A method for end-to-end comprehensive inversion of target infrared intrinsic characteristics

Abstract

This invention discloses a comprehensive inversion method for the entire link of target infrared intrinsic characteristics. Starting from the full-link radiation coupling imaging characteristics, a radiation coupling model of the target imaging link is constructed. Based on the coupling model and related infrared BRDF parameters, environmental radiation, atmospheric transmittance, equipment noise, and imaging effects, the comprehensive inversion of the target's intrinsic infrared characteristics is achieved, improving the accuracy of test data and providing accurate target and environmental infrared radiation data and coupling relationship data for model verification and simulation validation, thereby enhancing the credibility of the entire virtual simulation experiment.

Description

Technical Field

[0001] This invention belongs to the field of infrared imaging data inversion technology, specifically relating to a comprehensive inversion method for the entire link of target infrared intrinsic characteristics. Background Technology

[0002] Target field testing is generally conducted under highly dynamic conditions and at long measurement distances. Due to motion blur and optical dispersion, the effective data obtained from sampling is limited and the data is blurry, especially for point targets and small targets where dispersion is severe. Because the target's energy is dispersed, infrared imaging cannot reflect the true size of the target, and the edges are close to the background radiation, making it difficult to extract specific locations. Traditional brightness-temperature inversion models invert the target's brightness using grayscale output from infrared equipment, and then calculate the temperature using the target's emissivity based on radiation theory. This requires grayscale pixels to accurately reflect the radiation characteristics of different parts of the target; that is, grayscale pixels must represent both the target's geometric characteristics and its radiation characteristics. Correct inversion can only be carried out when the two satisfy the correspondence of geometric projection. When infrared imaging cannot reflect the true projection relationship of the target, the grayscale pixels after target extraction cannot correctly correspond to the target parts. When dispersion and motion blur occur, the inverted values ​​will also differ significantly from the true values.

[0003] The intrinsic infrared signature of a target is the final measurement value required for the field infrared signature test of weapon equipment. Only by obtaining accurate intrinsic infrared signatures of the target can accurate values ​​be given for model verification. However, the values ​​sampled during the test include the intrinsic radiation of the target, reflections caused by environmental radiation, atmospheric path path radiation, and optical dispersion, and are affected by atmospheric absorption. The coupling characteristics are very complex. At the same time, changes in the target's operating conditions during the sampling process lead to changes in attitude and tail jet radiation, resulting in inconsistent measurement values ​​under different conditions. Therefore, the inversion of target infrared signatures needs to consider the influence of the entire link on target radiation and decouple according to the transmission process in order to obtain accurate target infrared signature values.

[0004] However, current target feature inversion only focuses on atmospheric absorption and path radiation, without conducting targeted inversion based on the whole-link sampling mechanism. In particular, there is a lack of detailed research on the characteristics of targets to environmental radiation reflection and optical dispersion. The modeling accuracy of their reflection and dispersion models is low, and the inversion error is large. There is also a lack of corresponding research on the inversion of non-cooperative targets. Moreover, non-cooperative targets are foreign military equipment, which are urgently needed models for virtual system modeling and simulation experiments. Their accuracy directly affects the credibility of simulation experiments.

[0005] Therefore, only by improving the accuracy of target infrared feature testing equipment, infrared physical parameter calibration during field testing, real-time measurement of environmental factors, and high-precision extraction and inversion technology of target data can we comprehensively enhance the ability to acquire target infrared feature data. This will solve the problems of low accuracy of infrared feature data of current weapons and equipment, difficulty in model verification and simulation verification, improve the accuracy of typical targets and complex combat environment models of major combat targets, enhance the credibility verification level of infrared scene virtual simulation test, and provide technical support for the construction of virtual test environment for weapon systems and integrated comprehensive test verification technology. Summary of the Invention

[0006] To address the aforementioned problems, this invention provides a method for the comprehensive inversion of the intrinsic infrared characteristics of a target across the entire link. Starting from the full-link radiation-coupled imaging characteristics, a radiation-coupled model of the target imaging link is constructed. Based on the coupling model and relevant infrared BRDF parameters, environmental radiation, atmospheric transmittance, equipment noise, and imaging effects, the comprehensive inversion of the intrinsic infrared characteristics of the target is achieved.

