Non-uniformity correction method based on unsupervised machine learning

Abstract

The application discloses a non-uniformity correction method based on unsupervised machine learning, which comprises the following steps: system operating point calibration at different blackbody temperatures; establishing a unified calibration database; building a clustering-based unsupervised learning network and performing regression training; deploying the trained network parameters to perform energy domain non-uniformity correction on actual scene images. Based on the derived infrared physical model, a unified database including blackbody gray value, integration time and attenuation sheet is established. In the absence of blackbody theoretical radiation quantity labels, the clustering-based unsupervised learning method is used, the average predicted energy domain value at the same blackbody temperature is taken as the learning target, and the correlation between different operating point calibration data is fully mined through network training, so that the coefficients of the correction model are obtained. The application still maintains uniform and stable output after adjusting the integration time and the attenuation sheet.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of infrared image processing, and particularly relates to a non-uniformity correction method based on unsupervised machine learning. BACKGROUND

[0002] Infrared thermal imaging technology has a wide range of applications in night vision detection, autonomous driving, medical research, and power fault diagnosis. Compared with visible light imaging systems, infrared systems usually capture images with significantly more noise, which is caused by theoretical limitations, detector manufacturing technology, and the inherent characteristics of infrared imaging. One key factor affecting the quality of infrared image processing is the non-uniformity of the detector pixel response, which is represented as fixed pattern noise (FPN). Fixed pattern noise often appears as stripes, blocks, and gradual light and dark changes in the image, which has a significant impact on the quality of infrared imaging.

[0003] Common blackbody calibration-based correction methods include one-point correction, two-point correction, and multi-point correction algorithms. In addition, to expand the detection range of the system, the integration time is usually adjusted and an attenuation sheet is added. When the integration time or the transmittance of the attenuation sheet changes, the image gray scale also changes, which is not conducive to the subsequent image processing. At present, there are few studies on non-uniformity correction algorithms based on calibration that consider the joint effects of integration time and attenuation sheet. Adjusting the integration time of the system and adding an attenuation sheet both introduce additional non-uniformity. A relatively effective method to solve the above problems is to convert the gray scale domain correction to the energy domain correction, but the non-uniformity correction method based on the energy domain usually requires accurate knowledge of the blackbody temperature, emissivity, and working wavelength to calculate the theoretical radiation.

[0004] The conventional gray scale domain-based non-uniformity correction method has the following disadvantages:

[0005] 1. Fixed attenuation sheet and integration time: Due to factors such as storage capacity and calibration workload, the system can only store a limited number of attenuation sheet and integration time combination working points, and the best measurement working point cannot be dynamically selected during use, making it difficult to fully utilize the performance of the infrared radiation measurement system.

[0006] 2. Poor correction effect: The parameter calculation of the conventional method is performed independently for each attenuation sheet and integration time group, and the correlation between different calibration data is ignored. The calibration accuracy is easily affected by random errors in the system, resulting in poor non-uniformity correction effect when switching between working points.

[0007] 3. When the integration time or the transmittance of the attenuation sheet changes, the image gray scale also changes, which is not conducive to the subsequent image processing.

[0008] And the conventional non-uniformity correction method based on energy domain has the following disadvantages:

[0009] The non-uniformity correction method based on energy domain usually needs to accurately know the blackbody temperature, emissivity and working waveband to calculate the theoretical radiation. Deviation of the blackbody temperature, emissivity and working waveband from the true value will affect the calibration accuracy, and then affect the correction effect. SUMMARY

[0010] To solve the above technical problems, the application provides a non-uniformity correction method based on unsupervised machine learning, a unified database including infrared image gray, integration time and attenuation sheet information is established based on the derived infrared physical model, and the average value of the predicted radiance under the same blackbody temperature is used as the learning target to learn the correction coefficient of the model by using the clustering-based unsupervised learning method. The application does not need to know the true blackbody temperature, emissivity and working waveband of the calibration point. The correction network trained can maintain the uniform and stable output of the image after the integration time and attenuation sheet adjustment, which is beneficial to the post-processing of the infrared system.

[0011] Based on the derived infrared physical model, a unified database including blackbody gray value, integration time and attenuation sheet is established, and in the absence of blackbody theoretical radiation label, the average predicted energy domain value under the same blackbody temperature is used as the learning target by using the clustering-based unsupervised learning method, and the correlation between different working point calibration data is fully mined through network training, so as to obtain each coefficient of the correction model.

