Two-color and three-color combined colorimetric temperature measurement method based on ensemble learning

Abstract

The invention discloses a two-color and three-color combined colorimetric temperature measurement method based on ensemble learning, which considers the influence of a CCD (Charge Coupled Device) nonlinear response area, provides channel reliability calibration, considers the dominance of noise level occupying influence on temperature measurement precision, provides noise level calibration, and acquires a black-body furnace image with uniform temperature for a period of time. And fitting a linear function and a quadratic function by using the relationship between the spatial domain image and the time domain image, and obtaining noise magnitude statistical parameters introduced under the current measurement condition as priori knowledge of integrated learning. And finally, combining channel reliability, noise parameters, variation coefficients of an image space domain and a colorimetric gamut and the like to form a feature domain, and learning and predicting an optimal temperature measurement formula by using a random forest. When the CCD image is used for temperature measurement, firstly, the trained random forest is used for judging the optimal temperature measurement formula suitable for the current image, and then temperature measurement is carried out based on the colorimetric temperature measurement principle. According to the method, the advantages of a two-color temperature measurement method and a three-color temperature measurement method can be combined, the temperature measurement precision is improved while the temperature measurement dynamic range is ensured, and the overall performance of a temperature measurement system is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of radiation temperature measurement, and more specifically, relates to a two-color and three-color combined colorimetric temperature measurement method based on ensemble learning. Background Art

[0002] The measurement of high-temperature temperature field is to measure the temperature distribution of the entire field of the high-temperature target body. It has a wide range of applications and important significance in modern industrial production environments. In the metallurgical industry, accurate transient high-temperature furnace temperature field measurement is crucial to ensure the composition stability of steel and avoid material defects. The turbine blades of aircraft engines can withstand a maximum temperature of 2000°C during operation. Monitoring the temperature amplitude and heat flux distribution on their surface is the guarantee for the safe operation of the engine. The measurement of high-temperature temperature field is of great significance in ensuring the safe operation of industrial equipment, improving product quality and optimizing production processes.

[0003] High-temperature temperature field measurement technology can be divided into two categories: contact measurement and non-contact measurement. Contact measurement is susceptible to interference in high-temperature environments, and it is a single-point measurement that cannot cover the entire high-temperature full-field target. In recent years, in order to solve the problems of full-field temperature measurement and transient measurement, non-contact measurement has received widespread attention. Common non-contact measurement technologies include infrared temperature measurement, fiber optic sensor temperature measurement, and CCD-based colorimetric temperature measurement. Among them, the accuracy of infrared temperature measurement is affected by the surface emissivity and measurement environment, the fiber optic temperature measurement system is relatively complex and costly, and the CCD charge-coupled device has the advantages of high sensitivity, low signal distortion, stable and reliable operation, etc., so CCD-based colorimetric temperature measurement has become the best choice for high-temperature measurement.

[0004] However, in the process of temperature measurement using CCD charge coupled device, since it can collect the radiation of three channels of red, green and blue, the temperature can be calculated using the information of two channels, that is, two-color colorimetric temperature measurement, or the temperature can be calculated using the information of three channels, that is, three-color colorimetric temperature measurement. In the existing colorimetric temperature measurement, they are considered in an either-or relationship, that is, in the whole process of temperature measurement, either two channels are used throughout the whole process, or three channels are used throughout the whole process. In fact, two-color colorimetric temperature measurement and three-color colorimetric temperature measurement each have their corresponding advantages, for example, two-color colorimetric temperature measurement has a higher dynamic range, and three-color colorimetric temperature measurement has a higher theoretical accuracy. If the two methods can be organically combined and the most suitable temperature measurement method can be selected under appropriate conditions, it will be of great significance to improve the measurement accuracy, robustness, automation and temperature measurement quality of colorimetric temperature measurement based on CCD. However, under what conditions, which colorimetric method has better temperature measurement performance is a complex problem under the coupling of multiple factors, and it is difficult to manually determine a general criterion. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a two-color and three-color combined colorimetric temperature measurement method based on ensemble learning to achieve two-color and three-color combined colorimetric temperature measurement and improve temperature measurement performance.

[0006] To achieve the above-mentioned object of the invention, the present invention provides a two-color and three-color combined colorimetric temperature measurement method based on ensemble learning, which is characterized by comprising the following steps:

[0007] (1) Construction and training of a classifier model for selecting two-color or three-color temperature measurement

[0008] 1.1) Channel reliability calibration: Obtain the channel reliability range (V min ,V max );

[0009] 1.2) Noise level calibration: calculate M noise statistical parameters under the current temperature measurement conditions;

[0010] 1.3) Training set construction and generation

[0011] 1.3.1) Fix the temperature of the blackbody furnace at T℃ for a period of time until the temperature at the center of the blackbody furnace is uniform. Set up the CCD camera and set the exposure time in increments. Take a total of N images of the center of the blackbody furnace {I 1 ,I 2 ,…,I N};

[0012] 1.3.2) Take the center image of the blackbody furnace {I 1 ,I 2 ,…,I N} The image block of the central local area A×B size is calculated, and the mean value of the image block of each channel is used as the channel pixel value, that is, for the nth (n=1,…,N) image I n , whose red channel pixel value is:

[0013]

[0014] in, Indicates I n The pixel value of the red channel in row a and column b;

[0015] The green channel pixel values ​​are:

[0016]

[0017] in, Indicates I n The pixel value of the green channel in row a and column b;

[0018] The blue channel pixel values ​​are:

[0019]