[0007] The core idea of ​​this invention is:

[0008] Starting from the end-to-end infrared radiation transmission characteristics, the method decouples layer by layer to form analytical equations, giving the intrinsic infrared characteristics of the target. Based on a theoretical model, the method of this invention estimates and corrects the data, providing accurate infrared characteristics of various factors in the scene as inversion characteristics, rather than isolated radiation feature values. This gives the relationship between test values, target infrared characteristics, and environmental radiation at the system level, providing complete verification test values ​​for high-accuracy modeling and meeting the needs of infrared modeling and simulation of targets and scenes.

[0009] The technical solution to achieve the purpose of this invention is as follows:

[0010] A method for the full-link comprehensive inversion of the intrinsic infrared characteristics of a target, characterized by comprising the following steps:

[0011] Step 1: Obtain the infrared measurement image of the target through an optical system;

[0012] Step 2: Obtain the target radiation region using an improved method for extracting infrared imaging data of the target under optical dispersion;

[0013] Step 3: Estimate the operating conditions of the target under test within the radiation area;

[0014] Step 4: Input the target radiation area obtained in Step 2 and the working condition data obtained in Step 3 into the target full-link infrared imaging model, perform target infrared feature comprehensive inversion, and finally obtain the target infrared measurement image features.

[0015] Furthermore, the specific operation steps of the target infrared imaging data extraction method described in step 2 are as follows:

[0016] Step 21: Calculate the average brightness of the surrounding background region of the target infrared measurement image, and use this average value as the threshold;

[0017] Step 22: Determine whether the brightness of the target infrared measurement image is greater than the threshold. If it is, obtain the diffuse spot pixel area as the target radiation area.

[0018] Furthermore, the specific steps for step 3 are as follows:

[0019] Step 31: Estimation of Cooperation Objectives

[0020] By collecting the aircraft's operating conditions in real time through flight data sensors, the target's pitch angle, roll angle, attitude value, position, and velocity can be obtained.

[0021] Step 32: Estimation of non-cooperative objectives

[0022] A three-dimensional attitude information recognition framework is established. Based on this framework, the pitch angle, roll angle, attitude value, position, and velocity of the target are estimated.

[0023] Furthermore, the specific steps of step 32 are as follows:

[0024] Step 321: Convert the infrared measurement images of the target under different attitude conditions into binary target images through image segmentation;

[0025] Step 322: Extract the target contour from the obtained binarized target image, use the pose image database as a reference sample, and obtain the azimuth angle φ based on the image shape analysis method and the angle distance curve amplitude phase inference method;

[0026] Step 323: Extract RTS feature invariants from the binarized target image;

[0027] Step 324: Normalize the obtained RTS feature invariants;

[0028] Step 325: Construct a training dataset from the feature invariants obtained under different pose conditions, and train the neural network;

[0029] Step 326: Input the infrared measurement image of the target to be measured into the trained neural network to obtain the pitch angle θ and roll angle γ;

[0030] Step 327: Use a neural network to obtain the pose probability value of the current frame of the infrared measurement image of the target to be tested, sort the probability values, and take the first four as possible pose values.

[0031] Step 328: Use the improved Hausdorff distance metric-based contour matching method to examine any two pose values ​​and perform filtering to obtain the pose estimate of the current frame.

[0032] Step 329: Construct a sequence image position information filtering model based on the attitude estimation value, and use the model to estimate the position and velocity of the current frame to finally obtain position and velocity information.

[0033] Furthermore, the specific steps of step 329 are as follows:

[0034] Step 3291: Denote the position signal as X(t), then the following condition is met:

[0035]

[0036] Where α and β represent the start and end points over a period of time, W(t) is a white normal process, and the mean and covariance of W(t) are:

[0037]

[0038] Step 3292: Let the mean of X(t) satisfy:

[0039]

[0040] Meanwhile, let the variance of X(t) satisfy:

[0041]

[0042] Therefore, the average speed can be obtained as:

[0043]

[0044] Furthermore, the specific operational steps of step 4 include:

[0045] Step 41: Input the obtained pitch angle, roll angle, attitude value, position and velocity into the target full-link infrared imaging model. This imaging model is the transmission link model of target → background → interference → atmosphere → detection system.

[0046] Step 42: Based on the target's full-link infrared imaging model, establish the full-link imaging equation set and solve the equation set;

[0047] Step 43: Obtain the infrared radiation brightness value of the corresponding target part for each sampled data.

