A passive temperature measurement and distance measurement method and system based on multispectral infrared

Abstract

A kind of passive temperature measurement ranging method and system based on multispectral infrared, multispectral long-wave infrared detection system is designed, multispectral infrared temperature measurement ranging model is constructed, the difference between the minimum system observation value and model calculation value is taken as objective function, KKT and gradient descent method are used to solve model parameters preliminarily.Further, the temperature and distance correction model is established, the loss function fusing data error term and physical information term is designed, and the corrected temperature and distance are obtained.The multispectral long-wave infrared detection system is used to obtain infrared thermal image of specific wavelength, and the multispectral infrared detection model is established, the temperature and distance of target object are synchronously obtained, and the accuracy and reliability are high.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of passive temperature measurement and distance measurement, in particular to a multi-spectral infrared-based passive temperature measurement and distance measurement method and system. BACKGROUND

[0002] Infrared thermal imaging is a widely used non-contact temperature measurement technology, which has numerous applications in steel, agriculture, power electronics, medical auxiliary diagnosis, etc. The existing infrared thermal imager needs to set parameters such as target distance and emissivity in advance. When there are multiple targets with different distances and materials in the scene, the infrared thermal imager can only correct the temperature measurement results according to the fixed parameters set in advance, and the temperature measurement results obviously have errors. In the field of depth measurement, the existing depth sensing technology usually requires clear texture and is easily affected by light, making it difficult to achieve passive depth measurement in low light conditions. Using thermal signals for distance measurement has significant advantages, and it still has reliable distance measurement capability even in darkness and harsh environments. In the above-mentioned cases, the multi-spectral infrared temperature measurement and distance measurement method can achieve synchronous acquisition of the emissivity, temperature and distance information of the target object by receiving radiation information of multiple wavelengths combined with the spectral radiation and attenuation law. Due to its accuracy and robustness in temperature and depth measurement, it is gradually becoming a key technology path in the field of temperature and distance measurement.

[0003] The existing multi-spectral infrared temperature measurement and distance measurement method can be divided into analytical optimization method and data-driven method. The analytical optimization method usually includes two key components: objective function and constraint condition. The objective function defines the target to be optimized, and the constraint condition limits the feasible solution space of the problem. In this framework, the optimal solution is the solution that makes the objective function reach the maximum or minimum value under the given constraint condition, but this method is sensitive to initial value and has strong model dependence. Especially under nonlinear conditions, the solving accuracy may be limited, and even fall into local optimal solution, resulting in deviation of the inversion result. The feature extraction and parameter inversion in the data-driven method are usually based on the deep learning framework, and a complex network architecture is constructed to capture the mapping relationship between the input multi-spectral data and the target physical parameters. This method has high requirements for data quality and network structure design. In the case of insufficient labeled data or high complexity of feature space, it may face problems such as insufficient generalization ability or feature learning bias, which further affects the final parameter inversion accuracy. At the same time, the existing methods involve mechanism modeling, but most of them do not consider the source of the radiation energy received by the infrared thermal imager comprehensively. Most methods ignore the radiation energy from the atmosphere and surrounding objects, which leads to systematic deviation in the estimation of the radiation characteristics of the target object, thereby affecting the accuracy and reliability of the measurement.

[0004] The present application combines the advantages of fast decision and physical constraints of the optimization method with the powerful nonlinear fitting and mapping capabilities of the data-driven method, fully considers the source of radiant energy, and innovatively proposes a passive temperature measurement and distance measurement method and system based on multispectral infrared. The present application designs a multispectral long wavelength infrared (LWIR) detection system, constructs a multispectral infrared temperature measurement and distance measurement model, designs a minimization objective function between system observation values and model calculation values, constructs a fitting equation of emissivity versus wavelength, reduces the number of unknown parameters, constrains the solution of the objective function, simplifies the preliminary solution of the target object temperature, distance and emissivity, establishes a temperature and distance correction model based on a fully connected neural network (FCNN), designs a model loss function that combines data error terms and physical information error terms, overcomes the problem of incomplete consideration of the source of radiant energy and low solution accuracy in traditional temperature measurement and distance measurement methods, and realizes accurate detection of the temperature and distance of the target object.

[0005] The differences compared with the prior art are as follows:

[0006] Application (Patent) No.: CN202410174012.X

[0007] Application Publication No.: CN118067250A

[0008] Invention Name: Infrared Temperature Compensation Method, Device, Equipment and Medium Based on Ranging

[0009] The present application proposes an infrared temperature compensation method, device, equipment and medium based on ranging, which obtains the first distance between the infrared sensor and the target object through the ultrasonic sensor, and compensates the infrared temperature measurement result of the target object based on the first distance and the environmental temperature. However, this application needs to use an ultrasonic sensor for distance measurement, and the ultrasonic distance measurement must be performed before or after the infrared temperature measurement, so the temperature measurement and distance measurement cannot be performed simultaneously, which cannot meet the real-time requirement.

[0010] Application (Patent) No.: CN202010688429.X

[0011] Application Publication No.: CN112013964A

[0012] Invention Name: Infrared Temperature Distance Automatic Compensation Method, System, Device and Storage Medium

[0013] The application provides an infrared temperature measurement distance automatic compensation method, system, device and storage medium, target distance is acquired through a visible light image, and an infrared temperature measurement result is compensated based on the distance. However, the measurement system is complex, a double visible light camera and an infrared thermal imager need to be used at the same time, there are problems of multi-sensor calibration and time synchronization data acquisition difficulty, and the measurement effect is limited in a low-light environment. SUMMARY

[0014] In view of the above problems, the application provides a passive temperature measurement and distance measurement method and system based on multispectral infrared.

[0015] To achieve the above object, the technical scheme adopted by the application is as follows:

[0016] The application provides a passive temperature measurement and distance measurement system based on multispectral infrared, which comprises a signal line, a thermometer, a carrying handle, a protective shell, a focusing lens, a motor, an infrared thermal imager, a rotating rod, a filter wheel buckle and a filter wheel.