[0012] To achieve the above purpose, the technical scheme adopted by the application is:

[0013] A non-uniformity correction method based on unsupervised machine learning, comprising the following steps:

[0014] Step 1, system working point calibration under different blackbody temperatures;

[0015] Step 2, establishing a unified calibration database;

[0016] Step 3, building a clustering-based unsupervised learning network and performing regression training;

[0017] Step 4, deploying the trained network parameters to perform energy domain non-uniformity correction on the actual scene image.

[0018] Further, the step 1 comprises:

[0019] Calibration is performed on multiple attenuation sheets, integration times and blackbody temperature points of the system, and multiple frames of original data of the image gray response changing with the attenuation sheet, integration time and blackbody temperature are obtained.

[0020] Further, step 2 includes:

[0021] The multi-frame raw data from step 1 is preprocessed, including multi-frame denoising, averaging or replacing pixels around blind pixels.

[0022] Further, step 3 includes:

[0023] Establish a feedforward neural network for regression and set various network parameters, including: input layer, regression layer, output layer, loss function, backpropagation function, activation function, learning rate, and number of iterations;

[0024] The preprocessed normalized grayscale response data, normalized detector integration time, and attenuator enable signal obtained in step 2 are used as inputs to the regression neural network; the average value of the predicted energy domain at the same blackbody temperature is used as the expected value of the regression neural network.

[0025] Furthermore, the average value of the predicted energy range at the same blackbody temperature is shown in equation (1):

[0026] (1)

[0027] in, , where is the average value of the predicted energy domain at the same blackbody temperature, [i,j] is the pixel coordinate, W is the number of pixels in the detector width, H is the number of pixels in the detector height, and M is the number of calibration points at a certain blackbody temperature;

[0028] Establish a unified mathematical regression equation (2):

[0029] (2)

[0030] in, For the normalized average predicted value term, To normalize the pixel grayscale value, For normalized integral time, The regression coefficients represent the overall system gain. The regression coefficients for the system detector bias term. The stray radiation bias regression coefficient of the optical system. Let be the regression coefficient of the radiation bias of the k-th attenuator. is the regression coefficient of the transmittance term of the k-th attenuator.

[0031] Furthermore, based on equation (5), the relative mean square error loss function and its partial derivative are established:

[0032] (5)

[0033] in, The value of the loss function. Let be the relative error backpropagation function of the loss function, and let be the partial derivative of the loss function with respect to the radiance of the pixel at coordinate [i,j]. Let be the predicted regression value corresponding to the nth calibrated blackbody image at coordinate [i,j]. The average value of the predicted energy domain at the same blackbody temperature is given by N, where N is the number of radiation calibration points, W is the number of pixels in the detector width, and H is the number of pixels in the detector height. The reciprocals of the coefficients in the denominator of equation (2) are considered as a single term. The error backpropagation function is shown in equation (6).

[0034] (6)

[0035] in, Let [i,j] be the overall system gain at pixel [i,j]. Input a selection signal for the k-th attenuator. Let [i,j] be the gray value of the pixel at coordinates [i,j]. Let k be the transmittance of the attenuator. For integration time, The system detector offset is set at pixel coordinates [i,j]. The partial derivative of the mean square error loss function, Let W be the stray radiation from the optical system at pixel [i,j], W be the number of pixels in the detector width, and H be the number of pixels in the detector height. Let be the radiation amount of the k-th attenuator. yes right The partial derivative, yes right The partial derivative, yes right The partial derivative, It is the kth attenuator. right The partial derivative, It is the kth attenuator. right The partial derivative;

[0036] The iterative changes of each coefficient are shown in equation (7):

[0037] (7)

[0038] in, For the overall system gain, Let k be the transmittance of the attenuator. For system detector bias, For stray radiation from optical systems, Let be the radiation amount of the k-th attenuator. yes right The partial derivative, yes right The partial derivative, yes right The partial derivative, It is the kth attenuator. right The partial derivative, It is the kth attenuator. right The partial derivative, , , , , These represent the learning rates for each coefficient.