[0048] Preferably, the neural network mentioned in step 326 is any one of a BP neural network, a GRNN network classifier, or a v-SVM regression machine.

[0049] Preferably, the feature invariants obtained in step 323 include Hu moments, Fourier descriptors, and DCT descriptors.

[0050] Compared with existing technologies, this method has the following advantages:

[0051] This invention addresses the testing of infrared features of typical targets, the calibration of infrared physical parameters during field testing, real-time measurement of environmental factors, and high-precision extraction and inversion of target data. It establishes a standardized method for testing the infrared features of weapon systems in the field, aiming to comprehensively improve the ability to acquire target infrared feature data. This solves the problems of low accuracy of infrared feature data, difficulties in model verification and simulation validation applications in current weapon systems, improves the accuracy of models of typical targets and complex combat environments, enhances the credibility verification level of virtual simulation experiments in infrared scenarios, and provides technical support for the construction of virtual test environments for weapon systems and integrated comprehensive test verification technology.

[0052] This invention, based on the coupling model and combined with relevant infrared BRDF parameters, environmental radiation, atmospheric transmittance, equipment noise, and imaging effects, achieves a comprehensive inversion of the intrinsic infrared characteristics of the target. This effectively improves the accuracy of test data and provides accurate target and environmental infrared radiation data and coupling relationship data for the verification and simulation of traditional brightness temperature inversion models, thereby enhancing the credibility of the entire virtual simulation experiment. Attached Figure Description

[0053] Figure 1 A flowchart for attitude estimation of non-cooperative aerial targets;

[0054] Figure 2 This is a schematic diagram of a pose image database;

[0055] Figure 3 A schematic diagram of an infrared imaging link model for aerial targets. Detailed Implementation

[0056] To enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions of the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0057] I. Extraction of Infrared Imaging Data of Targets under Optical Dispersion

[0058] In actual field tests, when the aerial target is relatively far from the optical system, the image formed by the optical system is generally a point target, that is, its area on the detector pixel is about a few pixels. At this time, due to the diffraction effect of the optical system, its energy is diffused. In this case, the application of traditional radiation characteristic extraction methods will affect the accuracy of radiation characteristic measurement. Therefore, this invention proposes a new method for extracting infrared imaging data of targets under optical diffusion based on the relationship between energy, geometric projection and the interrelationship of multi-band radiation characteristics.

[0059] First, according to the law of conservation of energy, the radiant flux through the aperture is the same as the radiant flux forming the diffuse spot. The diffuse spot is formed by the projection of the target onto the detector. Since the size of the diffuse spot is related to the target energy, the larger the energy, the larger the diffuse spot. Therefore, this invention considers the size of the diffuse spot in different wavelength bands. Due to the large radiant energy of the skin, the diffuse spot in the lateral and head-on long-wavelength areas should be the strongest. In the tail nozzle, the short-wavelength energy is strong. Therefore, the union of the largest area (the pixel area of ​​the diffuse spot in the infrared imaging image of the target) is selected as the target radiation area. That is, the average brightness of the background area around the largest area is selected as the threshold. If it is greater than the threshold, it is the pixel area of ​​the diffuse spot. Thus, the pixel diffuse spot area is obtained, which is the target area.

[0060] II. Operating Condition Estimation for Non-Cooperative Objectives

[0061] The target's operating conditions are closely related to its infrared characteristics and are important parameters for infrared modeling and simulation. For traditional range testing, the aircraft's operating conditions can be collected in real time through flight parameters. However, for non-cooperative targets such as foreign aircraft, their flight data cannot be publicly obtained. To obtain relevant parameters such as attitude and position, it is necessary to reverse-engineer them from the test data. Therefore, a three-dimensional attitude information recognition framework is proposed to identify the attitude and position of non-cooperative targets.

[0062] Target attitude recognition falls under the categories of computer vision and pattern recognition. Inferring aircraft target attitude from images acquired by infrared imaging detectors is a monocular vision problem, which is quite challenging. Computer vision shows that if the 3D structure of a target is known, its projected images at different attitude angles can be easily obtained through affine or projective transformations. However, the images obtained by affine or projective transformations have a strong nonlinear relationship with the target-missile distance and viewing angle; images obtained from different distances and viewing angles are different. Many geometric properties are not invariant; for example, the projections of two parallel lines on an image are generally not parallel, and the length of line segments is also related to the target-missile distance and azimuth, leading to sample data conflicts. This makes it very difficult to solve for the 3D attitude from a 2D image. For infrared images, grayscale reflects the radiated energy of various parts of the aircraft, but not the depth information of structural components. Only shape features and the radiation characteristics of special areas such as engine nozzles can be utilized, and the solutions often have multiple solutions.