[0020] in, Indicates I n The pixel value of the blue channel in row a and column b;

[0021] 1.3.3) Calculate N images {I 1 ,I 2 ,…,I N} of the two-color colorimetric value set V two And the tristimulus colorimetric value set V three :

[0022]

[0023] 1.3.4) Calculate the proportional coefficient corresponding to each set of two-color colorimetric values The proportional coefficient corresponding to the three-color colorimetric value

[0024]

[0025] Where C2 = 1.4388 × 10 -2 m·K is the second radiation constant, λ r ,λ g ,λ b are the wavelengths of red, green and blue light respectively;

[0026] 1.3.5) Change the temperature of the black body furnace T℃ and repeat 1.3.1) to 1.3.4) several times to obtain a predetermined number of two-color colorimetric values ​​and their proportional coefficients. Three color ratio values ​​and their proportionality coefficients

[0027] The functional relationship between the two-color colorimetric value and its proportional coefficient is obtained by fitting using the least squares method:

[0028]

[0029] Where: S two Represents the functional relationship between the two-color colorimetric values ​​and their proportionality coefficients;

[0030] The functional relationship between the three-color colorimetric values ​​and their proportional coefficients is obtained by fitting using the least squares method:

[0031]

[0032] Where: S three Represents the functional relationship between the three-color colorimetric values ​​and their proportional coefficients;

[0033] 1.3.6) For each blackbody furnace image collected, construct a feature vector:

[0034] Based on the channel reliability interval (V min ,V max ), determine whether the red, green and blue channels of each image are within the channel reliability interval (V min ,V max ), if it is 1 and not 0, the judgment results of the three channels are used as the first 3 dimensions of the feature vector;

[0035] The M noise statistical parameters under the current temperature measurement conditions are used as the subsequent M dimensions of the feature vector;

[0036] Calculate the local coefficient of variation CV in the spatial domain I :

[0037]

[0038] Where I is the local area of ​​the blackbody furnace center of the blackbody furnace image, Var(I) represents the variance of the local image, E(I) is the average value of the local image, and the local coefficient of variation CV in the spatial domain is I The three channel values ​​of are taken as the 3+M+1, 3+M+2 and 3+M+3 dimensions of the feature vector;

[0039] Calculate the local coefficient of variation CV of the two-color ratio color gamut two :

[0040]

[0041] Among them, I two is a two-color colorimetric value local image obtained by calculating the two-color colorimetric value for each pixel in the local image I, E(I two ) is the two-color colorimetric value local image I two The spatial domain mean of Var(I two ) is the two-color colorimetric value local image I two The spatial domain variance of

[0042] CV two As the 3rd+M+4th dimension of the feature vector;

[0043] Calculate the local coefficient of variation CV of the three-color ratio color gamut three :

[0044]

[0045] Among them, I three The local image of three-color colorimetric values ​​obtained by calculating the three-color colorimetric values ​​of each pixel in the local image I, E(I three) is the local image I of the three-color colorimetric value three The spatial domain mean of Var(I three ) is the local image I of the three-color colorimetric value three The spatial domain variance of

[0046] CV three As the 3rd+M+5th dimension of the feature vector;

[0047] 1.3.7) For each blackbody furnace image I collected n , calculate its two-color colorimetric value and three-color colorimetric value respectively, and substitute the two-color colorimetric value and three-color colorimetric value into the fitted function relationship respectively:

[0048]

[0049] Get the corresponding proportional coefficient, recorded as the inversion proportional coefficient and Then the inversion scale factor and Substitute the following formula to obtain the inverted temperature value of the image:

[0050]

[0051] Compare Blackbody Furnace Images I n Two inversion temperatures T two ′ and T three ′ and collect the black body furnace image I n If the temperature inverted by the three-color method is closer to the real temperature, the temperature measurement label of the image is 1, otherwise the temperature measurement label is 0.

[0052] The feature vector and the temperature measurement label form a training sample. Each blackbody furnace image I n Can be used as a training sample;

[0053] Repeat steps 1.1) to 1.3) under different temperature measurement conditions to obtain the input training set of the random forest learner;

[0054] 1.4) Using the random forest-based ensemble learning method to train a classifier model to determine whether each blackbody furnace image should use two-color colorimetric temperature measurement or three-color colorimetric temperature measurement to achieve better temperature measurement accuracy;

[0055] (2) Measure the radiation temperature of the newly collected blackbody furnace image based on the trained classifier model

[0056] For each newly collected blackbody furnace image, a corresponding feature vector is obtained according to the method of 1.3.6), and then substituted into the classifier model to obtain the final prediction result. If it is 1, the blackbody furnace image is measured using the three-color temperature measurement method, otherwise the two-color temperature measurement method is used for temperature measurement.

