visible light band is roughly estimated based on the quantitative characteristics of dynamic water mist interference. Then, a prediction model is established with the water mist transmittance, saturation, brightness and channel difference contrast of the object in the visible light band as input and the infrared band transmittance as output. The water mist transmittance in the visible light band is corrected to the infrared band. (4) An intelligent compensation model for infrared temperature measurement is constructed, which can automatically identify whether the temperature measurement results of the infrared thermal imager are interfered by water mist and perform temperature compensation for the erroneous measurement results.

2. According to claim 1, the infrared thermal imaging temperature measurement method for dynamic water fog interference scenes is characterized by: The specific process of step (1) is as follows: a) Obtain infrared thermal images and visible light images of the object being measured, and calculate the temperature value of the area of ​​interest; Install the infrared thermal imager and visible light camera together in front of the object to be measured, maintain a suitable measuring distance and focal length so that the object to be measured is in the middle of the infrared thermal image and visible light image as much as possible and occupies a large field of view; Use an infrared thermal imager and a visible light camera to synchronously obtain infrared thermal images and visible light images of the object being measured for subsequent infrared temperature measurement; Based on the temperature difference between the region of interest and the background area, the following algorithm is proposed to locate the region of interest where the object is located and extract the temperature value; Step 1: Read the original infrared thermal image of the object to be measured, calculate the temperature histogram of the original infrared thermal image, and sort it in descending order according to the temperature value. Use the Otsu method to calculate the segmentation threshold T of the infrared thermal image. threshold , traverse the original infrared thermal image and retain the temperature value greater than T threshold The pixel points are considered to be the area of ​​interest of the object under test. Step 2: After locating the area of ​​interest of the object being measured, sort the temperature values ​​in the area of ​​interest in descending order and take the average temperature T in the upper half of the temperature distribution. avg As the raw temperature measurement value of the object being measured; b) Multimodal image registration; According to the infrared thermal image and visible light image pair obtained in a), a feature-based method is used for image registration. After key point detection, the points with significance and stability in the image are identified to provide a basis for subsequent feature extraction. The areas around the detected key points are encoded in combination with feature description to generate representative feature vectors so that the infrared thermal image and the visible light image have comparability and feature matching. After that, the best matching feature point pair is found by comparing the feature descriptors in different images, and the homography matrix is ​​calculated to obtain the transformed infrared thermal image and visible light image pair.

