A method for inverting spatial target parameters based on multi-angle sequence images

Abstract

The present invention discloses a method for inverse-inverting spatial target parameters based on multi-angle sequence images, including: establishing an infrared radiation characteristic model of a spatial target with any shape based on the facet method; simulating and generating a multi-view spatial target radiation image sequence according to the target motion model and the multi-platform detection model; based on the internal correlation between the time-series images of the spatial target under multiple views and parameters such as the target attitude, shape, size, and motion, generating big data samples through a high-precision simulation model, establishing a deep learning network model for training, and establishing an inverse-inversion model for spatial target parameters.

Description

Technical Field

[0001] The present invention relates to target characteristic analysis and feature extraction, deep learning networks, and image processing technologies, and particularly relates to a method for inverting spatial target parameters based on multi-angle sequence images. Background Art

[0002] Information such as shape, attitude, and size can reflect the structural characteristics and functions of spatial targets and are important features for target recognition. Infrared detectors have the advantages of high precision, small volume, and the ability to work day and night, and have been widely used on space-based detection platforms. However, due to the large observation distance and small target size, the target appears as a point target in the image, making it difficult to obtain information such as the shape and texture of the target. However, the multi-angle radiation intensity sequence of spatial targets provides an important way to extract information such as target attitude, shape, size, and motion. Conducting the inversion of parameters such as attitude, shape, and motion based on the infrared radiation intensity sequence is of great significance for improving the recognition ability of small targets at long distances.

[0003] At present, for the inversion of space target parameters, a lot of theoretical and practical research has been carried out at home and abroad, and shape inversion methods based on the characteristics of photometric curves, shape inversion methods based on Gaussian surface density, and shape inversion methods based on non-linear filtering technology have been formed. However, these methods mainly perform inversion on photometric data and have a high computational cost. For example, in the patent "Method for Identifying Characteristics of High-Orbit Small-Size Space Targets Based on Photometric Curves, Liang Yongqi, Shan Bin (ZL 201810167188.7, publication date of application: August 17, 2018)", through the analysis of the dynamic characteristics of high-orbit targets and the characteristics of photometric observation data, based on the target attitude kinematic model, a multi-model fusion algorithm is used to identify the characteristics of the target shape, size, material parameters, etc. This method cannot invert the high-dimensional characteristic parameters of space-based multi-platform approaching moving targets and perform target recognition. In the patent "Method for Obtaining the Scale of Space Objects Based on Photoelectric Observation, Wang Jianfeng (ZL 201410350320.X, publication date of application: October 15, 2014)", it is proposed to use observations to obtain high-precision photometric information of space objects, perform photometric correction processing, and calculate the optical scattering cross-section of space objects. This method does not correct the influence of factors such as target attitude and material, and cannot obtain the specific shape and size parameters of the target. In the patent "Inverting the Attitude Angle and Geometric Dimensions of a Dynamic Conical Target Based on Laser One-Dimensional Range Images, Mou Yuan, Wu Zhensen, Cao Yunhua, Li Yanhui (ZL 201410188024.4, publication date of application: July 30, 2014)", a method for using a radar array laser one-dimensional range image to identify the attitude angle and geometric dimensions of a dynamic conical target is proposed. This method is only applicable to the inversion of target parameters and attitude angles such as cones with analytical formulas. In the patent "Method for Detecting and Tracking Space Targets in Space-Based Optical Sequential Images, Ouyang Yan et al. (ZL 201610583483.1, publication date of application: January 4, 2017)", based on the continuous N-frame star maps obtained by space-based, a triangular matching method is used to identify reference stars. Taking the stars as a reference, the motion information of space targets is obtained, and the space targets are tracked in combination with Kalman filtering. This method can obtain the orbital motion information of space targets, but cannot obtain micro-motion information such as spin. Summary of the Invention

[0004] The object of the present invention is to provide a method for inverting space target parameters based on multi-angle sequential images, which improves the inversion efficiency while ensuring the inversion accuracy.

