Spectral imager radiance correction method

Abstract

According to the method for correcting the radiance of the spectral imager, the spectral imager subjected to radiometric calibration is combined with the BP neural network, radiation correction is completed on the spectral imager with known DN value response, and the problem that a traditional method depends on expensive calibration instruments, complex calibration operation and harsh test experiment conditions is solved. An uncalibrated spectral imager and a calibrated spectral imager are only needed to shoot a plurality of images of an unknown scene at the same time, and the spatial geometrical relationship between the shooting angle and the image resolution of the calibrated spectral imager and the uncalibrated spectral imager is analyzed and calculated. And fitting a mapping relation between the DN value image of the calibrated spectral imager and the radiance image of the calibrated spectral imager by using a BP neural network. And on the basis of the numerical relationship of the DN value images shot by the two spectral imagers and the analyzed spatial geometrical relationship, calculating the radiation brightness of the spectral imager by using the fitted mapping relationship and the DN value shot by the uncalibrated spectral imager.

Description

Technical Field

[0001] The present invention relates to the technical field of spectral imaging, and in particular to a method for correcting the radiation brightness of a spectral imager. Background Art

[0002] Radiation correction is mainly used to solve the quantitative conversion relationship between the radiation input received by the optical sensor and the digital signal (Digital Number, DN). It is the premise and basis for the quantification of spectral information. This technology can solve the image quality loss to a certain extent introduced in the links of spectral imaging from solar radiation to ground object reflection, optical imaging, signal acquisition, etc. Usually, spectral imaging instruments are completed through laboratory calibration. Radiation calibration is usually carried out in a darkroom with suitable and stable temperature and humidity. The standard radiation source, integrating sphere, and standard diffuse reflection plate produce uniform radiation brightness, which is then received by the system to be tested. Common radiation correction methods include long-distance small light source method, long-distance extended source method, short-distance extended source method, Jones method, etc. Most of these methods require high-precision distance measurement, high experimental equipment costs (such as collimators, large-size uniform light sources, transmission / reflection extenders), and appropriate external environments (such as calibration fields, lakes with known temperatures, good weather conditions, etc.), and usually only one instrument can be calibrated at a time, which is inefficient. In short, although there are relatively mature methods for radiation correction of spectral imaging instruments, they require expensive instruments, complex operating steps and strict experimental conditions.

[0003] In recent years, with the rise of computational imaging methods, many spectral imagers have directly or indirectly obtained DN values ​​with similar or higher resolution than standard instruments through fitting, super-resolution, deblurring, neural networks, deep learning and other methods. This idea of ​​combining algorithms with instrument design breaks the limits of spectral imaging systems and obtains spectral images that exceed the hardware capabilities of the system at low equipment cost or low sampling cost. It is gradually becoming an important method for spectral imaging equipment design. For such spectral imagers designed or optimized using computational imaging methods, if traditional radiation correction methods or systems are continued, it is neither possible to continue the design idea of ​​computational imaging methods that is mainly based on calculation and supplemented by instrument or system design, nor to give full play to the advantages of computational imaging design ideas in transferring instrument costs to algorithm design. Therefore, it is urgent to design a calibration method based on computational imaging methods to obtain instrument radiation brightness through collected DN value images. Summary of the invention

[0004] The present invention aims to solve at least one of the above-mentioned technical problems existing in the prior art.

[0005] To this end, the present invention provides a method for correcting the radiation brightness of a spectral imager.

[0006] The present invention provides a method for calibrating the radiation brightness of a spectral imager, comprising:

[0007] Acquire several groups of reference spectral images and spectral images to be calibrated taken at different locations; wherein the reference spectral images are taken by a reference spectral imager that has completed radiation correction, and the spectral images to be calibrated are taken by a spectral imager to be calibrated; the reference spectral images and spectral images to be calibrated in the same group are images whose screen centers match;

[0008] Calculate the number of pixels of the spectral image to be calibrated corresponding to the reference spectral image according to the field of view and the offset angle of the reference spectral imager and the spectral imager to be calibrated during shooting, as well as the pixel size of the formed spectral image;

[0009] The DN value of the pixel point on the reference spectrum image is used as the input of the BP neural network, and the radiant brightness of the pixel point after calibration is used as the output of the BP neural network to generate n1 point-to-point BP neural networks, wherein n1 is the number of pixel points of the reference spectrum image; the BP neural network is trained to obtain the mapping relationship between the DN value and the radiant brightness of the reference spectrum image;

[0010] For a pixel point at any spatial position on the spectral image to be calibrated, the coordinates of the pixel point corresponding to the point on the reference spectral image are determined according to the number and size of pixels of the reference spectral image and the spectral image to be calibrated;

[0011] Calculate the spatial correction parameter according to the difference in DN values ​​of the pixel points corresponding to the spatial positions on the reference spectral image and the spectral image to be calibrated;

[0012] Combined with the mapping relationship between the spatial correction parameters and the DN value and the radiation brightness in the BP neural network, the radiation brightness of any point in the spectral image newly taken by the spectral imager to be calibrated is calibrated; wherein the DN value of the image pixel point corrected by the spatial correction parameters is input into the trained BP neural network.