[0039] Furthermore, the training methods include:

[0040] Import the calibration data of the first attenuator from the training library into the learning network model, and iterate only the learning coefficients. , , When the value of the loss function changes slowly, i.e., less than 10... -4 The second decaying data is added to the learning network model; simultaneously, the coefficients are... , , The learning rate decreased to 10 -6 Only iterate through the learning coefficients , , , When the value of the loss function changes slowly, the coefficient of recovery... , , The learning rate is determined; finally, the network is trained to obtain globally optimized, non-uniform correction network parameters that meet accuracy requirements.

[0041] Furthermore, the hardware platform for the training network includes a central processing unit, a graphics processing unit, an advanced reduced instruction set computer architecture, or an application-specific integrated circuit.

[0042] Further, step 4 includes:

[0043] The trained parameters are deployed into a non-uniformity correction network, and real infrared scene images are captured under different attenuator settings and integration times. The normalized original grayscale response, integration time, and attenuator information are used as network inputs to perform inference calculations on the average predicted value of the pixel energy domain.

[0044] The advantages of this invention compared to the prior art are as follows:

[0045] (1) It is not necessary to know the blackbody temperature, emissivity, and operating band to calculate the theoretical radiation. The calibration accuracy will not be affected by the deviation of the blackbody temperature, emissivity, and operating band from the true value, and thus the correction effect will not be affected;

[0046] (2) The system operating point is continuously adjustable: The training correction network has good generalization within the system operating range and can perform non-uniform correction on actual infrared images with uncalibrated operating point settings within the operating range. This overcomes the shortcomings of the previous fixed attenuator and integration time settings and large adjustment range, making the adjustment of operating point parameters more delicate and flexible, and improving the non-uniform correction capability of the infrared system for targets with large dynamic range.

[0047] (3) The corrected image will not experience grayscale jumps due to changes in integration time or attenuator. It can maintain uniform and stable image output when the working point changes, which is beneficial for the post-processing of the infrared system. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of an unsupervised machine learning model.

[0049] Figure 2 Training curve for the loss function;

[0050] Figure 3 This is a schematic diagram showing the fitting results and relative error at the working point;

[0051] Figure 4 The images and their grayscale histograms for the two-point correction algorithm at different integration times are shown below. Among them, (a) is the image of the two-point correction algorithm with an integration time of 1920us, (b) is the image of the two-point correction algorithm with an integration time of 2240us, (c) is the image of the two-point correction algorithm with an integration time of 2560us, (d) is the histogram of the image of the two-point correction algorithm with an integration time of 1920us, (e) is the histogram of the image of the two-point correction algorithm with an integration time of 2240us, and (f) is the histogram of the image of the two-point correction algorithm with an integration time of 2560us.

[0052] Figure 5The following images and their histograms are provided for the algorithm proposed in this invention at different integration times: (a) is the image of the algorithm proposed in this invention at an integration time of 1920us; (b) is the image of the algorithm proposed in this invention at an integration time of 2240us; (c) is the image of the algorithm proposed in this invention at an integration time of 2560us; (d) is the histogram of the image of the algorithm proposed in this invention at an integration time of 1920us; (e) is the histogram of the image of the algorithm proposed in this invention at an integration time of 2240us; and (f) is the histogram of the image of the algorithm proposed in this invention at an integration time of 2560us.

[0053] Figure 6 The images and their grayscale histograms for the two-point correction algorithm under different attenuator settings are shown below. Among them, (a) is the image corrected by the two-point correction algorithm when the attenuator transmittance is 1, (b) is the image corrected by the two-point correction algorithm when the attenuator transmittance is 0.88, (c) is the histogram corresponding to the image corrected by the two-point correction algorithm when the attenuator transmittance is 1, and (d) is the histogram corresponding to the image corrected by the two-point correction algorithm when the attenuator transmittance is 0.88.

[0054] Figure 7 The present invention provides the corrected images and histograms of the algorithm proposed in this invention under different attenuation filter settings; wherein, (a) is the corrected image of the algorithm proposed in this invention when the attenuator transmittance is 1, (b) is the corrected image of the algorithm proposed in this invention when the attenuator transmittance is 0.88, (c) is the histogram of the corrected image of the algorithm proposed in this invention when the attenuator transmittance is 1, and (d) is the histogram of the corrected image of the algorithm proposed in this invention when the attenuator transmittance is 0.88.