[0057] The object of the invention of the present invention is achieved in this way:

[0058] The dual-color and three-color combined colorimetric temperature measurement method based on ensemble learning of the present invention extracts the influencing factors affecting the dual-color temperature measurement and three-color temperature measurement performance as feature vectors based on the colorimetric temperature measurement principle, and trains the random forest model as a classifier model in combination with the ensemble learning method to determine the optimal temperature measurement formula for the current blackbody furnace image and improve the overall temperature measurement performance. In the first stage of CCD imaging, considering the influence of the nonlinear response area of ​​the CCD, a channel reliability calibration is proposed to calibrate the reliable interval of the channel amplitude, and the channel reliability is used as one of the influencing features. In the second stage of CCD imaging, the noise level dominates the temperature measurement accuracy, and a noise level calibration is proposed. By collecting a blackbody furnace image with uniform temperature for a period of time, the relationship between the spatial domain image and the time domain image is used to fit the linear function and the quadratic function, and the statistical parameters of the noise size introduced under the current measurement conditions are obtained as the prior knowledge of ensemble learning. Finally, the feature domain is formed by combining the channel reliability, noise parameters, and the coefficient of variation of the image space domain and the colorimetric domain, and the random forest is used to learn and predict the optimal temperature measurement formula. When using CCD images for temperature measurement, the trained random forest is first used to determine the optimal temperature measurement formula applicable to the current image, and then the temperature is measured based on the colorimetric temperature measurement principle. This method can combine the advantages of two-color temperature measurement and three-color temperature measurement, ensure the dynamic range of temperature measurement, improve the temperature measurement accuracy, and improve the overall performance of the temperature measurement system.

[0059] At the same time, the dual-color and three-color combined colorimetric temperature measurement method based on ensemble learning of the present invention also has the following beneficial effects:

[0060] 1. The present invention combines the advantages of the two-color temperature measurement method and the three-color temperature measurement method, and integrates the integrated learning method into the traditional CCD radiation temperature measurement. The machine learning method is used to quickly determine the most suitable temperature measurement formula for the current image. The temperature measurement system can retain the advantages of the high dynamic range of the two-color temperature measurement method and combine the advantages of the small temperature measurement error of the three-color temperature measurement. Compared with the traditional single temperature measurement method system, it has better temperature measurement performance and robustness.

[0061] 2. In the process of constructing the characteristic factors that affect the optimal temperature measurement formula, the present invention takes into account that in the stage from receiving spectral radiation to generating photoelectrons, due to the existence of the nonlinear response interval of the CCD, the proportionality between photoelectrons and spectral radiation may be destroyed, so a channel reliability calibration method based on colorimetric values ​​is proposed. The colorimetric value at the same temperature should remain unchanged, and the upper and lower bounds of the reliability interval are searched by changing the shooting conditions, so as to determine whether the amplitude of a certain channel is reliable. The effectiveness of the feature vector and the classification accuracy of the learner are improved.

[0062] 3. In the process of constructing the characteristic factors that affect the optimal temperature measurement formula, the present invention takes into account the process from photoelectrons to the generation of pixel amplitude, and the influence of CCD noise on pixel amplitude begins to dominate. A noise level calibration method based on spatiotemporal characteristics is proposed. By collecting a period of images of uniform temperature in the center of a blackbody furnace, its time domain and space domain characteristic images are calculated. The least squares fitting linear function and quadratic function are used to derive the statistical parameters of the noise size under the current conditions. It provides a reference for whether to use the information of a certain channel for temperature measurement. Improve the temperature measurement characterization capability of the feature vector. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 It is a flow chart of a specific implementation method of the two-color and three-color combined colorimetric temperature measurement method based on ensemble learning of the present invention;

[0064] Figure 2 It is a specific flow chart of channel reliability calibration in the present invention;

[0065] Figure 3 It is an example result diagram of the dual-color critical value search process in the channel reliability calibration process of the present invention;

[0066] Figure 4 It is an example result diagram of the three-color critical value search process in the channel reliability calibration process of the present invention;

[0067] Figure 5 It is a specific flow chart of the noise statistical parameter calibration method in the present invention;

[0068] Figure 6 It is a fitting result diagram of a quadratic function in the noise statistical parameter calibration process of the present invention;

[0069] Figure 7 It is a fitting result diagram of a linear function in the noise statistical parameter calibration process of the present invention;

[0070] Figure 8 It is a specific flow chart of constructing ensemble learning training set samples in the present invention;

[0071] Fig. 9It is a schematic diagram of the coefficient of variation in the spatial domain of the sample set collected by the present invention;

[0072] Fig.10 It is a schematic diagram of the colorimetric gamut variation coefficient of the sample set collected by the present invention;

[0073] Fig.11 It is a comparison chart of the temperature measurement results of the two-color and three-color combined colorimetric temperature measurement method based on ensemble learning of the present invention and the temperature measurement results of the traditional pure two-color temperature measurement method and pure three-color temperature measurement method. DETAILED DESCRIPTION

[0074] The specific implementation of the present invention is described below in conjunction with the accompanying drawings so that those skilled in the art can better understand the present invention. It should be noted that in the following description, when the detailed description of known functions and designs may dilute the main content of the present invention, these descriptions will be omitted here.

[0075] Figure 1 It is a flow chart of a specific implementation of the two-color and three-color combined colorimetric temperature measurement method based on integrated learning of the present invention.

[0076] In this embodiment, if Figure 1 As shown, the dual-color and three-color combined colorimetric temperature measurement method based on integrated learning of the present invention comprises the following steps:

[0077] Step S1: Construction and training of a classifier model for selecting two-color or three-color temperature measurement Step S1.1: Channel reliability calibration

[0078] In this embodiment, if Figure 2 As shown, channel reliability calibration includes the following steps:

[0079] Step S1.1.1: Initialize the two-color critical value set Three color threshold sets Set the change significance threshold μ Two ,μ Three ;

[0080] Step S1.1.2: Fix the temperature of the blackbody furnace at T℃ for a period of time until the temperature at the center of the blackbody furnace is uniform, set up the CCD camera, and set the exposure time in increments. Take a total of N images of the center of the blackbody furnace {I 1 ,I 2 ,…,I N};