[0013] The method for calibrating the radiation brightness of a spectral imager according to the above technical solution of the present invention may also have the following additional technical features:

[0014] In the above technical solution, the spectral range of the reference spectral imager is larger than the spectral range of the spectral imager to be calibrated.

[0015] In the above technical solution, in the same group of reference spectral images and spectral images to be calibrated, the frame of the reference spectral image is larger than the frame of the spectral image to be calibrated, that is, the vertical field angle of the reference spectral imager when photographing the ground object is larger than the vertical field angle of the spectral image to be calibrated when photographing the ground object, and the horizontal field angle of the reference spectral imager when photographing the ground object is larger than the horizontal field angle of the spectral image to be calibrated when photographing the ground object.

[0016] In the above technical solution, a coordinate system is established with the spatial coordinates of the shooting position of the reference spectral imager as the origin, and the vertical offset angle and the horizontal offset angle of the spectral imager to be calibrated are recorded when shooting; generally, when shooting the same group of images, the shooting angles of the reference spectral imager and the spectral imager to be calibrated are consistent, which can effectively reduce the difficulty of calculation.

[0017] In the above technical solution, the vertical offset angle of the spectral imager to be calibrated is less than half of the vertical field of view angle of the spectral imager to be calibrated; the horizontal offset angle of the spectral imager to be calibrated is less than half of the horizontal field of view angle of the spectral imager to be calibrated.

[0018] In the above technical solution, the step of calculating the number of pixels of the spectral image to be calibrated corresponding to the reference spectral image includes:

[0019] X n2 =X n1 ·u1·(tan(α2 / 2+α3)+tan(α2 / 2-α3)) / 2(tan(α1 / 2)·u2)

[0020] Y n2 =Y n1 ·u1·(tan(β2 / 2+β3)+tan(β2 / 2-β3)) / 2(tan(β1 / 2)·u2)

[0021] Among them, X n1 ×Y n1 is the number of pixels of the reference spectral image, X n2 ×Y n2 is the number of pixels of the spectral image to be calibrated; u1 is the pixel size of the reference spectral image; u2 is the pixel size of the spectral image to be calibrated; α1 is the vertical field angle of the reference spectral imager when photographing the ground object; α2 is the vertical field angle of the spectral imager to be calibrated when photographing the ground object; α3 is the vertical offset angle of the spectral imager to be calibrated when photographing; β1 is the horizontal field angle of the reference spectral imager when photographing the ground object; β2 is the horizontal field angle of the spectral imager to be calibrated when photographing the ground object; β3 is the horizontal offset angle of the spectral imager to be calibrated when photographing.

[0022] In the above technical solution, for a pixel point at any spatial position on the spectral image to be calibrated, the coordinates of the pixel point corresponding to the point on the reference spectral image are determined according to the number and pixel size of the reference spectral image and the spectral image to be calibrated, including:

[0023] A first coordinate system is established with any point on the spectral image to be calibrated as the coordinate origin, and the coordinates (x′, y′) of each pixel point on the spectral image to be calibrated in the first coordinate system are determined;

[0024] The second coordinate system is established with the pixel point on the reference spectrum image that matches the origin of the first coordinate system as the origin of the coordinates. The pixel point on the reference spectrum image corresponding to the pixel point with coordinates (x′, y′) on the spectrum image to be calibrated has the coordinates (x, y) in the second coordinate system:

[0025]

[0026]

[0027] in, Indicates rounding up.

[0028] In the above technical solution, the calculation of the spatial correction parameter according to the difference in DN values ​​of the pixels corresponding to the spatial positions on the reference spectral image and the spectral image to be calibrated includes:

[0029] For any point (x′, y′) on the k′th image captured by the spectral imager to be calibrated, this point is defined as a calibration point. If there are p images that capture this calibration point at the same time in all the spectral images captured by the spectral imager to be calibrated, then the calibration parameter γ of this calibration point is (x′,y′) for:

[0030]

[0031] Among them, λ x ″ ,y′,k′,i represents the DN value of the calibration point on the i-th image among the p images taken by the spectral imager to be calibrated; x,y,k,j represents the DN value of the pixel point (x, y) corresponding to the spatial position of the calibration point on the j-th image in the reference spectral image corresponding to the p-th spectral images to be calibrated; k represents the number of the reference spectral image in the same group as the k′-th image taken by the spectral imager to be calibrated.