[0005] To achieve the above object, the present invention is realized through the following technical solutions:

[0006] A method for inverting space target parameters based on multi-angle sequential images is characterized by including the following steps:

[0007] S1. Approximate the shape and surface characteristic distribution of the space target with multiple small plane elements, establish the nodal balance equation, solve the surface temperature distribution of the target, and calculate the infrared radiation intensity of the space target with any complex shape using the reverse ray tracing method based on Planck's law;

[0008] S2. Establish a multi-view infrared radiation intensity sequence model of the space target according to the target motion model and the multi-platform detection model;

[0009] S3. Select typical space targets, traverse the target parameters to establish a training data set, and use the confidence network of deep learning to extract the target ontology features and cluster them;

[0010] S4. Through a multi-layer auto-encoding feature correlation network, establish a high-order, multi-dimensional, and non-linear mapping relationship network between the target parameters and the multi-angle radiation intensity, and generate the target parameters by inputting the radiation intensity sequence.

[0011] In step S1, any target is discretized into small plane elements, the nodal heat balance equation is established, the backward difference algorithm is used to solve the surface temperature, and the reverse ray tracing method is used to solve the infrared radiation intensity of the target with any complex shape.

[0012] In step S2, based on the target orbital motion and spin composite motion model and the space-based multi-detection platform orbital motion model, a multi-view infrared radiation intensity sequence model of the space target is established.

[0013] In step S3, the typical space targets include one or more of a cone, a frustum of a cone, a sphere, or rocket debris.

[0014] In step S4, the high-dimensional feature parameters of the target are used to label the target infrared sequence images corresponding to the scenes, and a two-dimensional feature map is generated based on the attention model using the infrared sequence images; based on the reversible deep auto-encoding network, a bidirectional mapping relationship is constructed using the high-dimensional feature parameters and the two-dimensional feature map; finally, the radiation intensity image sequence is input, the two-dimensional feature map is generated through the trained attention model, and then the high-dimensional feature parameters are generated from the two-dimensional feature map through the trained reversible deep auto-encoding network.

[0015] Compared with the prior art, the present invention has the following advantages:

[0016] Based on the internal correlation between the infrared radiation time series images of space targets from a space-based multi-perspective and parameters such as the target's attitude, shape, size, and motion, a reverse inversion model of space target parameters for a deep learning network is established through a data-driven method. The present invention introduces a multi-layer auto-encoding feature correlation network and an attention model to establish a high-order, multi-dimensional, non-linear mapping relationship network and an inversion model between target parameters and infrared radiation intensities from multiple angles, which can invert the parameters of targets with complex geometric shapes such as non-convex bodies and complex material targets. Compared with traditional inversion methods, the present invention has a wider application range, higher inversion accuracy, and higher efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a flowchart of a method for inverting space target parameters based on multi-angle sequence images according to the present invention;

[0018] Figure 2 is a diagram of an inversion model of target parameters for a data-driven deep network. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0019] The present invention will be further described below in conjunction with the drawings by elaborating on a preferred specific embodiment in detail.

[0020] As Figure 1 , 2 shown, a method for inverting space target parameters based on multi-angle sequence images includes the following steps:

[0021] S1. Approximate the shape and surface characteristic distribution of the space target with multiple small plane elements (the characteristics such as temperature, emissivity, and reflectivity are invariant at each plane), establish a nodal balance equation, solve the temperature distribution on the target surface, and calculate the infrared radiation intensity of any complex-shaped space target using the reverse ray tracing method based on Planck's law;

[0022] S2. Establish a multi-perspective space target infrared radiation intensity sequence model according to the target motion model and the space-based multi-platform detection model;

[0023] S3. Select typical space targets such as cones, frustums of cones, spheres, and rocket debris, traverse parameters such as the target's attitude, size, and motion to establish a training data set, and use a deep learning confidence network to extract the target body features and cluster them;

[0024] S4. Through a multi-layer auto-encoding feature correlation network and an attention model, establish a high-order, multi-dimensional, non-linear mapping relationship network between target parameters and multi-angle radiation intensities, and generate parameters such as target type, attitude, size, and motion by inputting the radiation intensity sequence.

[0025] Specifically, the above step S1 includes:

[0026] Discretize any target into small plane elements (with invariant characteristics such as temperature and emissivity across the plane), establish the nodal heat balance equation, solve the surface temperature using the backward difference algorithm, and solve the infrared radiation intensity of any target with a complex shape using the inverse ray tracing method.

[0027] The above step S2 is based on the target orbital motion and spin composite motion model and the space-based multi-detection platform orbital motion model to establish a multi-view spatial target infrared radiation intensity sequence model.