[0063] To address these issues, this invention designs a 3D pose information recognition framework based on machine learning methods, resolving the problems of sample data conflict and multiple solutions. By analyzing the distribution characteristics of pose image features, the multi-class recognition problem of angle identification is transformed into a nonlinear function approximation problem. Feature vectors are constructed using translation, rotation, and scaling invariants, and a pose estimator is built using three networks for validation studies. The specific recognition framework structure is as follows: Figure 1 As shown, it includes the following steps:

[0064] (1) Infrared images under different posture conditions are obtained by measuring through an optical system. The real-time infrared images obtained based on the infrared images are converted into binary target images by image segmentation. At this time, the influence of the tail flame is ignored, and the focus is on using the target features of the region and edge for detection and posture estimation.

[0065] (2) Extract the target contour from the binarized target image, using methods such as... Figure 2 The attitude image database shown is used as a reference sample. Based on the image shape analysis method and the angle distance curve amplitude phase inference method disclosed in the prior art, the angle between the projection of the machine axis in the imaging plane and the x-axis of the imaging plane is obtained, that is, the azimuth angle φ.

[0066] (3) Extract RTS feature invariants from the binarized target image, including Hu moments, Fourier descriptors, and DCT descriptors.

[0067] (4) Data normalization (standardization). Before training and testing the network, the invariants are normalized to prevent neuron output saturation caused by excessively large net input absolute values.

[0068] (5) Construct a training dataset from the feature invariants obtained under different posture conditions, and train a BP neural network or GRNN network classifier or v-SVM regression machine.

[0069] (6) Input the normalized feature vector obtained from the test image into the classifier / regressor and use the network approximation ability to fit the angle between the aircraft's spatial attitude and the imaging plane, namely the pitch angle θ and the roll angle γ.

[0070] (7) The estimated pose is obtained by classifying the probability values, sorting the probability values, taking the first four as the corresponding four pose values, using the contour matching method based on the publicly available improved Hausdorff distance metric to check the possible two pose values, and using the correlation of pose changes in the sequence images to filter and check to obtain the pose estimate of the current frame.

[0071] A position information filtering model for sequential images is constructed based on the attitude estimation values. The solution for information filtering is based on the continuity of position and velocity changes. That is, by utilizing the continuity of position and velocity in the sequential images, the position and velocity of the current frame are estimated using the preceding frames. The construction steps of this sequential image position information filtering model include:

[0072] Step 1: Assume the attitude signal changes over time as a continuous normal Markov process. If the position signal is denoted as X(t), then the following holds:

[0073]

[0074] Where W(t) is a white normal process, its mean and covariance are:

[0075]

[0076] Step 2: Let the mean of X(t) satisfy:

[0077]

[0078] If we let the variance of X(t)

[0079]

[0080] Equations (1)-(2) are the established sequence image position information filtering model. Using this sequence image position information filtering model, a set of state equations for attitude and position parameters can be established, and the position and velocity information can be estimated using the Kalman filtering model.

[0081] The engine operating condition is most clearly manifested in the mid-wave and short-wave characteristics of infrared targets. By combining range and attitude estimation analysis of changes in mid- and short-wave imaging data, the engine operating condition can be estimated.

[0082] III. Target Infrared Feature Comprehensive Inversion Method Based on Target End-Link Infrared Imaging Model

[0083] During the infrared signature testing of aerial targets, the signal transmission and conversion process of the optoelectronic system is a complex dynamic process influenced by factors across the entire link. Its link model is a transmission link model of "target → background → interference → atmosphere → detection system," as follows: Figure 3 As shown.

[0084] Typically, the intrinsic radiation of a target in infrared imaging involves numerous nonlinear and linear physical effects coupled between the environment and various unit modules. Therefore, during testing, it is necessary to obtain the corresponding quantities by measuring the target's physical parameters and the background environmental radiation parameters. Then, from the perspective of truly reflecting the actual testing process of the system, based on the principle of ray tracing, the infrared full-link imaging model is combined with the test values ​​to construct a set of full-link imaging equations and solve for the infrared radiation brightness value of the corresponding target part of each sampled data.