[0032] In the above technical solution, for the pixel points on the k′th image captured by the spectral imager to be calibrated that are not captured by other images, the correction parameters are:

[0033]

[0034] Wherein, N represents the number of pixels in the current spectral image to be calibrated that are captured by other images.

[0035] In the above technical solution, in each of the BP neural networks, it is assumed that the number of input neurons is IN M , the number of neurons in the hidden layer is HID L , the number of neurons in the output layer is OUT N; The transfer function used in the hidden layer is the Sigmoid function, and the transfer function in the output layer is the linear function;

[0036] Calculate the estimated value J of the number of neuron nodes in each hidden layer M :

[0037]

[0038] Where q represents an adjustable parameter, which is a constant with a value in [0,10];

[0039] Record the vertical and horizontal field angles of view of each spectral imager when photographing ground objects;

[0040] The training of the BP neural network to obtain a mapping relationship between the DN value and the radiant brightness of the reference spectrum image includes:

[0041] Based on the simulation software, the simulation was carried out, and the feedforward neural network was selected as the basic framework to create the network. The initial weights were defined as non-zero values, and the Levenberg-Marquardt algorithm was used to carry out network training.

[0042] Randomly select images to divide into training set, validation set and test set, and carry out training, set the evaluation function to MSE, and set the number of training iterations; among them, perform time and value normalization on network input, and perform value normalization on network output;

[0043] After training, the BP neural network of all spatial pixels of the benchmark spectral image can converge successfully; for each pixel point with spatial coordinates (x, y) in the benchmark spectral image, the BP neural network generated is N (x,y) ; Take any k-th spectral image, the radiation brightness of any point (x, y) on this image can be calculated through the neural network according to the DN value of the point, that is:

[0044] N (x,y) (λ x,y,k )

[0045] Among them, λ x,y,k Represents the DN value of the pixel (x, y) in the kth spectral image.

[0046] In summary, due to the adoption of the above technical features, the beneficial effects of the present invention are:

[0047] The present invention uses a spectral imager that has completed radiation calibration in combination with a BP neural network to complete radiation correction for a spectral imager with a known DN value response, thereby solving the problem of traditional methods relying on expensive calibration instruments, complex calibration operations, and harsh experimental conditions. It is only necessary to use an uncalibrated spectral imager and a calibrated spectral imager to simultaneously shoot several images of unknown scenes, and record the shooting angles of the two spectral imagers during shooting; analyze and calculate the spatial geometric relationship between the shooting angles and image resolutions of the calibrated spectral imager and the uncalibrated spectral imager. The BP (Back-Propagation) neural network is used to fit the mapping relationship between the DN value image of the calibrated spectral imager and its radiation brightness image. Based on the numerical relationship between the DN value images shot by the two spectral imagers and the analyzed spatial geometric relationship, the radiation brightness is calculated from the DN value shot by the uncalibrated spectral imager using the fitted mapping relationship.

[0048] Specifically, the present invention can calibrate multiple uncalibrated instruments at the same time while completing a single process; the calibration cost is low, and there is no need to use a large number of high-precision measuring instruments. The environmental requirements are low, and spectral images can be taken in an ordinary outdoor environment. There is no need for a specific laboratory environment with constant temperature and humidity, and the technical threshold required for operators is low. Previous methods usually require technicians to use multiple instruments, and the accuracy of operating the instruments has a great impact on the calibration results of the instruments. The present invention only needs to use the automatic acquisition function of the instrument as required. Subsequent calculation processes can be solidified into software functions through programming. The calibrated spectral imaging accuracy can be used in a wider range of commercial applications, such as environmental monitoring, agricultural production, etc.