[0055] Figure 8 The images and their corresponding histograms are shown for the calibrated and uncalibrated operating points of the proposed method of this invention. Among them, (a) is the image calibrated by the proposed algorithm of this invention when the integration time is 1472us, (b) is the image calibrated by the proposed algorithm of this invention when the integration time is 1920us, (c) is the image calibrated by the proposed algorithm of this invention when the integration time is 2368us, (d) is the histogram corresponding to the image calibrated by the proposed algorithm of this invention when the integration time is 1472us, (e) is the histogram corresponding to the image calibrated by the proposed algorithm of this invention when the integration time is 1920us, and (f) is the histogram corresponding to the image calibrated by the proposed algorithm of this invention when the integration time is 2368us. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other. The embodiments of this invention will now be described in detail with reference to the accompanying drawings.

[0057] like Figure 1 As shown, an embodiment of the present invention provides a non-uniformity correction method based on unsupervised machine learning, which includes the following steps:

[0058] Step 1: Calibrate the operating point of the mid-wave infrared system at different blackbody temperatures, including: calibrating at multiple attenuators, integration times, and blackbody temperatures of the system, and obtaining multiple frames of raw data showing the change of image grayscale response with attenuators and integration times.

[0059] Step 2: Establish a unified calibration database, including: preprocessing the multi-frame raw data for each working point by averaging and denoising, averaging or replacing pixels around blind cells, etc.; using the obtained preprocessed normalized grayscale response data, normalized detector integration time, and attenuator enable signal as input to the regression neural network; and simultaneously calculating the average value of the predicted energy domain at the same blackbody temperature, including:

[0060] The average predicted value is calculated according to equation (1) based on the initial values ​​of the system parameters:

[0061] (1)

[0062] in, Let [i,j] be the average predicted energy range value under the same blackbody temperature, [i,j] be the pixel coordinates, W be the number of pixels in the detector width, H be the number of pixels in the detector height, and M be the number of calibration points at a given blackbody temperature. The average predicted value is expressed as shown in equation (2):

[0063] (2)

[0064] in, For pixel grayscale values, For integration time, For the overall system gain, For system detector bias, For stray radiation from optical systems, Let be the radiation amount of the k-th attenuator. Let be the transmittance of the k-th attenuator.

[0065] The average value of the predicted energy range obtained at the same blackbody temperature is normalized and used as the expected value of the regression neural network.

[0066] Step 3: Construct an unsupervised learning network based on clustering and perform regression training, including:

[0067] Step 3.1: Build a feedforward neural network for regression and set various parameters of the neural network, including: input layer, regression layer, output layer, loss function, backpropagation function, activation function, learning rate, number of iterations, etc.

[0068] Step 3.2: The preprocessed grayscale response data, detector integration time, and attenuator information obtained in Step 2 are used as inputs to the regression neural network. To facilitate network training and accelerate training speed, the inputs are normalized according to equation (3):

[0069] (3)

[0070] in, Let be the grayscale response value of the pixel at coordinates [i, j]. and for The minimum and maximum values, This is the normalized grayscale response value. For the detector integration time, and for The minimum and maximum values, For normalized integral time, The input selection signal for the k-th attenuator is: the current attenuator is 1, and the rest are zero; X ​​is the input data matrix of the regression neural network.

[0071] Step 3.3: Use the average value of the predicted energy range at the same blackbody temperature obtained in Step 2 as the expected value of the regression neural network;

[0072] Step 3.4: Establish the regression equation and its backpropagation function, including:

[0073] Establish a mathematical regression equation according to equation (2);

[0074] Establish the relative mean square error loss function and its partial derivative according to equation (5):

[0075] (5)

[0076] in, The value of the loss function. The partial derivative of the loss function, Let be the predicted regression value corresponding to the nth calibrated blackbody image at coordinate [i,j]. is the average value of the predicted energy range at the same blackbody temperature, N is the number of radiation calibration points, W is the number of pixels in the detector width, and H is the number of pixels in the detector height.