[0081] Step S1.1.3: Take the center image of the blackbody furnace {I 1 ,I 2 ,…,I N} The image block of the central local area A×B size is calculated, and the mean value of the image block of each channel is used as the channel pixel value, that is, for the nth (n=1,…,N) image I n , whose red channel pixel value is:

[0082]

[0083] in, Indicates I n The pixel value of the red channel in row a and column b;

[0084] The green channel pixel values ​​are:

[0085]

[0086] in, Indicates I n The pixel value of the green channel in row a and column b;

[0087] The blue channel pixel values ​​are:

[0088]

[0089] in, Indicates I n The pixel value of the blue channel in row a and column b;

[0090] Step S1.1.4: Calculate N images {I 1 ,I 2 ,…,I N} of the two-color colorimetric value set V two And the tristimulus colorimetric value set V three :

[0091]

[0092] Step S1.1.5: Find the two-color colorimetric value set V two The n1 group of colorimetric values ​​with relatively stable colorimetric values ​​are calculated as the reference value of the two-color colorimetric value at the current temperature.

[0093]

[0094] Among them, V two {n1} represents the two-color colorimetric value set V two The middle n1 group has a stable set of colorimetric values;

[0095] Find the tristimulus colorimetric value set V three The n2 groups of colorimetric values ​​with relatively stable colorimetric values ​​are calculated as the reference value of the three-color colorimetric value at the current temperature.

[0096]

[0097] Among them, V three {n2} represents the set of three-color colorimetric values ​​V three The middle n2 group has a stable set of colorimetric values;

[0098] Step S1.1.6: Calculate the two-color colorimetric value set V two Each set of colorimetric values ​​and two-color colorimetric value reference values The difference between

[0099]

[0100] In this embodiment, if Figure 3 As shown, find a set of colorimetric values ​​less than V two The blackbody furnace center image with the minimum exposure time corresponding to the n1 group of blackbody furnace center images in {n1}, and its corresponding difference value is recorded as the difference value set In the difference value set Find the difference in Greater than the change significance threshold Δμ Two The image of the center of the blackbody furnace with the maximum exposure time The corresponding green channel amplitude Added two-color threshold set

[0101] Step S1.1.7: Calculate the two-color colorimetric value set V two Each set of colorimetric values ​​and two-color colorimetric value reference values The difference between

[0102]

[0103] In this embodiment, if Figure 4 As shown, find a colorimetric value set V greater than two The blackbody furnace center image with the maximum exposure time corresponding to the n1 group of blackbody furnace center images in {n1}, and the corresponding difference value is recorded as the difference value set In the difference value set Find the difference in Greater than the change significance threshold Δμ Two Image of the center of a blackbody furnace with the smallest exposure time The corresponding red channel amplitude Added two-color threshold set

[0104] Step S1.1.8: Calculate the three-color colorimetric value set Vthree Each set of colorimetric values ​​and three-color colorimetric reference values The difference between

[0105]

[0106] Find the colorimetric value set V that is less than three The blackbody furnace center image with the minimum exposure time corresponding to the n2 groups of blackbody furnace center images in {n2}, and the corresponding difference value is recorded as the difference value set In the difference value set Find the difference in Greater than the significant change threshold μ Three The image of the center of the blackbody furnace with the maximum exposure time The corresponding blue channel amplitude Add three color threshold set

[0107] Step S1.1.9: Calculate the three-color colorimetric value set V three Each set of colorimetric values ​​and three-color colorimetric reference values The difference between

[0108]

[0109] Find a set of colorimetric values ​​greater than V three The blackbody furnace center image with the maximum exposure time corresponding to the n2 groups of blackbody furnace center images in {n2}, and the corresponding difference value is recorded as the difference value set In the difference value set Find the difference in Greater than the significant change threshold μ Three Image of the center of a blackbody furnace with the smallest exposure time The corresponding red channel amplitude Add three color threshold set

[0110] Step S1.1.10: Change the temperature T of the black body furnace and repeat steps S1.1.2 to S1.1.9 several times to obtain the critical value set of two-color and three-color

[0111] Step S1.1.11: Calculate the maximum value of the lower critical value The lower limit of the amplitude reliability is

[0112] Step S1.1.12: Calculate the minimum value of the upper critical value The upper limit of the amplitude reliability is

[0113] Step S1.1.13: Obtain the channel reliability interval (V min ,V max )

[0114] When the amplitude of a channel of the blackbody furnace center image satisfies: V min <v R ,v G ,v B <V max (where v R ,v G ,v B When the pixel value is less affected by thermal noise and pixel saturation overflow, it can be considered that the pixel value is less affected by thermal noise and pixel saturation overflow. Using this pixel value as the input of the colorimetric temperature measurement formula can obtain a relatively reliable temperature measurement result. On the contrary, if the pixel value does not meet the above range, the colorimetric temperature value obtained based on the pixel value may have a large error.