[0049] Additional aspects and advantages of the present invention will become apparent from the following description or may be learned by practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which:

[0051] Figure 1 is a flow chart of a method for calibrating the radiation brightness of a spectral imager according to an embodiment of the present invention;

[0052] Figure 2 It is a module diagram of a method for radiant brightness correction of a spectral imager according to an embodiment of the present invention;

[0053] Figure 3 It is the convergence error distribution of the whole image after BP neural network training in the spectral imager radiation brightness correction method of one embodiment of the present invention;

[0054] Figure 4It is a schematic diagram of the changes of training, verification and test errors at a certain point in the method for radiance correction of a spectral imager according to an embodiment of the present invention;

[0055] Figure 5 is a radiance image captured by a reference spectral imager in a specific embodiment of the present invention;

[0056] Figure 6 It is a radiation brightness map obtained by the spectral imager to be calibrated in a specific embodiment of the present invention using the DN value image as input and according to the mapping relationship between the correction coefficient and the spatial correction;

[0057] Figure 7 is based on Figure 6 Schematic diagram of the difference between the RAD curve of the spectral imager to be calibrated obtained by fitting at point A and the standard RAD curve of the reference spectral imager;

[0058] Figure 8 is based on Figure 6 Schematic diagram of the difference between the RAD curve of the spectral imager to be calibrated obtained by fitting at point B and the standard RAD curve of the reference spectral imager. DETAILED DESCRIPTION

[0059] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0060] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited to the specific embodiments disclosed below.

[0061] Refer to the following Figures 1 to 8 The following describes a method for spectral imager radiation brightness correction according to some embodiments of the present invention.

[0062] Some embodiments of the present application provide a method for calibrating radiation brightness of a spectral imager.

[0063] like Figure 1 and Figure 2 As shown, the first embodiment of the present invention proposes a method for calibrating the radiation brightness of a spectral imager, including steps S1-S6.

[0064] S1. Acquire several groups of reference spectral images and spectral images to be calibrated taken at different locations; wherein the reference spectral images are taken by a reference spectral imager that has completed radiation correction, and the spectral images to be calibrated are taken by a spectral imager to be calibrated; the reference spectral images and spectral images to be calibrated in the same group are images whose centers match.

[0065] It can be understood that in each group of images, the spectral image to be calibrated corresponds to the reference spectral image, and each spectral image to be calibrated has a corresponding reference spectral image. Image center matching means that the image centers of the two images are the same or approximately the same. In general, the captured scene is required to contain as many different types of objects as possible, and the spectral range of the reference spectral imager is required to be larger than the spectral range of the spectral imager to be calibrated, to ensure that the reference spectral imager has a more accurate reference significance.

[0066] In addition, in the same group of reference spectral images and spectral images to be calibrated, the frame of the reference spectral image is larger than that of the spectral image to be calibrated, that is, the vertical field of view of the reference spectral imager when photographing the ground object is larger than that of the spectral image to be calibrated, and the horizontal field of view of the reference spectral imager when photographing the ground object is larger than that of the spectral image to be calibrated. This ensures that each pixel point in the spectral image to be calibrated can find a corresponding pixel point in the reference spectral image during subsequent calibration.

[0067] During each shooting, a coordinate system is established with the spatial coordinates of the shooting position of the reference spectral imager as the origin, and the vertical offset angle and horizontal offset angle of the spectral imager to be calibrated are recorded; when shooting the same group of images, the shooting angles of the reference spectral imager and the spectral imager to be calibrated are consistent. In some embodiments, the vertical offset angle of the spectral imager to be calibrated should be set to be less than half of the vertical field of view of the spectral imager to be calibrated; and the horizontal offset angle of the spectral imager to be calibrated should be less than half of the horizontal field of view of the spectral imager to be calibrated.

[0068] S2. Calculate the number of pixels of the spectral image to be calibrated corresponding to the reference spectral image according to the field angle and offset angle of the reference spectral imager and the spectral imager to be calibrated during shooting, as well as the pixel size of the formed spectral image.

[0069] It should be noted that the field of view angle and offset angle of the reference spectral imager and the spectral imager to be calibrated during shooting, as well as the pixel size of the spectral image are all known quantities obtained in S1. Under the condition that the centers of the reference spectral image and the spectral image to be calibrated match, the number of pixels of the spectral image to be calibrated corresponding to the reference spectral image can be calculated based on the number of pixels of the reference spectral image using the above-mentioned known quantities. The calculation method is as follows:

[0070] Xn2 =X n1 ·u1·(tan(α2 / 2+α3)+tan(α2 / 2-α3)) / 2(tan(α1 / 2)·u2)

[0071] Y n2 =Y n1 ·u1·(tan(β2 / 2+β3)+tan(β2 / 2-β3)) / 2(tan(β1 / 2)·u2)

[0072] Among them, X n1 ×Y n1 is the number of pixels of the reference spectral image, X n2 ×Y n2 is the number of pixels of the spectral image to be calibrated; u1 is the pixel size of the reference spectral image; u2 is the pixel size of the spectral image to be calibrated; α1 is the vertical field angle of the reference spectral imager when photographing the ground object; α2 is the vertical field angle of the spectral imager to be calibrated when photographing the ground object; α3 is the vertical offset angle of the spectral imager to be calibrated when photographing; β1 is the horizontal field angle of the reference spectral imager when photographing the ground object; β2 is the horizontal field angle of the spectral imager to be calibrated when photographing the ground object; β3 is the horizontal offset angle of the spectral imager to be calibrated when photographing.