[0077] The error backpropagation function is the partial derivative of the regression equation with respect to each regression parameter, as shown in equation (6):

[0078] (6)

[0079] in, Let [i,j] be the overall system gain at pixel [i,j]. Input a selection signal for the k-th attenuator. Let [i,j] be the gray value of the pixel at coordinates [i,j]. Let k be the transmittance of the attenuator. For integration time, The system detector offset is set at pixel coordinates [i,j]. The partial derivative of the mean square error loss function, Let W be the stray radiation from the optical system at pixel [i,j], W be the number of pixels in the detector width, and H be the number of pixels in the detector height. Let be the radiation amount of the k-th attenuator. yes right The partial derivative, yes right The partial derivative, yes right The partial derivative, It is the kth attenuator. right The partial derivative, It is the kth attenuator. right The partial derivative of the parameter. The change in each parameter in each iteration is shown in equation (7):

[0080] (7)

[0081] in, yes right The partial derivative, yes right The partial derivative, yes right The partial derivative, It is the kth attenuator. right The partial derivative, It is the kth attenuator. right The partial differential of . , , , , These represent the learning rates for each coefficient.

[0082] Step 3.5: Set the non-linear activation function, including:

[0083] The role of the nonlinear activation function is to introduce nonlinear factors and improve the regression ability of the network. Since equation (2) already has a certain nonlinear regression ability, the nonlinear activation function is not necessary. Whether to use it can be determined based on the detector response characteristics and the regression effect of equation (2). In order to maintain the continuity of the regression network and avoid singularities, the activation function should be a monotonic and continuous nonlinear function. Nonlinear activation functions include, but are not limited to, equations (8) and (9):

[0084] (8)

[0085] (9)

[0086] Where sig represents the output signal and x represents the input signal.

[0087] Step 3.6: Train the network to obtain network parameters that meet the accuracy requirements, including:

[0088] During training, data of the entire image size is input together, and the parameters of each pixel are learned in parallel during training, shortening the training time. Training stops once the expected accuracy is achieved. The hardware platform for training the network includes, but is not limited to, a central processing unit (CPU), a graphics processing unit (GPU), an advanced reduced instruction set computer architecture (ARM), and an application-specific integrated circuit (ASIC).

[0089] Step 4: Deploy network parameters and capture images of the actual scene, including:

[0090] The trained parameters are deployed into the non-uniformity correction network, and the normalized original gray response, integration time and attenuation plate information are used as inputs. Equation (10) is used to infer and calculate the average predicted value of the pixel energy domain.

[0091] (10)

[0092] in, This is a predicted value for the energy domain. For a well-trained regression network, It is about equation (2). According to the regression network Perform reasoning and calculation;

[0093] Example:

[0094] An embodiment of the present invention provides a non-uniformity correction method based on unsupervised machine learning, comprising:

[0095] Step 1: Calibrate the system at multiple attenuator, integration time, and blackbody temperature points to obtain multiple frames of raw data showing the change in image grayscale response with attenuator, integration time, and blackbody temperature.

[0096] Step 2: Perform average denoising and blind pixel replacement preprocessing on the multi-frame raw data of the calibration working point.

[0097] Step 3, as follows Figure 2 , Figure 3 As shown, an unsupervised learning network based on clustering is constructed and regression training is performed:

[0098] Step 3.1: Build the feedforward neural network for regression. The initial settings for the parameter learning network are as follows:

[0099]

[0100] in, The number of pixels is the width of the detector. The number of pixels at the detector height. For the dimension of the input matrix, For network learning rate, For momentum parameters, This specifies the dimension of the output matrix.

[0101] Step 3.2, the preprocessed grayscale response data, detector integration time and attenuator information obtained in step 2 are used as input to the regression neural network. In order to facilitate network training and speed up training, the input is normalized according to equation (3).

[0102] The non-uniformity correction image display effect of this embodiment is as follows:

[0103] (1) The corrected images and their grayscale histograms of the two-point correction algorithm at different integration times are as follows: Figure 4 As shown, the corrected images and their histograms of the algorithm proposed in this invention at different integration times are as follows: Figure 5 As shown, the results indicate that, compared to the two-point correction method, the corrected image obtained by this method is more stable when the integration time is changed. Specifically, Figure 4 (a) shows the image corrected by the two-point correction algorithm when the integration time is 1920µs. Figure 4(b) shows the image corrected by the two-point correction algorithm when the integration time is 2240µs. Figure 4 (c) represents the image corrected by the two-point correction algorithm when the integration time is 2560µs. Figure 4 (d) is the histogram of the image corrected by the two-point correction algorithm when the integration time is 1920us. Figure 4 (e) is the histogram of the image corrected by the two-point correction algorithm when the integration time is 2240us. Figure 4 (f) is the histogram of the image corrected by the two-point correction algorithm when the integration time is 2560us; Figure 5 (a) shows the image corrected by the algorithm proposed in this invention when the integration time is 1920µs. Figure 5 (b) shows the image corrected by the algorithm proposed in this invention when the integration time is 2240µs. Figure 5 (c) represents the image correction algorithm proposed in this invention when the integration time is 2560µs. Figure 5 (d) represents the histogram of the image corrected by the algorithm proposed in this invention when the integration time is 1920us. Figure 5 (e) represents the histogram of the image corrected by the algorithm proposed in this invention when the integration time is 2240µs. Figure 5 (f) is the histogram corresponding to the image corrected by the algorithm proposed in this invention when the integration time is 2560us.