3. The infrared thermal imaging temperature measurement method for dynamic water fog interference scenes according to claim 1 is characterized by: The specific process of step (2) is as follows: a) Water mist interference identification; Use a visible light camera to identify whether there is water fog interference in the light path. Let the visible light image interfered by water fog be I. According to the characteristics of hue difference and channel difference mean, a reasonable threshold can be set to realize the identification of water fog interference. 1) Hue difference; Where c∈{r,g,b} is the RGB channel of the visible light image; h is the hue channel; H(I) is the hue difference; The hue difference indicates whether there is water mist interference in a certain area of ​​the image. It can be obtained by calculating the hue channel difference between the original image and its semi-inverse image. When the pixels in the image are not interfered by water mist, the hue difference changes very little, while when the pixels in the image are interfered by water mist, the hue difference changes greatly, which can be used to detect whether there is water mist interference. 2) Channel difference mean; In the formula, max is the maximum value of each pixel position of the visible light image I on the RGB channel; min is the minimum value of each pixel position of the visible light image I on the RGB channel; mean is the mean operation; The channel difference mean is used to characterize whether there is water fog interference in the current visible light image. It can be obtained by calculating the mean of the difference between the maximum channel and the minimum channel of the original image. When there is no water fog interference in the image, the channel difference mean is large, and when there is water fog interference in the image, the channel difference mean will become smaller and close to 0; b) Quantification of water mist interference: When there is no water mist on the optical path between the visible light camera and the object to be measured, it is not affected by the water mist, and the brightness and texture features of the object to be measured are not destroyed; when there is water mist on the optical path between the visible light camera and the object to be measured, the visible light imaging is affected by the water mist, and the reflected light of the object to be measured is affected by the suspended particles of the water mist. The proportion of global atmospheric light participating in the imaging increases, resulting in higher brightness, lower saturation and more blurred texture in the area with denser water mist in the visible light image; 1) Multi-scale channel differences; In the formula, Ω r (x) is a local block with x as the center and radius r; f r (x) is the multi-scale channel difference; Multi-scale channel difference characterizes the concentration of water mist in a certain area of ​​the image. It is a metric parameter used to describe the difference between the maximum channel and the minimum channel in the image. For the light fog area, there is a significant difference between the maximum channel and the minimum channel, and the corresponding channel difference is large. For the dense fog area, due to the serious decrease in contrast, the channel difference tends to 0, and different water mist concentration areas can be adaptively distinguished. 2) Multi-scale dark channel; In the formula, Ω r (x) is a local block with x as the center and radius r; D r (x) is a multi-scale dark channel; The multi-scale dark channel represents the minimum value of all pixel colors in a local block of an image. It is a metric parameter used to describe the water mist content in a local area of ​​an image. The size of the local block affects the characterization performance of the dark channel. If the local block is too large, it will lead to excessive defogging, and if the local block is too small, it will not be able to play the role of feature description. 3) Multi-scale local gradients; In the formula, Ω r (x) is a local block with x as the center and radius r; ▽I is the gradient map of the visible light image; T r (x) is the multi-scale local gradient; The gradient represents the intensity change of a certain point in the image. It is a metric parameter used to describe the rate and direction of change of pixel intensity in the image. It can be obtained by taking the derivative of the image and expressed in vector form. The larger the gradient amplitude, the greater the intensity change of a certain pixel point in the image along the gradient direction. When the image is affected by water mist, the texture features in the visible light image change from complex uneven distribution to relatively uniform distribution, and the gradient decreases. 4) Multi-scale local maximum contrast; In the formula, Ω r (x) is a local block with x as the center and radius r; Ω s (y) is a local block with y as the center and radius s; C r (x) is the multi-scale local maximum contrast; The local contrast is the variance of the pixel intensity in the local block of the image and the central pixel. The local maximum is further calculated in this area to obtain the multi-scale local maximum contrast. When the image is affected by water fog, the visibility of the corresponding area decreases and the local maximum contrast also decreases. 5) Multi-scale local maximum saturation; In the formula, Ω r (x) is a local block with x as the center and radius r; S r (x) is the multi-scale local maximum saturation; Saturation represents the purity or intensity of color. It is a metric parameter that describes the vividness of color in the HSV color space. Its value range is 0 to 1. The greater the saturation, the greater the intensity of the color. When the image is affected by water mist, the color intensity of the visible light image will be weakened and the saturation will decrease due to the interference of suspended particles in the water mist. 6) Multi-scale local maximum brightness; In the formula, Ω r (x) is a local block with x as the center and radius r; L I is the brightness map of the visible light image; L r (x) is the multi-scale local maximum brightness; Brightness represents the brightness of the color and affects the brightness of the color visually. Its value range is 0 to 1. When the image is affected by water mist, the visible light image will change from normal brightness to very bright, and the local maximum brightness of the image will also increase accordingly. 7) Channel difference contrast; R c (x)=255-I c (x)(10) In the formula, R c (x) is the inverted image of the visible light image; Con(x) is the channel difference contrast; Channel difference contrast is the ratio of the difference between the maximum channel and the minimum channel of the image to the maximum channel of the inverse image. Its value range is 0 to 1. When the image is affected by water mist, the brightness of the blurred area increases, and the corresponding channel difference will also increase. At the same time, the maximum channel of the inverse image obtained by exchanging the light and dark areas will become smaller, so the channel difference contrast will increase, and vice versa.