[0073] S3, using the DN value of the pixel point on the reference spectrum image as the input of the BP neural network, and using the radiant brightness of the pixel point after calibration as the output of the BP neural network, generate n1 point-to-point BP neural networks, where n1 is the number of pixels in the reference spectrum image, that is, X n1 ×Y n1 ; Train the BP neural network to obtain the mapping relationship between the DN value and the radiation brightness of the reference spectrum image.

[0074] It should be noted that the radiation brightness of the pixel points in the reference spectral image is usually obtained by laboratory calibration to obtain the corresponding parameters. The specific calibration method can adopt any of the existing methods and can be expressed in the form of calculation through the standard correction equation. However, an accurate calibration method should be selected as much as possible to improve the accuracy of the reference. The specific calibration method is not limited here.

[0075] In some embodiments, in each of the BP neural networks, it is assumed that the number of input neurons is IN M , the number of neurons in the hidden layer is HID L , the number of neurons in the output layer is OUT N ; The transfer function used in the hidden layer is the Sigmoid function, and the transfer function in the output layer is the linear function;

[0076] Calculate the estimated value J of the number of neuron nodes in each hidden layerM :

[0077]

[0078] Where q is an adjustable parameter, which is a constant with a value between [0,10] and can be adjusted according to the specific problem.

[0079] Record the vertical and horizontal field angles of view of each spectral imager when photographing ground objects;

[0080] The training of the BP neural network to obtain a mapping relationship between the DN value and the radiant brightness of the reference spectrum image includes:

[0081] Based on simulation software (such as MATLAB, Python, etc.), simulation is carried out, and a feedforward neural network is selected as the basic framework to create a network. The initial weight is defined as a small non-zero value, and the Levenberg-Marquardt algorithm is used to carry out network training;

[0082] Randomly select images to divide into training set, validation set and test set, and carry out training, set the evaluation function to MSE, and set the number of training iterations; wherein, perform time and value normalization processing on network input, and perform value normalization processing on network output; in a specific embodiment, the ratio of training set, validation set and test set is about 7:2:1, and the maximum number of training iterations is set to 1000 times;

[0083] After training, the BP neural network of all spatial pixels of the benchmark spectral image can converge successfully; for each pixel point with spatial coordinates (x, y) in the benchmark spectral image, the BP neural network generated is N (x,y) ; Take any k-th spectral image, the radiation brightness of any point (x, y) on this image can be calculated through the neural network according to the DN value of the point, that is:

[0084] N (x,y) (λ x,y,k )

[0085] Among them, λ x,y,k Represents the DN value of the pixel (x, y) in the kth spectral image.

[0086] In a specific embodiment, Figure 3 The technical effect diagram when the BP neural network converges after the above training is shown, that is, the convergence error distribution of the whole diagram; Figure 4Taking the coordinate point (158, 137) as an example and MSE as the evaluation standard, the changes in training, validation and test errors are shown, where the horizontal axis is the training round, the vertical axis is MSE (mean square error), the blue line represents the training error, the green line represents the validation error, and the red line represents the test error. The best validation performance is 0.00010718 at 13 rounds.

[0087] It is understandable that through step S3, the corresponding relationship between the DN value and the radiation brightness of each pixel point of the image captured by the reference spectral imager can be obtained, and this corresponding relationship is associated with the coordinates of the pixel point.

[0088] S4. For any pixel point at a spatial position on the spectral image to be calibrated, the pixel coordinates corresponding to the point on the reference spectral image are determined according to the number and size of pixels of the reference spectral image and the spectral image to be calibrated.

[0089] In some embodiments, step S4 includes:

[0090] A first coordinate system is established with any point on the spectral image to be calibrated as the coordinate origin, and the coordinates (x′, y′) of each pixel point on the spectral image to be calibrated in the first coordinate system are determined;

[0091] The second coordinate system is established with the pixel point on the reference spectrum image that matches the origin of the first coordinate system as the origin of the coordinates. The pixel point on the reference spectrum image corresponding to the pixel point with coordinates (x′, y′) on the spectrum image to be calibrated has the coordinates (x, y) in the second coordinate system:

[0092]

[0093]

[0094] in, Indicates rounding up.