[0104] (2) Corrected images and their grayscale histograms of the two-point correction algorithm under different attenuation filter transmittances are shown below. Figure 6 As shown, the algorithm proposed in this invention provides the corrected images and their histograms under different attenuation filter transmittances. Figure 7 As shown, the results indicate that, compared to the two-point correction method, the corrected image obtained by this method is more stable when the transmittance of the attenuator is changed. Specifically, Figure 6 (a) shows the image corrected by the two-point correction algorithm when the transmittance of the attenuator is 1. Figure 6 (b) shows the image corrected by the two-point correction algorithm when the attenuator transmittance is 0.88. Figure 6 (c) is the histogram of the image corrected by the two-point correction algorithm when the transmittance of the attenuator is 1. Figure 6 (d) is the histogram of the image corrected by the two-point correction algorithm when the transmittance of the attenuator is 0.88; Figure 7 (a) represents the image corrected by the algorithm proposed in this invention when the transmittance of the attenuator is 1. Figure 7 (b) shows the image corrected by the algorithm proposed in this invention when the transmittance of the attenuator is 0.88. Figure 7 (c) represents the histogram corresponding to the image corrected by the algorithm proposed in this invention when the transmittance of the attenuator is 1. Figure 7 (d) is the histogram corresponding to the image corrected by the algorithm proposed in this invention when the transmittance of the attenuator is 0.88.

[0105] (3) The images and their corresponding histograms after correction using the proposed method at the calibrated and uncalibrated operating points are as follows: Figure 8 As shown, 1920µs is the calibration point, and 1472µs and 2368µs are the uncalibrated points. The results demonstrate that the method proposed in this invention can maintain good correction and display effects at both calibrated and uncalibrated operating points. Figure 8 (a) shows the image corrected by the algorithm proposed in this invention when the integration time is 1472 μs. Figure 8 (b) shows the image corrected by the algorithm proposed in this invention when the integration time is 1920µs. Figure 8 (c) represents the image corrected by the algorithm proposed in this invention when the integration time is 2368µs. Figure 8 (d) represents the histogram corresponding to the image corrected by the algorithm proposed in this invention when the integration time is 1472us. Figure 8 (e) represents the histogram corresponding to the image corrected by the algorithm proposed in this invention when the integration time is 1920µs. Figure 8 (f) represents the histogram corresponding to the image corrected by the algorithm proposed in this invention when the integration time is 2368us.

[0106] Numerous specific details are set forth in the specification provided herein. However, it will be understood that embodiments of the invention may be practiced without these specific details. In some instances, well-known methods, structures, and techniques have not been shown in detail so as not to obscure the understanding of this specification.

[0107] Although the invention has been described with respect to a limited number of embodiments, those skilled in the art will understand from the foregoing description that other embodiments are conceivable within the scope of the invention described herein. Furthermore, it should be noted that the language used in this specification has been chosen primarily for readability and instructional purposes, and not for the purpose of explaining or limiting the subject matter of the invention.

Claims

1. A non-uniformity correction method based on unsupervised machine learning, characterized in that, Includes the following steps: Step 1: Calibrate the system operating point at different blackbody temperatures; Step 2: Establish a unified calibration database; Step 3: Construct an unsupervised learning network based on clustering and perform regression training, including: Establish a feedforward neural network for regression and set various network parameters, including: input layer, regression layer, output layer, loss function, backpropagation function, activation function, learning rate, and number of iterations; The preprocessed normalized grayscale response data, normalized detector integration time, and attenuator enable signal obtained in step 2 are used as inputs to the regression neural network; the average value of the predicted energy domain at the same blackbody temperature is used as the expected value of the regression neural network. The average value of the predicted energy range at the same blackbody temperature is shown in equation (1): （1） in, , where is the average value of the predicted energy domain at the same blackbody temperature, [i,j] is the pixel coordinate, W is the number of pixels in the detector width, H is the number of pixels in the detector height, and M is the number of calibration points at a certain blackbody temperature; Establish a unified mathematical regression equation (2): （2） in, For the normalized average predicted value term, To normalize the pixel grayscale value, For normalized integral time, The regression coefficients represent the overall system gain. The regression coefficients for the system detector bias term. The stray radiation bias regression coefficient of the optical system. Let be the regression coefficient of the radiation bias of the k-th attenuator. The regression coefficient for the transmittance term of the k-th attenuator is given. Step 4: Deploy the trained network parameters and perform energy domain non-uniformity correction on the real-world scene images.