[0017] The application provides a passive temperature measurement and distance measurement method of the passive temperature measurement and distance measurement system based on multispectral infrared, which comprises the following steps:

[0018] (1) a multispectral LWIR detection system is designed, the detection system is composed of an infrared thermal imager, a filter wheel with four narrow-band filters, one full-transparency filter and one light-shielding filter, a thermometer and a focusing lens, the filter wheel is controlled to rotate by a motor, so as to switch the narrow-band filters and the light-shielding filter, the atmospheric temperature is measured by the thermometer, and the infrared radiation is converged to the surface of the system by the focusing lens;

[0019] (2) based on the infrared radiation temperature measurement principle and the Lambert-Beer law, the radiation energy of the target object itself, the radiation energy of the target object reflected by the surrounding environment object, the atmospheric radiation energy and the radiation attenuation are fully considered, and a multispectral infrared temperature measurement and distance measurement model is constructed;

[0020] (3) Construct a minimization objective function between the system observation value and the model calculation value, and construct a fitting equation of the emissivity with respect to the wavelength according to the wavelength variation law of the spectral emissivity, take the fitting equation as a constraint condition of the objective function, and iteratively optimize the objective function by using KKT and gradient descent method, so that the temperature, distance and emissivity are finally solved;

[0021] (4) Construct a temperature and distance correction model based on FCNN, correct the temperature and distance by using the solved emissivity information, and in addition, design a loss function fusing a data error term and a physical information error term to provide feedback guidance for updating the FCNN parameters.

[0022] As a further improvement of the method of the application, the step (1) of constructing a multi-spectral LWIR detection system is specifically as follows:

[0023] The system is composed of an infrared thermal imager, a filter wheel with four narrow-band filters, a full-transmission filter and a light-shielding filter, a thermometer and a focusing lens. According to the specific spectral response curve of the infrared thermal imager in the long-wave infrared band, four long-wave infrared narrow-band filters at the response peak are selected. The spectral band allowed to pass through the narrow-band filter is very narrow, that is, for the spectral channel i = 1, 2, …, K, it can be approximately regarded as all incident radiation being absorbed only at the discrete wavelength λ i The detection system rotates the filter wheel by motor control to switch the narrow-band filter and the light-shielding filter; the atmospheric temperature is measured by the thermometer; and the infrared radiation is converged to the system surface by the focusing lens;

[0024] The light-shielding filter is used to reduce two special noises existing in the LWIR imaging process: "narcissus effect" and 1 / f noise. In the system design, the light-shielding filter is introduced. By quickly switching the narrow-band filter and the light-shielding filter, the camera captures a pair of images containing / without scene radiation, i.e. scene and reference. By subtracting the two images, as shown in formula (1), the image containing only scene radiation can be obtained, and the "narcissus effect" is effectively eliminated:

[0025] I(λ i )=I s (λi)-I r (λ i ) (1)

[0026] Wherein, I s (λ i ) and I r (λ i ) are the scene image and the reference image affected by the "narcissus effect", respectively, and I(λ i ) is the image containing only scene radiation;

[0027] By quickly switching narrow-band filters and light-shielding pieces, a plurality of pairs of images are acquired in a short time, and the 1 / f noise can be significantly suppressed by subtracting and averaging the images:

[0028]

[0029] wherein N is the number of pairs of images taken, is the image obtained by averaging N I(λ i ).

[0030] As a further improvement of the method of the present application, step (2) establishes a multi-spectral infrared temperature measurement and ranging model, which is specifically as follows:

[0031] Step 1: Starting from the object emission term, the ideal black body emits light of wavelength λ at temperature T, according to Planck's black body radiation law:

[0032]

[0033] where λ is the wavelength of light, c is the speed of light, h is the Planck constant, and k is the Boltzmann constant;

[0034] The emissivity ε(λ) of the object is defined as the ratio between the radiation emitted by the object and the radiation emitted by the black body source at that temperature, so the radiation energy emitted by the object with emissivity ε(λ) and temperature T is represented as:

[0035] E(λ;T)=ε(λ)M e (λ;T) (4)

[0036] When the light propagates through the atmosphere to the sensor, some of it is absorbed. The ratio between the radiation reaching the sensor and the radiation emitted by the object is defined as the transmittance T air (λ), then the contribution of the object to the observed radiation is given by:

[0037] E obj (λ;T)=τ air (λ)ε(λ)M e (λ;T) (5)

[0038] According to Kirchhoff's law of thermal radiation, at thermal equilibrium, the absorptivity is equal to the emissivity, and the emissivity of the air is related to the transmittance, i.e.:

[0039] ε air (λ)=1-τ air (λ) (6)

[0040] According to (6), the emission of air at temperature T air is:

[0041] E air (λ;Tair ) = (1 - τ air (λ))M e (λ; T air ) (7)

[0042] where T air is the air temperature, obtained from a thermometer;

[0043] The target will reflect the radiant energy of the surrounding objects, quantified as follows:

[0044] E ref (λ) = τ air (λ)(1 - ε(λ))X λ (8)

[0045] where X λ represents the radiant energy of the surrounding objects at wavelength λ;

[0046] Considering the contributions of the target object, the atmosphere, and the surrounding environment, the spectrum observed by the measurement system is represented as:

[0047] I(λ) = E obj (λ; T) + E air (λ; T air ) + E ref (λ) (9)

[0048] Using the Lambert-Beer law, the transmittance of the medium can be represented as:

[0049] i out (λ) = exp(-σ air (λ)d) in (λ) (10)

[0050] where i in (λ) and i out (λ) are the input and transmitted intensities of the spectrum, respectively, σ air (λ) is the extinction coefficient of the medium, and d is the thickness of the medium, τ air (λ) = exp(-σ air (λ)d);

[0051] Based on equations (3)-(10), the radiant intensity I(λ) at wavelength λ can be represented as:

[0052]

[0053] where R v (λ) is the sensitivity that relates the energy density to the observed intensity;

[0054] To simplify the calculation, it is assumed that the atmospheric pressure and water vapor content are known, and the extinction coefficient is fixed for all λ, then the parameter σ air (λ) can be obtained from the air temperature, humidity and pressure readings of the weather station using high-resolution spectral modeling software SpectralCalc; a blackbody furnace with known temperature, ε(λ) = 1, is placed in front of the thermal imager at 0 m, formula (11) can be simplified as:

[0055] I(λ) = R v (λ)M e (λ; T) (12)

[0056] where I(λ) is the observed radiation intensity, M e (λ; T) can be calculated according to formula (3), and all are known values, so R v (λ) can be calculated according to formula (12)

[0057] Step 2: The radiation energy X λ of the surrounding environment objects at wavelength λ is represented as follows:

[0058]