4. The infrared thermal imaging temperature measurement method for dynamic water fog interference scenes according to claim 1 is characterized by: The specific process of step (3) is as follows: a) Stacked asymmetric autoencoders; The mapping function used by the encoder can be expressed as follows; In the formula, h is the hidden layer after encoding; s1 and s2 are sigmoid activation functions; W1, W2, b1 and b2 are weight coefficients and intercept terms respectively; is the input variable; The decoder reconstructs the hidden layer h into the original input through a mapping function. The mapping function used by the decoder can be expressed as follows: In the formula, is the reconstructed original input after decoding; s3 is the sigmoid activation function; W3 and b3 are the weight coefficient and intercept term respectively; The asymmetric autoencoder optimizes the model parameters through the loss function L in the following formula; Where m is the number of input variables. The asymmetric autoencoder can be regarded as a three-layer neural network consisting of an encoder and a decoder. At the same time, the size of the output layer is the same as the size of the input layer. The asymmetric autoencoder performs unsupervised training on the encoder by making the reconstructed output as equal to the input as possible, thereby improving the feature extraction ability of the model. A feature extraction layer is added during stacking connection so that the next asymmetric autoencoder can better learn the output of the previous asymmetric autoencoder and improve the prediction ability of the model. The feature extraction layer can be expressed as follows; k=sigmoid(W4f in )(15) f out =ReLU(k⊙(W5f in ))(16) Where k is the weight vector; ⊙ is the element multiplication operation; W4 and W5 are weight coefficients; f in and f out are the input feature vector and the output feature vector respectively; For the hidden layer features output by the asymmetric autoencoder, two learnable weight vectors of equal size are used to weight the input features respectively. One branch uses the sigmoid function as the activation function, performs element-wise multiplication with the other branch, and then uses the ReLU activation function to obtain the final output features. b) Two-stage infrared band water mist transmittance estimation; According to the dynamic water fog interference quantification method proposed in (2), due to the influence of water fog suspended particles in the air, when the water fog concentration is higher, the brightness in the corresponding area is higher and the saturation is lower. Therefore, the blurred area in the foggy image has the characteristics of high brightness and low saturation. In general, the water fog concentration will increase with the change of scene depth. The above phenomenon can be expressed by the following formula: d(x)∝c(x)∝function(f r (x),D r (x),T r (x),C r (x),S r (x),L r (x),Con(x))(17)where d(x) is the scene depth; c(x) is the water mist concentration; The relationship between transmittance and scene depth is as follows: t(x)=e -βd(x) (18) Where t(x) is the transmittance in the visible light band; d(x) is the scene depth; β is the atmospheric scattering coefficient; Therefore, by quantifying the multi-scale channel differences, dark channels, gradients and other features, a rough estimate of the water mist transmittance in the visible light band can be made; In addition, according to Lambert-Beer's law, the mass extinction coefficient of water mist is related to the wavelength; t wm (λ)=exp[-β′d′C′k′(λ)](19) In the formula, t wm (λ) is the water mist transmittance in the visible light band; β′ is the proportionality coefficient; d′ is the influence distance of the water mist; C′ is the relative water mist concentration; k′(λ) is the mass extinction coefficient of the water mist; It can be seen from formula (17) that the saturation S and brightness V of the visible light image are related to most of the quantitative features of water mist concentration, and the channel difference contrast feature can quantify the water mist concentration. The water mist transmittance in the visible light band is corrected to the infrared band through the features such as saturation, brightness and channel difference contrast. t inf (λ1)=function(t vis (λ2),S,V,Con)(20) Where, t inf (λ1) is the water mist transmittance in the infrared band; t vis (λ2) is the water mist transmittance in the visible light band; λ1 and λ2 are the imaging bands of the infrared thermal imager and the visible light camera, respectively; S is the saturation; V is the brightness; Con is the channel difference contrast; Saturation S and brightness V are defined as follows: V=max(R,G,B)(22) Where R, G and B correspond to the red, green and blue color channels in the image respectively.

5. The infrared thermal imaging temperature measurement method for dynamic water fog interference scenes according to claim 1 is characterized by: The specific process of step (4) is as follows: Step 1: Collect infrared thermal images and visible light images of the object under test at different temperatures, different scenes, and different water mist concentrations. At the same time, collect the real temperature of the object under test and the measured temperature affected by water mist, and calculate the true value of the water mist transmittance label in the infrared band according to the following formula; Where, t label is the true value of the water mist transmittance label; T effect , T truth and T u are the measured temperature affected by water mist, the real temperature of the measured object and the ambient temperature respectively; n is the fitting coefficient related to the detector material of the infrared thermal imager; Randomly extract two-thirds of the collected paired infrared thermal images and visible light image datasets as training sets, and the rest of the datasets as test sets; Step 2: Establish a prediction model based on stacked asymmetric autoencoders to extract the quantitative features of water mist concentration in the visible light image of the object under test, convert the visible light image into the HSV color space, extract the saturation and brightness channels, and use the proposed two-stage infrared band water mist transmittance estimation method to obtain the infrared band water mist transmittance; Step 3: Input the water mist transmittance, saturation, brightness, channel difference contrast and measured temperature in the infrared band into the prediction model for training, and test the prediction effect of the model on the temperature measurement error on the test set; Step 4: Use the trained intelligent compensation model to compensate the measured temperature of the infrared thermal imager under the interference of water fog; It is expressed by the following formula: ΔT=function(t inf ,S,V,Con,T origin )(25) In the formula, T and T origin are the output temperature and the original temperature measurement result respectively; ΔT is the prediction error; The loss function constraint based on physical information ensures that the intelligent model can accurately estimate the water mist transmittance in the infrared band while also estimating the measurement error. The loss function of the intelligent compensation model composed of physical loss and label loss can be expressed as follows: L total =L label +L physi (28) Where, L total , L label and L physi They are the total loss function, label loss and physical loss respectively; m is the number of input variables; is the predicted water mist transmittance in the infrared band; T i ′、T i and T u are the affected measured temperature, true temperature and ambient temperature respectively; and are the predicted error and the true error respectively; The collected infrared thermal images and visible light images are divided into data sets, and the established intelligent compensation model is trained using the training set. The temperature measurement error under water mist interference is predicted, thereby achieving accurate infrared temperature measurement under dynamic water mist interference.

Images