[0095] In a specific embodiment, the coordinate origin of the first coordinate system is usually a vertex of the spectral image to be calibrated, and the coordinate origin of the second coordinate system is a vertex on the reference spectral image corresponding to the direction of the vertex.

[0096] According to step S4, the spatial correspondence between the pixel points on the spectral image to be calibrated and the pixel points on the reference spectral image can be determined.

[0097] S5. Calculate the spatial correction parameter according to the difference in DN values ​​between the pixel points corresponding to the spatial positions on the reference spectral image and the spectral image to be calibrated.

[0098] Specifically, step S5 includes:

[0099] For any point (x′, y′) on the k′th image captured by the spectral imager to be calibrated, this point is defined as a calibration point. If there are p images that capture this calibration point at the same time in all the spectral images captured by the spectral imager to be calibrated, then the calibration parameter γ of this calibration point is (x′,y′) for:

[0100]

[0101] Among them, λ x ″ ,y′,k′,i represents the DN value of the calibration point on the i-th image among the p images taken by the spectral imager to be calibrated; x,y,k,j represents the DN value of the pixel point (x, y) corresponding to the spatial position of the calibration point on the j-th image in the reference spectral image corresponding to (in the same group as) the p-th spectral images to be calibrated; k represents the number of the reference spectral image in the same group as the k′-th image taken by the spectral imager to be calibrated.

[0102] In addition, when there are some pixels on the current spectral image to be calibrated that are not captured by other spectral images to be calibrated at the same time, the correction parameters of these pixels can be taken as the average of the correction parameters of the remaining pixels on the current spectral image to be calibrated. Specifically: For the pixels on the k′th image captured by the spectral imager to be calibrated that are not captured by other images, their correction parameters are:

[0103]

[0104] Wherein, N represents the number of pixels in the current spectral image to be calibrated that are captured by other images.

[0105] S6. Combine the spatial correction parameters and the mapping relationship between DN value and radiation brightness in the BP neural network to calibrate the radiation brightness of any point in the spectral image newly taken by the spectral imager to be calibrated; wherein the DN value of the image pixel point corrected by the spatial correction parameters is input into the trained BP neural network.

[0106] Specifically, for any point (x0, y0) in the spectral image newly taken by the spectral imager to be calibrated, according to step S4, it can be determined that the corresponding pixel point on the reference spectral image is (x, y), and the corresponding trained BP neural network is N (x,y) According to step S5, the correction parameter of this point can be determined as Set the DN value of the point (x0, y0) First use the correction parameters to correct it and then use it as the input of the BP neural network to get the radiation brightness of the corresponding point: By traversing the entire image in accordance with this step, a radiation brightness map of the scene photographed by the spectral imager to be calibrated is obtained.

[0107] In some embodiments, when executing step S1, each group of images is composed of a reference spectrum image and multiple spectra to be calibrated taken by different spectral imagers to be calibrated, so that calibration of multiple spectral imagers to be calibrated can be completed at one time.

[0108] The present invention provides a specific example of using the spectral imager radiation brightness correction method disclosed in the present invention to realize the calibration of the spectral imager, as shown below.

[0109] The reference spectral imager is the calibrated HySpex Baldur V-1024 (device model), and the spectral imager to be calibrated is the uncalibrated PSD-HIS (device model) of Spectrum Vision Corporation.

[0110] The instrument parameters corresponding to the reference spectral imager are α1=β1=16.33°, X n1 =Y n1 =1024;

[0111] The instrument parameters corresponding to the spectral imager to be calibrated are α2=β2=9.28°;

[0112] The angles of the two images are the same when shooting, i.e., α3 = β3 = 0°, and the pixel sizes of the two images are u1 = u2 = 6.5um;

[0113] According to step S2, the spatial pixel size of the image captured by the spectral imager to be calibrated at equal distance is calculated as X n2 ×Y n2 =568×568;

[0114] The spectral range of the instrument HySpex Baldur V-1024 (400-1000nm) is larger than the spectral range of PSD-HIS (450-720nm). Since PSD-HIS has 55 bands in its spectral range, the dimension of the network input variable is set to 55 in the BP neural network.

[0115] In the BP neural network training, MATLAB is selected as the simulation platform. Take the cumulative number of scene picture groups P = 10, that is, IN M =OUT N =10. Take q=5, calculate the number of neuron nodes in each hidden layer, and round it up to J M = 10. The ratio of training set, validation set, and test set is 7:2:1.