[0059] where α represents the target object, β represents all other objects in the surrounding environment, V αβ is the view factor, representing the proportion of radiation from β reaching α, E βλ represents the infrared radiation emitted by object β, V αβ satisfies:

[0060]

[0061] Considering the radiation emitted by a finite number of dominant objects, the E λ part of X βλ is observed as I(λ), and I(λ) is sampled into m spectra to approximately describe the m most important environmental objects. Assuming that the size of each thermal image collected is H x W x C, H is the image height, W is the image width, and C is the number of spectral channels, considering m environmental objects, the image H x W is divided into m x 1 quadrants in space, and the dimension of each quadrant is

[0062] Then, each quadrant is averaged in space to a spectrum with length C, that is, for each of the C channels, the sub-image with size is averaged to obtain its average value. Considering m = 2 environmental objects, the two spectra obtained according to the above method are represented as E 1λ and E 2λ , which are the equivalent objects of the environment, and now X λ is given by:

[0063] X λ =V α1 E 1λ +V α2 E 2λ (15)

[0064] By substituting equation (15) into (11), the unknown parameter set to be estimated becomes {d, T, ε(λ i ), V α1 , V α2}, where i = 1, 2, …, K.

[0065] As a further improvement of the method of the present application, the construction and solving of the objective function of step (3) are as follows:

[0066] In equation (11), there are K+4 unknowns for the measurement of K spectral channels, i.e., K unknown values of the emissivity of the object, 1 unknown value of the temperature, 1 unknown value of the distance, and 2 unknown values of the view factor, which is an underdetermined problem. The objective function is used to solve the problem, a minimization objective function between the system observation value and the model calculation value is constructed, a fitting equation of the emissivity with respect to the wavelength is constructed, the fitting equation is taken as a constraint condition of the objective function, and the number of unknown parameters and the search range of the solution are reduced. Specifically, the following steps are included:

[0067] Step 1: Construct the objective function. The underdetermined problem has a small number of equations compared to the number of unknowns, resulting in insufficient constraints. The solution set of the underdetermined equation group has a high degree of freedom, and there may be an infinite number of solutions. Therefore, the objective function is introduced for optimization, and the selection process of the solution is changed from disorder to order. The objective function is as follows:

[0068]

[0069] where argmin is the input point that makes the function output value minimum, ||·||2 represents the 2-norm of the matrix, I(λ i ) is the observed radiation intensity, I(λ i ; d, ε(λ i ), T, V α1 , V α2 ) is the radiation intensity calculated according to equation (11);

[0070] Step 2: Increase the constraint condition. A commonly accepted assumption is that the spectral emissivity changes with the wavelength. The spectral emissivity is fitted using a wavelength function containing M, M < K, adjustable parameters. The assumption equation is as follows:

[0071]

[0072] lnε(λ) = a + bλ (18)

[0073] ε(λ) = a0+ ai λ (19)

[0074]

[0075] ε(λ) = exp[-(a0+ ai λ) 2 ] (21)

[0076] where K represents the number of spectral channels;

[0077] Using equation (20) as the constraint condition of emissivity, equation (16) becomes:

[0078]

[0079] Step 3: Solution of the objective function. First, to avoid the model falling into a local optimal solution, the model parameters are initialized to be as close to the true value as possible. The initialization parameters are as follows:

[0080]

[0081]

[0082] where M e -1 is the inverse function of Planck blackbody radiation law, K is the number of spectral channels, and U(0, 10) represents random extraction from a uniform distribution between 0 and 10m;

[0083] Subsequently, the Lagrange function is constructed:

[0084]

[0085] where η i , p i are the Lagrange multipliers of the inequality constraints of ε(λ i ), v, ξ, γ α1 , γ α2 are the Lagrange multipliers of other inequality constraints, and μ is the Lagrange multiplier of the equality constraint V α1 + V α2 = 1, all of which are non-negative values. For each inequality constraint term, it satisfies the complementary relaxation condition:

[0086]

[0087] The partial derivative of each variable is taken to obtain the KKT gradient condition:

[0088]

[0089]

[0090] Finally, the model parameters are solved by using the gradient descent method combined with the above conditions.

[0091] As a further improvement of the method of the application, the step (4) constructs a temperature and distance correction model, specifically as follows:

[0092] The emissivity information obtained in step (3) is used to correct the temperature and distance. First, a correction model network architecture is built. The network used is FCNN. The network input features are the emissivity, temperature and distance information obtained by solving the objective function. The network main body is composed of multiple fully connected layers to extract the deep nonlinear relationship of the input features. An activation function ReLU is introduced in each layer to enhance the model's fitting ability to complex distribution. Finally, the corrected temperature and distance values are generated through the output layer. To ensure physical consistency, a loss function that combines data error terms and physical information error terms is designed to provide feedback guidance for the update of FCNN parameters. Specifically, the following steps are included:

[0093] Step 1: Verify the rationality of using emissivity to correct temperature and distance. First, calculate the derivatives of formula (11) with respect to distance, emissivity and temperature:

[0094]

[0095] Eliminate common terms R v (λ)·exp(-σ air (λ)d), the core comparison of the three derivatives is simplified as:

[0096]

[0097] Based on formulas (22), (34)-(39), the derivatives of formula (22) with respect to each parameter are calculated and simplified. It is found that the order of magnitude of the derivative of formula (22) with respect to emissivity is much larger than that of the derivative of formula (22) with respect to distance and temperature. This indicates that the accuracy of emissivity is much higher than that of temperature and distance. Therefore, it is reasonable to use emissivity to correct temperature and distance.

[0098] Step 2: To solve the problem that traditional regression prediction only takes the accuracy of predicted values as the evaluation index, a loss function that combines data error terms and physical information error terms is designed:

[0099] Loss=ω1·MSE data +ω2·MSE physics (40)

[0100] Where, MSE data is the data error term, which measures the error between the network predicted value and the true value, and MSE physicsThe physical information error term is the constraint residual based on the physical model, and ω1 and ω2 are the weight hyperparameters of the data error term and the physical information error term, respectively.

[0101] The data error term is the parameter value d predicted by the network. pred and T pred Compared with the true value d true and T true The error between them is calculated as follows:

[0102]

[0103] Where N is the number of samples;

[0104] The physical information error term is the parameter value d predicted by the neural network. pred T pred The I obtained after substituting into the mechanism model (11) model (λ;d pred T pred The error between the observed value I(λ) and the observed value I(λ) is calculated as follows:

[0105]

[0106] Finally, the FCNN parameters are updated by minimizing the total loss using gradient descent.