[0115] Step S1.2: Noise level calibration

[0116] In this embodiment, if Figure 5 As shown, the noise level calibration includes the following steps

[0117] Step S1.2.1: Initialize the spatial domain mean E(μ(I)) set X=Φ of the mean image μ(I), the spatial domain variance Var(μ(I)) set Y1=Φ and the variance image σ 2 The spatial domain mean E(σ 2 (I)) Set Y2 = Φ;

[0118] Step S1.2.2: Initialize the CCD image acquisition system and keep the temperature of the blackbody furnace at T for a period of time to ensure that the temperature inside the blackbody furnace is uniform. Continuously collect blackbody furnace images for a total of F frames, use the circle detection algorithm to obtain the center area of ​​the blackbody furnace, and take a local image of size A×B near the center of the circle as a uniform brightness image, recorded as L i (i=1,2,…,F), and then calculate the spatial domain mean E(μ(I)), spatial domain variance Var(μ(I)) and spatial domain mean E(σ 2 (I)):

[0119]

[0120] Step S1.2.3: The calculated spatial domain mean E(μ(I)), spatial domain variance Var(μ(I)) and spatial domain mean E(σ 2(I)) are added to the corresponding sets: X = X∪E(μ(I)), Y1 = Y1∪Var(μ(I)), Y2 = Y2∪E(σ 2 (I));

[0121] Step S1.2.4: Change the temperature T of the blackbody furnace, and repeat steps S1.2.2 to S1.2.3 until the set to be fitted X, Y1, Y2 has the set amount of data;

[0122] Step S1.2.5: Figure 6 As shown, the least squares method is used to fit the data in Y1 and X as a quadratic function, such as Figure 7 As shown, the data in Y2 and X are fitted as a linear function, and the functional relationship between them is obtained as y1=a1x 2 +b1x+c1andy2=k2x+d2.

[0123] Step S1.2.6: Add:

[0124]

[0125] As the noise statistical parameter under the current temperature measurement conditions.

[0126] Subsequently, these statistical parameters will be combined as one of the features for determining the optimal temperature measurement formula, and the accuracy of the learner in learning the optimal formula under different shooting conditions will be enhanced.

[0127] Step S1.3: Training set construction and generation

[0128] In this embodiment, if Figure 8 As shown, the following steps are included:

[0129] Step S1.3.1: Fix the temperature of the blackbody furnace at T℃ for a period of time until the temperature at the center of the blackbody furnace is uniform. Set up the CCD camera and set the exposure time in increments. Take a total of N images of the center of the blackbody furnace {I 1 ,I 2 ,…,I N};

[0130] Step S1.3.2: Take the center image of the blackbody furnace {I 1 ,I 2 ,…,I N} The image block of the central local area A×B size is calculated, and the mean value of the image block of each channel is used as the channel pixel value, that is, for the nth (n=1,…,N) image I n , whose red channel pixel value is:

[0131]

[0132] in, Indicates I n The pixel value of the red channel in row a and column b;

[0133] The green channel pixel values ​​are:

[0134]

[0135] in, Indicates I n The pixel value of the green channel in row a and column b;

[0136] The blue channel pixel values ​​are:

[0137]

[0138] in, Indicates I n The pixel value of the blue channel at row a and column b.

[0139] Step S1.3.3: Calculate N images {I 1 ,I 2 ,…,I N} of the two-color colorimetric value set V two And the tristimulus colorimetric value set V three :

[0140]

[0141] Step S1.3.4: Calculate the proportional coefficient corresponding to each set of two-color colorimetric values The proportional coefficient corresponding to the three-color colorimetric value

[0142]

[0143] Where C2 = 1.4388 × 10 -2 m·K is the second radiation constant, λ r ,λ g ,λ b are the wavelengths of red, green and blue light respectively;

[0144] Step S1.3.5: Change the black body furnace temperature T℃ and repeat steps S1.3.1 to S1.3.4 several times to obtain a predetermined number of two-color colorimetric values ​​and their proportional coefficients. Three color ratio values ​​and their proportionality coefficients

[0145]

[0146] The functional relationship between the two-color colorimetric value and its proportional coefficient is obtained by fitting using the least squares method:

[0147]

[0148] Where: S two Represents the functional relationship between the two-color colorimetric values ​​and their proportionality coefficients;

[0149] The functional relationship between the three-color colorimetric values ​​and their proportional coefficients is obtained by fitting using the least squares method:

[0150]

[0151] Where: S three Represents the functional relationship between the three-color colorimetric values ​​and their proportional coefficients;

[0152] Step S1.3.6: For each blackbody furnace image collected, construct a feature vector:

[0153] Based on the channel reliability interval (V min ,V max ), determine whether the red, green and blue channels of each image are within the channel reliability interval (V min ,V max ), if it is 1 and not 0, the judgment results of the three channels are used as the first 3 dimensions of the feature vector;

[0154] The M noise statistical parameters under the current temperature measurement conditions are used as the subsequent M dimensions of the feature vector;

[0155] Calculate the local coefficient of variation CV in the spatial domain I :

[0156]

[0157] Wherein, I is the local area of ​​the center of the blackbody furnace of the blackbody furnace image, Var(I) represents the variance of the local image, and E(I) is the average value of the local image.

[0158] In this embodiment, the spatial domain local variation coefficient CV I like Fig. 9 shown.

[0159] The local coefficient of variation CV in the spatial domain I The three channel values ​​of are taken as the 3+M+1, 3+M+2 and 3+M+3 dimensions of the feature vector;

[0160] Calculate the local coefficient of variation CV of the two-color ratio color gamut two :

[0161]

[0162] Among them, I twois a two-color colorimetric value local image obtained by calculating the two-color colorimetric value for each pixel in the local image I, E(I two ) is the two-color colorimetric value local image I two The spatial domain mean of Var(I two ) is the two-color colorimetric value local image I two The spatial domain variance of .

[0163] In this embodiment, the local coefficient of variation of the two-color ratio color gamut CV two like Fig. 9 shown.