[0116] After 1000 trainings, the mean of the best test error is 0.0044 and the maximum is 0.9475. Since the shooting angles are consistent, N = 568*568 = 322624. It should be noted that since the computational spectral imaging instrument PSD-HIS uses HySpex Baldur V-1024 to complete the DN value correction, there is a γ (x′,y′) =1, no further calculation is required.

[0117] The radiance of the corrected HySpex Baldur V-1024 and the PSD-HIS calculated by the present invention are compared on the test spectrum image. The results are shown in the attached Figure 5-8 middle. Figure 5 Standard radiance obtained for HySpex Baldur V-1024, Figure 6 is the radiant brightness of PSD-HIS obtained by the method of the present invention. Figure 6 The feature points with spatial coordinates A(115,180) and B(242,122) are Figure 5 The corresponding coordinate points are (286,318) and (349,284) respectively. That is, select N in the trained BP neural network. (286,318) and N (349，284) Two neural networks, using the DN value of each point as network input, obtain the corrected radiation brightness N (286,318) (λ (115,180) ) and N (349，284) (λ (242,122) ). The results are shown in Figure 7 and 8 In the experiment, the difference between the algorithm-corrected radiation brightness curve and the standard radiation brightness curve is small, and the curve shapes are basically the same, which can be used for low-cost batch calibration of radiation brightness of commercial instruments.

[0118] In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0119] Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for calibrating the radiation brightness of a spectral imager, characterized in that: include: Acquire several groups of reference spectral images and spectral images to be calibrated taken at different locations; wherein the reference spectral images are taken by a reference spectral imager that has completed radiation correction, and the spectral images to be calibrated are taken by a spectral imager to be calibrated; the reference spectral images and spectral images to be calibrated in the same group are images whose screen centers match; Calculate the number of pixels of the spectral image to be calibrated corresponding to the reference spectral image according to the field of view and the offset angle of the reference spectral imager and the spectral imager to be calibrated during shooting, as well as the pixel size of the formed spectral image; The DN value of the pixel point on the reference spectrum image is used as the input of the BP neural network, and the radiant brightness of the pixel point after calibration is used as the output of the BP neural network to generate n1 point-to-point BP neural networks, wherein n1 is the number of pixel points of the reference spectrum image; the BP neural network is trained to obtain the mapping relationship between the DN value and the radiant brightness of the reference spectrum image; For a pixel point at any spatial position on the spectral image to be calibrated, the coordinates of the pixel point corresponding to the point on the reference spectral image are determined according to the number and size of pixels of the reference spectral image and the spectral image to be calibrated; Calculate the spatial correction parameter according to the difference in DN values ​​of the pixel points corresponding to the spatial positions on the reference spectral image and the spectral image to be calibrated; Combined with the mapping relationship between the spatial correction parameters and the DN value and the radiation brightness in the BP neural network, the radiation brightness of any point in the spectral image newly taken by the spectral imager to be calibrated is calibrated; wherein the DN value of the image pixel point corrected by the spatial correction parameters is input into the trained BP neural network.

2. The method for calibrating the radiation brightness of a spectral imager according to claim 1, characterized in that: The spectral range of the reference spectral imager is greater than the spectral range of the spectral imager to be calibrated.

3. The method for calibrating the radiation brightness of a spectral imager according to claim 1, characterized in that: In the same group of reference spectral images and spectral images to be calibrated, the frame of the reference spectral image is larger than that of the spectral image to be calibrated, that is, the vertical field angle of the reference spectral imager when photographing the ground object is larger than the vertical field angle of the spectral image to be calibrated when photographing the ground object, and the horizontal field angle of the reference spectral imager when photographing the ground object is larger than the horizontal field angle of the spectral image to be calibrated when photographing the ground object.

4. The method for calibrating the radiation brightness of a spectral imager according to claim 3, characterized in that: A coordinate system is established with the spatial coordinates of the shooting position of the reference spectral imager as the origin, and the vertical offset angle and the horizontal offset angle of the spectral imager to be calibrated when shooting are recorded.

5. The method for calibrating the radiation brightness of a spectral imager according to claim 4, characterized in that: The vertical offset angle of the spectral imager to be calibrated is less than half of the vertical field angle of the spectral imager to be calibrated; the horizontal offset angle of the spectral imager to be calibrated is less than half of the horizontal field angle of the spectral imager to be calibrated.