[0107] Beneficial effects:

[0108] (1) The present invention designs a multispectral LWIR detection system, which captures specific wavelength spectral information through an infrared thermal imager and a narrowband filter, fully considering the specific contributions of different wavelengths, effectively overcoming the problem that existing infrared thermal imagers only consider the cumulative radiation of specific band intervals, which is conducive to improving the measurement accuracy of temperature and distance.

[0109] (2) Based on the principle of infrared radiation thermometry and Lambert-Beer's law, this invention fully considers the radiation energy of the target object itself, the radiation energy reflected by the target object from the surrounding environment, the atmospheric radiation energy and radiation attenuation, and constructs a multispectral infrared thermometry and ranging model. This effectively overcomes the limitation of traditional methods that only focus on the radiation of the target object itself, and significantly improves the accuracy and stability of multispectral infrared thermometry and ranging, providing a reliable technical solution for high-precision measurement in complex environments.

[0110] (3) This invention constructs a minimum objective function between the system observation value and the model calculation value, and constructs a fitting equation of emissivity to wavelength based on the law of spectral emissivity variation with wavelength. The fitting equation is used as a constraint condition of the objective function, and the objective function is iteratively optimized using KKT and gradient descent methods. Finally, the solution of parameters such as temperature, distance and emissivity is realized.

[0111] (4) The present application constructs a temperature and distance correction model based on FCNN, and corrects the temperature and distance by using the obtained emissivity information. In addition, a loss function fusing data error items and physical information error items is designed, and two target information are optimized at the same time, thereby providing feedback guidance for the update of FCNN parameters. BRIEF DESCRIPTION OF DRAWINGS

[0112] Figure 1 A multispectral LWIR detection system is shown;

[0113] Figure 2 A multispectral LWIR detection system is shown;

[0114] Figure 3 A multispectral LWIR detection system is shown;

[0115] Figure 4 A method flowchart is shown;

[0116] Figure 5 A distance regression prediction result graph is shown;

[0117] Figure 6 A temperature regression prediction result graph is shown.

[0118] Reference signs are as follows:

[0119] 1, signal line; 2, thermometer; 3, carrying handle; 4, protective shell; 5, focusing lens; 6, motor; 7, infrared thermal imager; 8, rotating rod; 9, filter wheel buckle; 10, filter wheel; 11, narrowband filter; 12, light shield; 13, full transparent sheet. DETAILED DESCRIPTION

[0120] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. The following embodiments are used to illustrate the present application, but are not used to limit the scope of the present application.

[0121] The present application proposes a multispectral infrared-based passive temperature measurement and distance measurement method and system. Figures 1-3 is a system schematic diagram, Figure 4 is an implementation step diagram, including the following steps:

[0122] (3) A multispectral LWIR detection system is designed, which is composed of an infrared thermal imager, a filter wheel with four narrowband filters, a full transparent sheet and a light shield, a thermometer and a focusing lens. The filter wheel is controlled to rotate by a motor to switch the narrowband filter and the light shield. The atmospheric temperature is measured by the thermometer, and the infrared radiation is converged to the surface of the system by the focusing lens;

[0123] (4) Based on the infrared radiation temperature measurement principle and Lambert-Beer law, fully considering the target object's own radiation energy, the target object's reflection of the surrounding environment object's radiation energy, atmospheric radiation energy and radiation attenuation, a multispectral infrared temperature measurement and ranging model is constructed;

[0124] (5) A minimum objective function between the system observation value and the model calculation value is constructed, and according to the wavelength variation law of the spectral emissivity, a fitting equation of the emissivity with respect to the wavelength is constructed, the fitting equation is taken as a constraint condition of the objective function, the KKT and gradient descent method are used to iteratively optimize the objective function, and finally the temperature, distance and emissivity and other parameters are solved;

[0125] (6) A temperature and distance correction model based on FCNN is constructed, and the emissivity information obtained by solving is used to correct the temperature and distance. In addition, a loss function fusing data error items and physical information error items is designed to provide feedback guidance for the update of FCNN parameters.

[0126] The specific implementation scheme is as follows:

[0127] (1) Multispectral LWIR detection system construction;

[0128] In view of the problem that the existing infrared thermal imager only considers the cumulative radiation in a specific waveband interval, the present application acquires multiple specific wavelength radiation information, and the designed multispectral LWIR detection system is as shown in Figures 1-3 The system is composed of an infrared thermal imager 7, a filter wheel 10 with four narrow-band filters 11, a full-transmission filter 13 and a light-shielding filter 12, a thermometer 2 and a focusing lens 5. According to the specific spectral response curve of the infrared thermal imager in the long-wave infrared band, four long-wave infrared narrow-band filters at the response peak are selected, and the narrow-band filter allows a very narrow spectral band to pass, that is, for the spectral channel i=1, 2, …, K, it can be approximately regarded as all incident radiation is only absorbed at the discrete wavelength λ i The detection system controls the rotation of the filter wheel through the motor, switches the narrow-band filter and the light-shielding filter, measures the atmospheric temperature through the thermometer, and converges the infrared radiation to the surface of the system through the focusing lens.

[0129] The application discloses a passive temperature measurement and distance measurement system based on multispectral infrared, a motor 6, an infrared thermal imager 7 and a filter wheel 10 are arranged in a protective shell 4, the motor 6 and the infrared thermal imager 7 are connected with an upper computer through a signal line 1, a rotating rod 8 is arranged at the output end of the motor 6, one end of the rotating rod 8 is connected with the filter wheel 10, the filter wheel 10 is fixed in the protective shell 4 through a filter wheel buckle 9, four narrow-band filters 11, one full-transparency sheet 13 and one light-shielding sheet 12 are arranged at equal angles on the filter wheel 10, the motor 6 outputs a control signal to rotate the rotating rod 8 so as to drive the filter wheel 10 to rotate, and the narrow-band filters 11 and the light-shielding sheet 12 are switched, a lifting handle is arranged on the upper end face of the protective shell 4, a thermometer 2 and a focusing lens 5 are arranged outside the protective shell 4, the thermometer 2 is also connected with the upper computer through the signal line 1, and the focusing lens 5 is arranged in front of the protective shell 4, so that external radiant energy is converged and enters the inside of the protective shell 4. The light-shielding sheet is used to reduce two special noises, namely, the “narcissus effect” and 1 / f noise, in the LWIR imaging process. In the LWIR observation, the radiation from the camera body is reflected by the surface of the optical element (including the narrow-band filter and the light-shielding sheet), and is imaged again on the plane where the sensor is located, thereby greatly affecting the imaging quality of the system. This is the “narcissus effect”. In order to eliminate the “narcissus effect”, the light-shielding sheet is introduced in the system design, a pair of images containing / containing no scene radiation are captured by the camera through the fast switching of the narrow-band filter and the light-shielding sheet, that is, the scene and the reference, and the two images are subtracted, so that the image containing only the scene radiation is obtained:

[0130] I(λ i )=I s (λ i )-I r (λ i ) (1)

[0131] Wherein, I s (λ i ) and I r (λ i ) are the scene image and the reference image affected by the “narcissus effect”, respectively, and I(λ i ) is the image containing only the scene radiation.