2. The non-uniformity correction method based on unsupervised machine learning as described in claim 1, characterized in that, Step 1 includes: The system is calibrated at multiple attenuator, integration time, and blackbody temperature points to obtain multiple frames of raw data showing the change in image grayscale response with attenuator, integration time, and blackbody temperature.

3. The non-uniformity correction method based on unsupervised machine learning as described in claim 2, characterized in that, Step 2 includes: The multi-frame raw data from step 1 is preprocessed, including multi-frame denoising, averaging or replacing pixels around blind pixels.

4. The non-uniformity correction method based on unsupervised machine learning as described in claim 1, characterized in that, Based on equation (5), establish the relative mean square error loss function and its partial derivative: （5） in, The value of the loss function. Let be the relative error backpropagation function of the loss function, and let be the partial derivative of the loss function with respect to the radiance of the pixel at coordinate [i,j]. Let be the predicted regression value corresponding to the nth calibrated blackbody image at coordinate [i,j]. The average value of the predicted energy domain at the same blackbody temperature is given by N, where N is the number of radiation calibration points, W is the number of pixels in the detector width, and H is the number of pixels in the detector height. The reciprocals of the coefficients in the denominator of equation (2) are considered as a single term. The error backpropagation function is shown in equation (6). （6） in, Let [i,j] be the overall system gain at pixel [i,j]. Input a selection signal for the k-th attenuator. Let [i,j] be the gray value of the pixel at coordinates [i,j]. Let k be the transmittance of the attenuator. For integration time, The system detector offset is set at pixel coordinates [i,j]. The partial derivative of the mean square error loss function, Let W be the stray radiation of the optical system at pixel [i,j], W be the number of pixels in the detector width, and H be the number of pixels in the detector height. Let be the radiation amount of the k-th attenuator. yes right The partial derivative, yes right The partial derivative, yes right The partial derivative, It is the kth attenuator. right The partial derivative, It is the kth attenuator. right The partial derivative; The iterative changes of each coefficient are shown in equation (7): （7） in, For the overall system gain, Let k be the transmittance of the attenuator. For system detector bias, For stray radiation from optical systems, Let be the radiation amount of the k-th attenuator. yes right The partial derivative, yes right The partial derivative, yes right The partial derivative, It is the kth attenuator. right The partial derivative, It is the kth attenuator. right The partial derivative, , , , , These represent the learning rates for each coefficient.

5. The non-uniformity correction method based on unsupervised machine learning as described in claim 4, characterized in that, Training methods include: Import the calibration data of the first attenuator from the training library into the learning network model, and iterate only the learning coefficients. , , When the value of the loss function changes slowly, i.e., less than 10... -4 The second decaying data is added to the learning network model; simultaneously, the coefficients are... , , The learning rate decreased to 10 -6 Only iterate through the learning coefficients , , , When the value of the loss function changes slowly, the coefficient of recovery... , , The learning rate is determined; finally, the network is trained to obtain globally optimized, non-uniform correction network parameters that meet accuracy requirements.

6. The non-uniformity correction method based on unsupervised machine learning as described in claim 1, characterized in that, The hardware platform for training networks includes a central processing unit, a graphics processing unit, an advanced reduced instruction set computer architecture, or an application-specific integrated circuit.

7. The non-uniformity correction method based on unsupervised machine learning as described in claim 1, characterized in that, Step 4 includes: The trained parameters are deployed into a non-uniformity correction network, and real infrared scene images are captured under different attenuator settings and integration times. The normalized original grayscale response, integration time, and attenuator information are used as network inputs to perform inference calculations on the average predicted value of the pixel energy domain.

Images