[0085] Considering the existence of testing errors, this set of equations may be contradictory. By solving the contradictory equations using the least squares method, an approximate radiance value can be obtained.

[0086] Contents not described in detail in this specification are existing technologies known to those skilled in the art. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for end-to-end comprehensive inversion of the intrinsic infrared characteristics of a target, characterized in that, Includes the following steps: Step 1: Obtain the infrared measurement image of the target through an optical system; Step 2: Obtain the target radiation region using an improved method for extracting infrared imaging data of the target under optical dispersion; Step 3: Estimate the operating conditions of the target under test within the radiation area; Step 4: Input the target radiation area obtained in Step 2 and the working condition data obtained in Step 3 into the target full-link infrared imaging model, perform target infrared feature comprehensive inversion, and finally obtain the target infrared measurement image features. The specific steps in step 3 are as follows: Step 31: Estimation of Cooperation Objectives By collecting the aircraft's operating conditions in real time through flight data sensors, the target's pitch angle, roll angle, attitude value, position, and velocity can be obtained. Step 32: Estimation of non-cooperative objectives A three-dimensional attitude information recognition framework is established. Based on this framework, the pitch angle, roll angle, attitude value, position, and velocity of the target are estimated. The specific steps in step 32 are as follows: Step 321: Convert the infrared measurement images of the target under different attitude conditions into binary target images through image segmentation; Step 322: Extract the target contour from the obtained binarized target image, use the pose image database as a reference sample, and obtain the azimuth angle φ based on the image shape analysis method and the angle distance curve amplitude phase inference method; Step 323: Extract RTS feature invariants from the binarized target image; Step 324: Normalize the obtained RTS feature invariants; Step 325: Construct a training dataset from the feature invariants obtained under different pose conditions, and train the neural network; Step 326: Input the infrared measurement image of the target to be measured into the trained neural network to obtain the pitch angle θ and roll angle γ; Step 327: Use a neural network to obtain the pose probability value of the current frame of the infrared measurement image of the target to be tested, sort the probability values, and take the first four as possible pose values. Step 328: Use the improved Hausdorff distance metric-based contour matching method to examine any two pose values ​​and perform filtering to obtain the pose estimate of the current frame. Step 329: Construct a sequence image position information filtering model based on the attitude estimation value, and use the model to estimate the position and velocity of the current frame to finally obtain the position and velocity information; The specific steps of step 329 are as follows: Step 3291: Denote the position signal as X(t), then the following condition is met: （1） in, , Let W(t) represent the start and end points over a period of time, and let W(t) be a white normal process with mean and covariance as follows: ； Step 3292: Let the mean of X(t) satisfy: ； At the same time, let the variance of X(t) satisfy ; （2） Equations (1)-(2) are the established sequence image position information filtering model. Using this sequence image position information filtering model, the state equation set of attitude and position parameters can be established, and the position and velocity information can be estimated using the Kalman filtering model. The specific steps in step 4 include: Step 41: Input the obtained pitch angle, roll angle, attitude value, position, and velocity into the target end-to-end infrared imaging model. This imaging model is the transmission link model of target → background → interference → atmosphere → detection system; the transmission link model is simulated based on the ray tracing model. The ray tracing model is as follows: ； Step 42: Based on the target end-link infrared imaging model, establish the end-link imaging equation set and solve the equation set; Step 43: Obtain the infrared radiation brightness value of the corresponding target part for each sampled data.

2. The method for full-link comprehensive inversion of target infrared intrinsic characteristics according to claim 1, characterized in that, The specific operation steps of the target infrared imaging data extraction method described in step 2 are as follows: Step 21: Calculate the average brightness of the surrounding background region of the target infrared measurement image, and use this average value as the threshold; Step 22: Determine whether the brightness of the target infrared measurement image is greater than the threshold. If it is, obtain the diffuse spot pixel area as the target radiation area.

3. The method for full-link comprehensive inversion of target infrared intrinsic characteristics according to claim 1, characterized in that, The neural network mentioned in step 326 can be any one of a BP neural network, a GRNN network classifier, or a v-SVM regression machine.

4. The method for full-link comprehensive inversion of target infrared intrinsic characteristics according to claim 1, characterized in that, The feature invariants obtained in step 323 include Hu moments, Fourier descriptors, and DCT descriptors.

Images