[0132] Another unique noise characteristic in the LWIR imaging process is 1 / f noise. 1 / f noise is a low-frequency fluctuation that appears strongly when processing images taken at long intervals. By quickly switching the narrow-band filter and the light-shielding sheet, a plurality of pairs of images are obtained in a short time, and after subtraction and averaging, the 1 / f noise can be significantly suppressed:

[0133]

[0134] Wherein, N is the number of images taken, is the image obtained after averaging N I(λ i )。

[0135] In summary, the present application designs a multispectral LWIR detection system, which suppresses the typical noise that may occur in the imaging process, and creates a prerequisite for subsequent model establishment.

[0136] (2) Establish a multispectral infrared temperature measurement and distance measurement model;

[0137] In order to solve the problem that the existing infrared temperature measurement and distance measurement method focuses on the radiation of the target object itself and ignores other radiation factors, the present application is based on the principle of infrared radiation temperature measurement and Lambert-Beer law, fully considers all the radiation energy sources that the infrared thermal imager may receive, and constructs a multispectral infrared temperature measurement and distance measurement model. Specifically, the following steps are included:

[0138] Step 1: Starting from the object emission term, the ideal black body emits light of wavelength λ at temperature T, according to Planck's black body radiation law:

[0139]

[0140] Where λ is the wavelength of light, c is the speed of light, h is the Planck constant, and k is the Boltzmann constant.

[0141] Real-world materials emit less thermal radiation than true black bodies. The emissivity ε(λ) of an object is defined as the ratio between the radiation emitted by the object and the radiation emitted by a black body source at that temperature. Therefore, the radiation energy emitted by an object with emissivity ε(λ) and temperature T can be represented as:

[0142] E(λ;T)=ε(λ)M e (λ;T) (4)

[0143] When light propagates through the atmosphere to the sensor, some of it is absorbed. The ratio between the radiation reaching the sensor and the radiation emitted by the object is defined as the transmittance τ air (λ). Then, the contribution of the object to the observed radiation is given by:

[0144] E obj (λ;T)=τ air (λ)ε(λ)M e (λ;T) (5)

[0145] According to Kirchhoff's law of thermal radiation, at thermal equilibrium, the absorptivity is equal to the emissivity, and the emissivity of air is related to the transmittance, that is:

[0146] ε air (λ)=1-τ air(λ) (6)

[0147] According to (6), the amount of air emitted at temperature T air is:

[0148] E air (λ; T air ) = (1 - τ air (λ)) M e (λ; T air ) (7)

[0149] where T air is the air temperature, obtained from the thermometer.

[0150] The target will reflect the radiant energy of the surrounding objects, quantified as follows:

[0151] E ref (λ) = τ air (λ)(1 - ε(λ)) X λ (8)

[0152] where X λ represents the radiant energy of the surrounding objects at wavelength λ.

[0153] Taking into account the contributions of the target object, the atmosphere and the surrounding environment, the spectrum observed by the measurement system can be represented as:

[0154] I(λ) = E obj (λ; T) + E air (λ; T air ) + E ref (λ) (9)

[0155] Using the Lambert-Beer law, the transmittance of the medium can be represented as:

[0156] i out (λ) = exp(-σ air (λ)d) i in (λ) (10)

[0157] where i in (λ) and i out (λ) are the input and transmitted intensities of the spectrum, σ air (λ) is the extinction coefficient of the medium, d is the thickness of the medium, and τ air (λ) = exp(-σ air (λ)d). Thermal cameras either eliminate this attenuation by allowing the user to manually input the depth or simply ignore it. In contrast, the present invention will take this into account in the subsequent solving process to estimate the depth of the target.

[0158] Based on equations (3)-(10), the radiance I(λ) at wavelength λ can be expressed as:

[0159]

[0160] where R v (λ) is the sensitivity that relates the energy density to the observed intensity.

[0161] To simplify the calculation, it is assumed that the atmospheric pressure and water vapor content are known, and the extinction coefficient is fixed for all λ, then the parameter σ air (λ) can be obtained from the air temperature, humidity, and pressure readings at the weather station using high-resolution spectral modeling software, SpectralCalc; a blackbody furnace with known temperature (ε(λ) = 1) is placed 0 m in front of the thermal imager, equation (11) can be simplified as:

[0162] I(λ) = R v (λ)M e (λ; T) (12)

[0163] where I(λ) is the observed radiance, M e (λ; T) can be calculated from equation (3), and both are known values, so R v (λ) can be calculated from equation (12).

[0164] Step 2: The radiated energy X λ of the surrounding environment objects at wavelength λ is expressed as:

[0165]

[0166] where α represents the target object, β represents all other objects in the surrounding environment, V αβ is the view factor, representing the proportion of radiation from β that reaches α, and E βλ represents the infrared radiation emitted by object β. V αβ satisfies:

[0167]

[0168] Although it is difficult to fully consider the radiation emitted by all objects in the surrounding environment of the target object, it is worth noting that there are usually several objects that dominate the scattering contribution, such as the sky, the ground, and buildings. Only the radiation emitted by a limited number of dominant objects is considered, X λ , where E βλPart of the observation is I (lambda), I (lambda) is sampled into m spectra to approximate the description of the m most important environmental objects. The present application assumes that the size of each thermal image collected is HxWxC, H is the image height, W is the image width, C is the number of spectral channels, considering m environmental objects, the image HxW is divided into m x 1 quadrant in space, the dimension of each quadrant is

[0169] Then, each quadrant is averaged in space to a spectrum with length C, that is, for each of the C channels, the sub-image with size is averaged to obtain its average value, considering m = 2 environmental objects, two spectra obtained according to the above method are represented as E 1λ and E 2λ , which are the equivalent objects of the environment, and X λ is given by

[0170] X λ = V α1 E 1λ + V α2 E 2λ (15)

[0171] By substituting equation (15) into (11), the unknown parameter set to be estimated becomes {d, T, epsilon (lambda i ), V α1 , V α2}, where i = 1, 2, …, K.