[0164] CV two As the 3rd+M+4th dimension of the feature vector;

[0165] Calculate the local coefficient of variation CV of the three-color ratio color gamut three :

[0166]

[0167] Among them, I three The local image of three-color colorimetric values ​​obtained by calculating the three-color colorimetric values ​​of each pixel in the local image I, E(I three ) is the local image I of the three-color colorimetric value three The spatial domain mean of Var(I three ) is the local image I of the three-color colorimetric value three The spatial domain variance

[0168] In this embodiment, the local coefficient of variation of the three-color ratio color gamut CV three like Fig.10 Shown

[0169] CV three As the 3rd+M+5th dimension of the feature vector;

[0170] Step S1.3.7: For each blackbody furnace image I collected n , calculate its two-color colorimetric value and three-color colorimetric value respectively, and substitute the two-color colorimetric value and three-color colorimetric value into the fitted function relationship respectively:

[0171]

[0172] Get the corresponding proportional coefficient, recorded as the inversion proportional coefficient and Then the inversion scale factor and Substitute the following formula to obtain the inverted temperature value of the image:

[0173]

[0174] Compare the two inversion temperatures T n ′ and T two ′ of the blackbody furnace image I three with the true temperature T of the blackbody furnace when collecting the blackbody furnace image I n If the temperature inverted by the three-color method is closer to the true temperature, record the temperature measurement label of this image as 1; otherwise, the temperature measurement label is 0.

[0175] Combine the feature vector and the temperature measurement label to form a training sample. Each blackbody furnace image I n can be used as a training sample.

[0176] Under different temperature measurement conditions, repeat steps 1.1) - 1.3) to obtain the input training set of the random forest learner.

[0177] Step S1.4: Use the ensemble learning method based on random forest to train a classifier model to determine whether better temperature measurement accuracy can be achieved by using two-color or three-color colorimetric temperature measurement for each blackbody furnace image

[0178] Step S1.4.1: Set the number of decision trees as n tree , the maximum depth as depth max , the minimum number of samples required for node splitting as min sp , and the leaf node merging threshold as min leaf .

[0179] Step S1.4.2: Randomly select |D| samples of the same size as the original sample set from the original training set D with replacement based on bootstrap, and repeat n tree times to obtain training subsets Build a decision tree based on each training subset. Each decision tree randomly selects m out of M feature vectors (m < M) to form a feature subset for training, forming a complete random forest.

[0180] Step S2: Perform radiation temperature measurement on newly collected blackbody furnace images based on the trained classifier model

[0181] Step S2.1: For each newly collected blackbody furnace image, obtain a corresponding feature vector according to the method in step S1.3.6, and substitute the feature vector of the new image into n tree decision trees in the random forest to classify and predict the optimal temperature measurement formula for this image, and finally obtain n tree prediction results where is the output result of the n tree th decision tree.

[0182] Step S2.2: Take the mode of the prediction results of each decision tree as the final prediction result of the image

[0183]

[0184] Where y = 0 or 1, which is the predicted temperature measurement result, and I(·) is the indicator function, which takes the value of 1 or 0.

[0185] Step S2.3: If the predicted temperature measurement result is 1, the image is measured using the three-color temperature measurement method, otherwise the image is measured using the two-color temperature measurement method.

[0186] Examples

[0187] In this example, the reliability of the CCD channel is first calibrated. Blackbody furnace images with different exposures are taken at the same temperature, and the average value of the central area is used as the pixel amplitude. The process of calculating the two-color critical value at 1100°C is as follows: Figure 3 As shown. Each point represents a set of two-color colorimetric values, and the dotted line represents the two-color colorimetric value reference value.

[0188] The process of calculating the three-color critical value at 1100℃ is as follows Figure 4 As shown in the figure, each point represents a set of three-color colorimetric values, and the dotted line represents the three-color colorimetric value reference value.

[0189] Then the noise level of the current shooting condition of the CCD is calibrated. The blackbody furnace videos were collected at the temperatures of 800℃, 810℃, 820℃, 830℃, 840℃, 850℃ and 900℃, and each video had F=100 frames. The image block with a size of 51×51 in the center area of ​​the blackbody furnace was collected as a uniform brightness image. The quadratic function of the red channel fitting is as follows: Figure 6 As shown, the fitted linear function is Figure 7 shown.

[0190] Then, new training set images are collected, and feature vectors are constructed based on the channel reliability calibration and noise level calibration mentioned above, as well as the spatial domain and colorimetric domain variation coefficients of the captured images. The spatial domain variation coefficient of the training set images is as follows: Fig. 9 As shown, the color gamut variation coefficient is as follows Fig.10 shown.

[0191] Finally, the random forest is trained to learn and predict the optimal temperature measurement formula. The comparison between the predicted results and the pure two-color measurement and pure three-color measurement results is shown in the figure below. Fig.11 The performance results of the model training are shown in the following table, and the prediction accuracy rate reaches 93.3%.