6. The method for calibrating the radiation brightness of a spectral imager according to claim 1, characterized in that: The step of calculating the number of pixels of the spectral image to be calibrated corresponding to the reference spectral image comprises: X n2 =X n1 ·u1·(tan(α2 / 2+α3)+tan(α2 / 2-α3)) / 2(tan(α1 / 2)·u2) AND n2 =And n1 ·u1·(tan(β2 / 2+β3)+tan(β2 / 2-β3)) / 2(tan(β1 / 2)·u2) Among them, X n1 ×Y n1 is the number of pixels of the reference spectral image, X n2 ×Y n2 is the number of pixels of the spectral image to be calibrated; u1 is the pixel size of the reference spectral image; u2 is the pixel size of the spectral image to be calibrated; α1 is the vertical field angle of the reference spectral imager when photographing the ground object; α2 is the vertical field angle of the spectral imager to be calibrated when photographing the ground object; α3 is the vertical offset angle of the spectral imager to be calibrated when photographing; β1 is the horizontal field angle of the reference spectral imager when photographing the ground object; β2 is the horizontal field angle of the spectral imager to be calibrated when photographing the ground object; β3 is the horizontal offset angle of the spectral imager to be calibrated when photographing.

7. The method for calibrating the radiation brightness of a spectral imager according to claim 6, characterized in that: For a pixel point at any spatial position on the spectral image to be calibrated, determining the pixel point coordinates corresponding to the point on the reference spectral image according to the number and pixel size of the reference spectral image and the spectral image to be calibrated, comprises: A first coordinate system is established with any point on the spectral image to be calibrated as the coordinate origin, and the coordinates (x′, y′) of each pixel point on the spectral image to be calibrated in the first coordinate system are determined; The second coordinate system is established with the pixel point on the reference spectrum image that matches the origin of the first coordinate system as the origin of the coordinates. The pixel point on the reference spectrum image corresponding to the pixel point with coordinates (x′, y′) on the spectrum image to be calibrated has the coordinates (x, y) in the second coordinate system: in, Indicates rounding up.

8. The method for calibrating the radiation brightness of a spectral imager according to claim 7, characterized in that: The calculating of the spatial correction parameter according to the difference in DN values ​​of the pixel points corresponding to the spatial positions on the reference spectral image and the spectral image to be calibrated comprises: For any point (x′, y′) on the k′th image captured by the spectral imager to be calibrated, this point is defined as a calibration point. If there are p images that capture this calibration point at the same time in all the spectral images captured by the spectral imager to be calibrated, then the calibration parameter γ of this calibration point is (x′,y′) for: Among them, λ′ x′,y′,k′,i represents the DN value of the calibration point on the i-th image among the p images taken by the spectral imager to be calibrated; x,y,k,j represents the DN value of the pixel point (x, y) corresponding to the spatial position of the calibration point on the j-th image in the reference spectral image corresponding to the p-th spectral images to be calibrated; k represents the number of the reference spectral image in the same group as the k′-th image taken by the spectral imager to be calibrated.

9. The method for calibrating the radiation brightness of a spectral imager according to claim 8, characterized in that: For the pixel points captured by the spectral imager to be calibrated on the k′th image that are not captured by other images, the calibration parameters are: Wherein, N represents the number of pixels in the current spectral image to be calibrated that are captured by other images.

10. The method for calibrating the radiation brightness of a spectral imager according to claim 1, characterized in that: In each of the BP neural networks, assume that the number of input neurons is IN M , the number of neurons in the hidden layer is HID L , the number of neurons in the output layer is OUT N ; The transfer function used in the hidden layer is the Sigmoid function, and the transfer function in the output layer is the linear function; Calculate the estimated value J of the number of neuron nodes in each hidden layer M : Where q represents an adjustable parameter, which is a constant with a value in [0,10]; Record the vertical and horizontal field angles of view of each spectral imager when photographing ground objects; The training of the BP neural network to obtain a mapping relationship between the DN value and the radiant brightness of the reference spectrum image includes: Based on the simulation software, the simulation was carried out, and the feedforward neural network was selected as the basic framework to create the network. The initial weights were defined as non-zero values, and the Levenberg-Marquardt algorithm was used to carry out network training. Randomly select images to divide into training set, validation set and test set, and carry out training, set the evaluation function to MSE, and set the number of training iterations; among them, perform time and value normalization on network input, and perform value normalization on network output; After training, the BP neural network of all spatial pixels of the benchmark spectral image can converge successfully; for each pixel point with spatial coordinates (x, y) in the benchmark spectral image, the BP neural network generated is N (x,y) ; Take any k-th spectral image, the radiation brightness of any point (x, y) on this image can be calculated through the neural network according to the DN value of the point, that is: N (x,y) (l x,y,k ) Among them, λ x,y,k Represents the DN value of the pixel (x, y) in the kth spectral image.

Images