[0172] In summary, based on the principle of infrared radiation temperature measurement and the Lambert-Beer law, the present application constructs a multispectral infrared temperature measurement and ranging model, fully considers all the radiation that the infrared thermal imager can receive, accurately reflects the source of radiation energy, improves the reliability and interpretability of the model, and at the same time, the environmental radiation is approximately processed, and the complexity of the model is reduced.

[0173] (3) Construction and solution of the objective function;

[0174] The inversion of the forward model in equation (11) is underdetermined. For K spectral channel measurements, there are K+4 unknowns (K object emissivity unknown values, 1 temperature unknown value, 1 distance unknown value, and 2 view factor unknown values), which belongs to an underdetermined problem. In order to solve the inverse problem, the present application constructs a minimization objective function between the system observation value and the model calculation value, in view of the problem of redundant calculation or meaningless solution exploration caused by the too wide solution space of the objective function, the present application constructs a fitting equation of the emissivity with respect to the wavelength based on the wavelength variation law of the spectral emissivity, takes the fitting equation as a constraint condition of the objective function, and reduces the number of unknown parameters and the search range of the solution. Specifically, the following steps are included:

[0175] Step1: Construct the objective function. Underdetermined problem due to the number of equations less than the number of unknowns, resulting in insufficient constraints, the solution set of underdetermined equations has high degree of freedom, there may be infinite solutions. Therefore, the objective function is introduced for optimization, and the solution selection process is changed from disorder to order. The objective function is as follows:

[0176]

[0177] Where, argmin is the input point that makes the function output value minimum, ||·||2 represents the 2-norm of the matrix, I(λ i ) is the observed radiation intensity, I(λ i ; d, ε(λ i ), T, V α1 , V α2 ) is the radiation intensity calculated according to formula (11).

[0178] Step2: Increase the constraint condition. In order to further simplify the problem solving, the present application uses a wavelength function containing M (M < K) adjustable parameters to fit the spectral emissivity, which reduces the number of unknown parameters and the search range of the solution. At present, one of the widely recognized assumptions in the field of multispectral radiation thermometry is that the spectral emissivity changes with the change of wavelength, and some famous assumption equations are:

[0179]

[0180] lnε(λ)=a+bλ (18)

[0181] ε(λ)=a0+a1λ (19)

[0182]

[0183] ε(λ)=exp[-(a0+a1λ) 2 ] (21)

[0184] Where, K represents the number of spectral channels.

[0185] The present application uses formula (20) as the constraint condition of emissivity. Formula (16) becomes:

[0186]

[0187] Step3: Solution of the objective function. First, in order to avoid the model falling into local optimal solution, the model parameters are initialized to be as close to the true value as possible, and the initialization parameters are as follows:

[0188]

[0189] Where, Me -1 is the inverse function of Planck blackbody radiation law, K is the number of spectral channels, U(0, 10) represents random extraction from a uniform distribution between 0 and 10m.

[0190] Subsequently, a Lagrange function is constructed:

[0191]

[0192] Wherein, η i , ρ i is the Lagrange multiplier of inequality constraint of ε(λ i ), v, ξ, γ α1 , γ α2 is the Lagrange multiplier of other inequality constraints, μ is the Lagrange multiplier of equality constraint V α1 +V α2 =1, all are non-negative values.For each inequality constraint term, it satisfies the complementary relaxation condition:

[0193]

[0194] The partial derivative of each variable is obtained, and the KKT gradient condition is obtained:

[0195]

[0196]

[0197] Finally, combined with the above conditions, the gradient descent method is used to solve the model parameters.

[0198] In summary, the minimum objective function between the system observation value and the model calculation value is constructed, and based on the wavelength variation law of the spectral emissivity, the fitting equation of the emissivity to the wavelength is constructed, the fitting equation is taken as the constraint condition of the objective function, the number of unknown parameters and the search range of the solution are reduced, the model solving is effectively simplified, at the same time, the KKT and gradient descent method are used to solve the model parameters, which provides a basis for the realization of subsequent temperature and distance correction model.

[0199] (4) constructing a temperature and distance correction model;

[0200] To further improve the accuracy of temperature and distance, the present application uses the emissivity information obtained by solving to correct the temperature and distance. First, the correction model network architecture is built, the network used by the present application is FCNN, the network input features are the emissivity, temperature and distance information obtained by solving the objective function, the network main body is composed of multiple fully connected layers, which is used to extract the deep nonlinear relationship of the input features, and an activation function (ReLU) is introduced in each layer to enhance the fitting ability of the model to complex distribution, and finally the corrected temperature and distance values are generated through the output layer. In order to ensure physical consistency, the present application designs a loss function that combines data error term and physical information error term to provide feedback guidance for the update of FCNN parameters. Specifically, the following steps are included:

[0201] Step1: Verify the rationality of using emissivity to correct temperature and distance. First, calculate the derivative of formula (11) with respect to distance, emissivity and temperature:

[0202]

[0203]

[0204] Eliminate common terms R v (λ)·exp(-σ air (λ)d), the core comparison of the three derivatives is simplified as:

[0205]

[0206] Based on formula (22), (34)-(39), calculate the derivative of formula (22) with respect to each parameter and simplify, find that the order of magnitude of the derivative of formula (22) with respect to emissivity is much larger than the order of magnitude of the derivative of formula (22) with respect to distance and temperature, which shows that the accuracy of emissivity is much higher than the accuracy of temperature and distance, so it is reasonable to use emissivity to correct temperature and distance.

[0207] Step2: In order to solve the problem that traditional regression prediction only takes the accuracy of predicted value as the evaluation index, the present application designs a loss function that combines data error term and physical information error term:

[0208] Loss=ω1·MSE data +ω2·MSE physics (40)

[0209] Where, MSE data is the data error term, which measures the error between the network predicted value and the true value, MSE physics is the physical information error term, which is the constraint residual based on the physical model, and ω1, ω2 are the weight hyperparameters of the data error term and the physical information error term respectively.