[0192] Table 1 Random forest model classification performance

[0193] Classification accuracy ACC Precision Recall F1-score Random Forest Model 0.9162011173184358 0.9047619047619048 0.7755102040816326 0.8351648351648352 Performance on new data 0.9333333333333333 0.9166666666666666 1.0 0.9565217391304348

[0194] Table 1

[0195] Although the above describes an illustrative specific embodiment of the present invention to facilitate the understanding of the present invention by those skilled in the art, it should be clear that the present invention is not limited to the scope of the specific embodiment. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

Claims

1. A two-color and three-color combined colorimetric temperature measurement method based on ensemble learning, characterized in that it comprises the following steps: (1) Construction and training of a classifier model for selecting two-color or three-color temperature measurement 1.1) Channel reliability calibration: Obtain the channel reliability range (V min ,V max ); 1.2) Noise level calibration: calculate M noise statistical parameters under the current temperature measurement conditions; 1.3) Training set construction and generation 1.3.1) Fix the temperature of the blackbody furnace at T℃ for a period of time until the temperature in the center of the blackbody furnace is uniform, set up the CCD camera, and set the exposure time in increments. Take a total of N images of the center of the blackbody furnace {I 1 ,I 2 ,…,I N }; 1.3.2) Take the center image of the blackbody furnace {I 1 ,I 2 ,…,I N } The image block of the central local area A×B size is calculated, and the mean value of the image block of each channel is used as the channel pixel value, that is, for the nth (n=1,…,N) image I n , whose red channel pixel value is: in, Indicates I n The pixel value of the red channel in row a and column b; The green channel pixel values ​​are: in, Indicates I n The pixel value of the green channel in row a and column b; The blue channel pixel values ​​are: in, Indicates I n The pixel value of the blue channel in row a and column b; 1.3.3) Calculate N images {I 1 ,I 2 ,…,I N } of the two-color colorimetric value set V two And the tristimulus colorimetric value set V three : 1.3.4) Calculate the proportional coefficient corresponding to each set of two-color colorimetric values The proportional coefficient corresponding to the three-color colorimetric value Where C2 = 1.4388 × 10 -2 m·K is the second radiation constant, λ r ,λ g ,λ b are the wavelengths of red, green and blue light respectively; 1.3.5) Change the temperature of the black body furnace T℃ and repeat 1.3.1) to 1.3.4) several times to obtain a predetermined number of two-color colorimetric values ​​and their proportional coefficients. Three color ratio values ​​and their proportionality coefficients The functional relationship between the two-color colorimetric value and its proportional coefficient is obtained by fitting using the least squares method: Where: S two Represents the functional relationship between the two-color colorimetric values ​​and their proportionality coefficients; The functional relationship between the three-color colorimetric values ​​and their proportional coefficients is obtained by fitting using the least squares method: Where: S three Represents the functional relationship between the three-color colorimetric values ​​and their proportional coefficients; 1.3.6) For each blackbody furnace image collected, construct a feature vector: Based on the channel reliability interval (V min ,V max ), determine whether the red, green and blue channels of each image are within the channel reliability interval (V min ,V max ), if it is 1 and not 0, the judgment results of the three channels are used as the first 3 dimensions of the feature vector; The M noise statistical parameters under the current temperature measurement conditions are used as the subsequent M dimensions of the feature vector; Calculate the local coefficient of variation CV in the spatial domain I : Where I is the local area of ​​the blackbody furnace center of the blackbody furnace image, Var(I) represents the variance of the local image, E(I) is the average value of the local image, and the local coefficient of variation CV in the spatial domain is I The three channel values ​​of are taken as the 3+M+1, 3+M+2 and 3+M+3 dimensions of the feature vector; Calculate the local coefficient of variation CV of the two-color ratio color gamut two : Among them, I two is a two-color colorimetric value local image obtained by calculating the two-color colorimetric value for each pixel in the local image I, E(I two ) is the two-color colorimetric value local image I two The spatial domain mean of Var(I two ) is the two-color colorimetric value local image I two The spatial domain variance of CV two As the 3rd+M+4th dimension of the feature vector; Calculate the local coefficient of variation CV of the three-color ratio color gamut three : Among them, I three The local image of three-color colorimetric values ​​obtained by calculating the three-color colorimetric values ​​of each pixel in the local image I, E(I three ) is the local image I of the three-color colorimetric value three The spatial domain mean of Var(I three ) is the local image I of the three-color colorimetric value three The spatial domain variance of CV three As the 3rd+M+5th dimension of the feature vector; 1.3.7) For each blackbody furnace image I collected n , calculate its two-color colorimetric value and three-color colorimetric value respectively, and substitute the two-color colorimetric value and the three-color colorimetric value into the fitted function relationship respectively: Get the corresponding proportional coefficient, recorded as the inversion proportional coefficient and Then the inversion scale factor and Substitute the following formula to obtain the inverted temperature value of the image: Compare Blackbody Furnace Images I n Two inversion temperatures T two ′ and T three ′ and collect the black body furnace image I n If the temperature inverted by the three-color method is closer to the real temperature, the temperature measurement label of the image is 1, otherwise the temperature measurement label is 0. The feature vector and the temperature measurement label form a training sample. Each blackbody furnace image I n Can be used as a training sample; Repeat steps 1.1) to 1.3) under different temperature measurement conditions to obtain the input training set of the random forest learner; 1.4) Using the random forest-based ensemble learning method to train a classifier model to determine whether each blackbody furnace image should use two-color colorimetric temperature measurement or three-color colorimetric temperature measurement to achieve better temperature measurement accuracy; (2) Measure the radiation temperature of the newly collected blackbody furnace image based on the trained classifier model For each newly collected blackbody furnace image, a corresponding feature vector is obtained according to the method of 1.3.6), and then substituted into the classifier model to obtain the final prediction result. If it is 1, the blackbody furnace image is measured using the three-color temperature measurement method, otherwise the two-color temperature measurement method is used for temperature measurement.