[0210] The data error term is the error between the predicted parameter value (d pred and T pred ) of the network and the true value (d true and T true ), and is specifically calculated as follows:

[0211]

[0212] Wherein, N is the sample quantity.

[0213] The physical information error term is the error between I model (λ; d pred , T pred ) obtained by substituting the predicted parameter value (d pred , T pred ) of the neural network into the mechanism model (11) and the observed value I (λ), and is specifically calculated as follows:

[0214]

[0215] Finally, the total loss is minimized using the gradient descent method to update the FCNN parameters.

[0216] In summary, the temperature and distance correction model is constructed based on the FCNN in the application, and the emissivity information obtained is used to correct the temperature and distance. In addition, the application designs a loss function that combines the data loss term and the physical information error term, which provides feedback guidance for the update of the FCNN parameters, so that the model optimizes the two target information at the same time during the training process, and ensures the physical consistency.

[0217] Embodiment 1

[0218] In this embodiment, the blackbody furnace is taken as the object, and the application of the multispectral infrared temperature and distance measurement method and system to the temperature and distance measurement of the blackbody furnace is carried out in the laboratory environment (temperature: 24℃, relative humidity: 53%). The temperature of the blackbody furnace is set to 50℃, the distance range is 0-10m, and the step is 1m. Figure 5 、 6 The absolute error and the relative error of the distance measurement and the temperature measurement results of the application are shown respectively. Figure 3 It can be seen that the root mean square error of the distance measurement result of the application is 0.210, and the average absolute error is 0.191. Figure 4 It can be seen that the root mean square error of the temperature measurement result of the application is 0.689, and the average absolute error is 0.611. The experimental results fully prove that the method proposed in the application performs well in temperature and distance measurement, and can meet the demand of measurement accuracy.

[0219] The above merely describes the preferred embodiments of the present application, but does not constitute any other form of limitation to the present application, and any modification or equivalent change made according to the technical essence of the present application still falls within the scope of the present application.

Claims

1. A passive temperature measurement ranging method based on a multispectral infrared passive temperature measurement ranging system, characterized in that: The method comprises the following steps: (1) designing a multi-spectral LWIR detection system, the detection system being composed of an infrared thermal imager, a filter wheel with four narrow-band filters, one full-transmission filter and one light-shielding filter, a thermometer and a focusing lens, the filter wheel being controlled to rotate by a motor to switch the narrow-band filters and the light-shielding filter, the atmospheric temperature being measured by the thermometer, and the infrared radiation being converged to the surface of the system by the focusing lens; (2) based on the infrared radiation temperature measurement principle and the Lambert-Beer law, fully considering the radiation energy of the target object, the radiation energy of the target object reflecting the surrounding environment object, the atmospheric radiation energy and the radiation attenuation, a multi-spectral infrared temperature measurement and distance measurement model is constructed; The step (2) of establishing the multi-spectral infrared temperature measurement and distance measurement model is specifically as follows: Step 1: Starting from the object emission term, the temperature is The ideal blackbody emission wavelength for a temperature of is light according to Planck's law of blackbody radiation: （3） wherein is the wavelength of light, is the speed of light, is Planck's constant, is Boltzmann's constant; having an emissivity and temperature The radiant energy emitted by an object having an emissivity （4） When light propagates through the atmosphere to the sensor, some of it is absorbed, and the ratio between the radiation that reaches the sensor and the radiation emitted by the object defines the transmittance The contribution of the object to the observed radiation is then given by the following equation: （5） According to the Kirchhoff's law of thermal radiation, under thermal equilibrium, the absorption rate is equal to the emission rate, the emission rate of air is related to the transmission rate, that is: （6） According to (6), the amount of air discharged at a temperature of 20°C is: （7） wherein, for the air temperature, obtained from a thermometer; The target object will reflect the radiation energy of the surrounding environment object, which is quantified as follows: （8） wherein, represents the radiant energy of the surrounding environment object at the wavelength of 0.9 μm. Considering the contributions of the target object, the atmosphere and the surrounding environment, the spectrum observed by the measurement system is represented as: （9） Using the Lambert-Beer law, the transmission rate of the medium is represented as: （10） wherein and Iin and Itr are the input and transmitted intensities of the spectrum, respectively, k is the extinction coefficient of the medium, d is the thickness of the medium, ; Based on equations (3)-(10), the radiation intensity at the wavelength is expressed as: is expressed as: （11） wherein, is the sensitivity that relates the energy density to the observed intensity; Assuming the atmospheric pressure and water vapor content are known, the extinction coefficient is constant for all wavelengths and the parameter is obtained from the meteorological station air temperature, humidity and pressure readings using high resolution spectral modeling software, SpectralCalc; a blackbody furnace with known temperature is placed in front of the thermal camera at a distance of 0 meters and equation (11) simplifies to: （12） Thus, according to equation (12) the calculation is ; Step 2: The surrounding environment objects are in the wavelength of the radiation energy is represented as follows:​ （13） wherein, represents the target object, represents all other objects of the surrounding environment, is a view factor representing the proportion of the radiation leaving object Oj, reaching object Oi, represents the infrared radiation emitted by object Oj, satisfies:​​ （14） Consider the radiation emitted by a finite number of dominant objects. In Some were observed as ,Will Sampled In the spectrum, assuming the size of each acquired thermal image is... , Image height, Image width, For the number of spectral channels, consider An environment object, will the image Spatially divided into There are 1 quadrant, and the dimensions of each quadrant are 1. ; Each quadrant is then spatially averaged to a length of spectra, i.e. for each of the channels, a sub-image of size is averaged to obtain its average value, taking into account environmental objects, the two spectra obtained are denoted respectively and is the equivalent object of the environment, now is given by the following formula: （15） By substituting equation (15) into equation (11), the set of unknown parameters to be estimated becomes where ; (3) constructing a minimization objective function between the system observation value and the model calculation value, and constructing a fitting equation of the emission rate with respect to the wavelength according to the wavelength variation law of the spectral emission rate, taking the fitting equation as the constraint condition of the objective function, and using the KKT and gradient descent method to iteratively optimize the objective function, so as to finally solve the temperature, distance and emission rate; The step (3) of constructing and solving the objective function is specifically as follows: Step 1: constructing the objective function, the objective function being as follows: （16） wherein is the input point that makes the function output value take a minimum value, denotes the 2-norm of a matrix, is the observed radiation intensity, is the radiation intensity calculated according to equation (11); Step 2: adding the constraint condition, using the wavelength function with M, M < K adjustable parameters to fit the spectral emission rate, the assumed equation being as follows: (17) (18) （19） （20） （21） Using equation (20) as the constraint condition of the emission rate, equation (16) becomes: （22） Step 3: solving the objective function, the initialization parameters being as follows: （23） （24） （25） （26） wherein is the inverse function of Planck's blackbody radiation law, is the number of spectral channels, denotes a random draw from a uniform distribution between 0 and 10 . Subsequently, the Lagrange function is constructed as follows: （27） wherein , is a Lagrange multiplier for the inequality constraint of , , , , is a Lagrange multiplier for other inequality constraints, is a Lagrange multiplier for the equality constraint , all of which are non-negative values, and for each inequality constraint term, it satisfies a complementary slackness condition: （28） Taking the partial derivative of each variable, the KKT gradient condition is obtained as follows: （29） （30） （31） （32） （33） Finally, all conditions are combined, and the gradient descent method is used to solve the model parameters; (4) constructing a temperature and distance correction model based on FCNN, using the solved emission rate information to correct the temperature and distance, and in addition, designing a loss function fusing a data error term and a physical information error term to provide feedback guidance for the update of FCNN parameters.

2. The method of passive temperature measurement and ranging according to claim 1, wherein: The multispectral infrared-based passive temperature measurement and distance measurement system comprises a signal line (1), a thermometer (2), a carrying handle (3), a protective shell (4), a focusing lens (5), a motor (6), an infrared thermal imager (7), a rotating rod (8), a filter wheel buckle (9) and a filter wheel (10), the protective shell (4) is internally provided with the motor (6), the infrared thermal imager (7) and the filter wheel (10), the motor (6) and the infrared thermal imager (7) are connected with an upper computer through the signal line (1), the motor (6) is provided with the rotating rod (8) at the output end, one end of the rotating rod (8) is connected with the filter wheel (10), the filter wheel (10) is fixed in the protective shell (4) through the filter wheel buckle (9), the filter wheel (10) is provided with four narrow-band filters (11), one full-transparency piece (13) and one light-shielding piece (12) at equal angles, the motor (6) outputs a control signal to make the rotating rod (8) rotate and drive the filter wheel (10) to rotate, so as to switch the narrow-band filter (11) and the light-shielding piece (12), the carrying handle is arranged on the upper end face of the protective shell (4), the thermometer (2) and the focusing lens (5) are arranged outside the protective shell (4), the thermometer (2) is also connected with the upper computer through the signal line (1), and the focusing lens (5) is arranged in front of the protective shell (4), so that external radiant energy is converged and enters the inside of the protective shell (4).

3. The method of passive temperature and range finding according to claim 2, wherein: The step (1) is to build a multispectral LWIR detection system, and the specific process is as follows. The multispectral LWIR detection system is composed of an infrared thermal imager, a filter wheel with four narrow-band filters, a full-transmission filter and a light-shield filter, a thermometer and a focusing lens. According to the specific spectral response curve of the infrared thermal imager in the long-wave infrared band, four long-wave infrared narrow-band filters at the response peak are selected. The spectral band allowed to pass through the narrow-band filter is very narrow. For the spectral channel , the incident radiation is considered to be absorbed only at discrete wavelengths . The detection system switches the narrow-band filter and the light-shield filter by rotating the filter wheel under the control of a motor, measures the atmospheric temperature by the thermometer, and converges the infrared radiation to the system surface by the focusing lens. The use of a chopper to mitigate two special noises present in the LWIR imaging process: "Narcissus effect" and Noise, in the system design, the chopper is introduced, by fast switching narrow band filter and chopper, the camera captures a pair of images containing / without scene radiation, that is, scene and reference, subtracts the two images, see formula (1), obtains the image containing only scene radiation, effectively eliminates "Narcissus effect": （1） wherein and are respectively a scene image and a reference image affected by the "narcissus effect", is an image containing only the scene radiation; By quickly switching narrow-band filters and diaphragms, multiple pairs of images are acquired in a short time, and after subtraction and averaging, the noise is significantly suppressed : （2） wherein, is the number of images taken, is the number of images taken, the image obtained after averaging.

4. The method of passive temperature measurement and ranging according to claim 2, wherein: The step (4) is to build a temperature and distance correction model, and the specific process is as follows. The emissivity information obtained in step (3) is used to correct the temperature and distance, first, a correction model network architecture is built, the network used is FCNN, the network input features are the emissivity, temperature and distance information obtained by solving the target function, the network main body is composed of multiple fully connected layers, which are used to extract the deep nonlinear relationship of the input features, and the activation function ReLU is introduced in each layer to enhance the fitting ability of the model to complex distribution, and finally the corrected temperature and distance values are generated through the output layer, in order to ensure the physical consistency, a loss function is designed to fuse the data error term and the physical information error term, which provides feedback guidance for the update of FCNN parameters, and the specific steps include the following steps: Step 1: verify the rationality of correcting the temperature and distance by using the emissivity, first, calculate the derivative of formula (11) with respect to distance, emissivity and temperature: （34） （35） （36） Eliminate common terms The core contrast of three derivatives simplifies to: （37） (38) (39) Based on formula (22), (34)-(39), the derivative of formula (22) with respect to each parameter is calculated and simplified, it is found that the order of magnitude of the derivative of formula (22) with respect to the emissivity is much larger than that of formula (22) with respect to the distance and the temperature, which indicates that the accuracy of the emissivity is much higher than that of the temperature and the distance, so it is reasonable to correct the temperature and distance by using the emissivity; Step 2: a loss function is designed to fuse the data error term and the physical information error term: (40) wherein, is a data error term measuring the error between the network predicted value and the true value, is a physical information error term, which is a constraint residual based on a physical model, , are weight hyperparameters for the data error term and the physical information error term, respectively. Data error terms are the parameter values predicted by the network and the error between the true values and are computed as follows: (41) wherein, is the number of samples; The physical information error term is the parameter value predicted by the neural network. , The result obtained by substituting into formula (11) and observed values The error is calculated as follows: (42) Finally, the gradient descent method is used to minimize the total loss to update the FCNN parameters.

Images