2. The two-color and three-color combined colorimetric temperature measurement method based on ensemble learning according to claim 1 is characterized in that: The channel reliability calibration is: 1.1.1) Initialize the two-color critical value set Three color threshold sets Set the change significance threshold μ Two ,μ Three ; 1.1.2) Fix the temperature of the blackbody furnace at T℃ for a period of time until the temperature at the center of the blackbody furnace is uniform, set up the CCD camera, and set the incremental exposure time respectively. Take a total of N images of the center of the blackbody furnace {I 1 ,I 2 ,…,I N }; 1.1.3) Take the center image of the blackbody furnace {I 1 ,I 2 ,…,I N } The image block of the central local area A×B size is calculated, and the mean value of the image block of each channel is used as the channel pixel value, that is, for the nth (n=1,…,N) image I n , whose red channel pixel value is: in, Indicates I n The pixel value of the red channel in row a and column b; The green channel pixel values ​​are: in, Indicates I n The pixel value of the green channel in row a and column b; The blue channel pixel values ​​are: in, Indicates I n The pixel value of the blue channel in row a and column b; 1.1.4) Calculate N images {I 1 ,I 2 ,…,I N } of the two-color colorimetric value set V two And the tristimulus colorimetric value set V three : 1.1.5) Find the two-color colorimetric value set V two The n1 group of colorimetric values ​​with relatively stable colorimetric values ​​are calculated as the reference value of the two-color colorimetric value at the current temperature. in, V two {n1} represents the two-color colorimetric value set V two The middle n1 group has a stable set of colorimetric values; Find the tristimulus colorimetric value set V three The n2 groups of colorimetric values ​​with relatively stable colorimetric values ​​are calculated as the reference value of the three-color colorimetric value at the current temperature. Among them, V three {n2} represents the set of three-color colorimetric values ​​V three The middle n2 group has a stable set of colorimetric values; 1.1.6) Calculate the two-color colorimetric value set V two Each set of colorimetric values ​​and two-color colorimetric value reference values The difference between Find the colorimetric value set V that is less than two The blackbody furnace center image with the minimum exposure time corresponding to the n1 group of blackbody furnace center images in {n1}, and its corresponding difference value is recorded as the difference value set In the difference value set Find the difference in Greater than the change significance threshold Δμ Two The image of the center of the blackbody furnace with the maximum exposure time The corresponding green channel amplitude Added two-color threshold set 1.1.7) Calculate the two-color colorimetric value set V two Each set of colorimetric values ​​and two-color colorimetric value reference values The difference between 1.1.8) Calculate the three-color colorimetric value set V three Each set of colorimetric values ​​and three-color colorimetric reference values The difference between Find the colorimetric value set V that is less than three The blackbody furnace center image with the minimum exposure time corresponding to the n2 groups of blackbody furnace center images in {n2}, and the corresponding difference value is recorded as the difference value set In the difference value set Find the difference in Greater than the significant change threshold μ Three The image of the center of the blackbody furnace with the maximum exposure time The corresponding blue channel amplitude Add three color threshold set 1.1.9) Calculate the three-color colorimetric value set V three Each set of colorimetric values ​​and three-color colorimetric reference values The difference between Find a set of colorimetric values ​​greater than V three The blackbody furnace center image with the maximum exposure time corresponding to the n2 groups of blackbody furnace center images in {n2}, and the corresponding difference value is recorded as the difference value set In the difference value set Find the difference in Greater than the significant change threshold μ Three Image of the center of a blackbody furnace with the smallest exposure time The corresponding red channel amplitude Add three color threshold set 1.1.10) Change the temperature T of the blackbody furnace and repeat steps 1.1.2) to 1.1.9) several times to obtain the critical value set of two-color and three-color 1.1.11) Calculate the maximum value of the lower critical value The lower limit of the amplitude reliability is 1.1.12) Calculate the minimum value of the upper critical value The upper limit of the amplitude reliability is 1.1.13) The channel reliability interval is obtained as (V min ,V max ).

3. The two-color and three-color combined colorimetric temperature measurement method based on ensemble learning according to claim 1 is characterized in that: The noise level is calibrated as: 1.2.1) Initialize the spatial domain mean E(μ(I)) set X=Φ of the mean image μ(I), the spatial domain variance Var(μ(I)) set Y1=Φ and the variance image σ 2 The spatial domain mean E(σ 2 (I)) Set Y2 = Φ; 1.2.2) Initialize the CCD image acquisition system, keep the temperature of the blackbody furnace at T for a period of time to ensure that the temperature inside the blackbody furnace is uniform, continuously collect blackbody furnace images for a total of F frames, use the circle detection algorithm to obtain the center area of ​​the blackbody furnace, and take a local image of size A×B near the center of the circle as a uniform brightness image, recorded as L i (i=1,2,…,F), and then calculate the spatial domain mean E(μ(I)), spatial domain variance Var(μ(I)) and spatial domain mean E(σ 2 (I)): 1.2.3) The calculated spatial domain mean E(μ(I)), spatial domain variance Var(μ(I)) and spatial domain mean E(σ 2 (I)) are added to the corresponding sets: X = X∪E(μ(I)), Y1 = Y1UVar(μ(I)), Y2 = Y2∪E(σ 2 (I)); 1.2.4) Change the temperature T of the blackbody furnace and repeat steps 1.2.2) to 1.2.3) until the set to be fitted X, Y1, Y2 has the set amount of data; 1.2.5) Use the least squares method to fit the data in Y1 and X as a quadratic function, and fit the data in Y2 and X as a linear function, and obtain the functional relationship between them y1 = a1x 2 +b1x+c1andy2=k2x+d2; 1.2.6) will: <h2 style=";text-align:left;direction:ltr">k2, d2, a1<h2 style=";text-align:left;direction:ltr"> As the noise statistical parameter under the current temperature measurement conditions.

Images