[0028] In the above step S3, the target feature clustering model based on the deep belief network is divided into a pre-training part and a fine-tuning part. The pre-training part consists of an unsupervised deep learning structure layer and a pre-clustering layer. First, input the target feature parameters. The input data obtains effective and simple deep features of the target through unsupervised deep learning. Train the stacked restricted Boltzmann machine (DRBM) by the gradient descent method. The deep feature representation of the target can be obtained, and the initial weights of the network can be obtained. Then, the k-means algorithm of the pre-clustering layer performs preliminary clustering on the deep features obtained previously and generates preliminary clustering centers. The main function of this part is to perform feature learning on the target data. Through the mining of deep features, the self-organized class structure inside the target data is fully exposed. The deep structure can well avoid falling into the trap of local optimal solutions through unsupervised pre-training.

[0029] The fine-tuning part also consists of a multi-layer autoencoder network and a clustering layer. However, different from the pre-training module, fine-tuning further performs cross-iteration on the results of pre-training to obtain the final result. The fine-tuning part inputs the background environment elements and the target ontology feature parameters, and then uses the initial weights and initial clustering centers obtained in the pre-training part for deep fine-tuning. Finally, use the k-means algorithm to obtain the final clustering result. The objective function of the fine-tuning part is to maximize the likelihood probability function and minimize the within-class distance. The stopping condition is to reach the preset number of iterations or when the objective function value is less than a certain threshold. Through cross-iteration of the objective function, optimize the depth structure of the distribution and the clustering result of the target ontology characteristics to obtain the target ontology clustering result, extract the clustering centers of each type of target ontology and the range of each feature parameter.

[0030] In the above step S4, the high-dimensional feature parameters of the target form a matching data set with the target infrared sequence images, that is, the target infrared sequence images of the corresponding scene are labeled with the high-dimensional feature parameters of the target; based on the attention model, the infrared sequence images are used to generate two-dimensional feature maps. Since there is a one-to-one correspondence between the infrared sequence images and the two-dimensional feature maps, a matching data is formed between the generated two-dimensional feature maps and the labeled high-dimensional feature parameters; based on the reversible deep autoencoder network, the high-dimensional feature parameters and the two-dimensional feature maps are used to construct a bijective mapping relationship, that is, the high-dimensional feature parameters are generated through the two-dimensional feature maps, and the two-dimensional feature maps can also be generated through the high-dimensional feature parameters; finally, the radiation intensity image sequence is input, the two-dimensional feature maps are generated through the trained attention model, and then the high-dimensional feature parameters are generated from the two-dimensional feature maps through the trained reversible deep autoencoder network.

[0031] In summary, the spatial target parameter inversion method based on multi-angle sequence images of the present invention improves the inversion efficiency while ensuring the inversion accuracy.

[0032] Although the content of the present invention has been described in detail through the above preferred embodiments, it should be understood that the above description should not be construed as a limitation of the present invention. After those skilled in the art have read the above content, various modifications and substitutions to the present invention will be obvious. Therefore, the protection scope of the present invention should be defined by the appended claims.

Claims

1. A method for inverting spatial target parameters based on multi-angle sequence images, characterized in that, It includes the following steps: S1. Approximate the shape and surface characteristic distribution of the space target with multiple small plane elements, establish the nodal balance equation, solve the surface temperature distribution of the target, and calculate the infrared radiation intensity of the space target with any complex shape by using the backward ray tracing method based on Planck's law; S2. Establish a multi-view infrared radiation intensity sequence model of the space target according to the target motion model and the multi-platform detection model; S3. Select typical space targets, traverse the target parameters to establish a training data set, and use the confidence network of deep learning to extract the target body features and cluster them; S4. Through the multi-layer auto-encoding feature correlation network, establish a high-order, multi-dimensional, and non-linear mapping relationship network between the target parameters and the multi-angle radiation intensity, and generate the target parameters by inputting the radiation intensity sequence.

2. The method for inverting spatial target parameters based on multi-angle sequence images according to claim 1, wherein In step S1, any target is discretized into small plane elements, the nodal heat balance equation is established, the backward difference algorithm is used to solve the surface temperature, and the backward ray tracing method is used to solve the infrared radiation intensity of the target with any complex shape.

3. The method for inverting spatial target parameters based on multi-angle sequence images according to claim 1, wherein In step S2, based on the target orbital motion and spin composite motion model and the space-based multi-detection platform orbital motion model, a multi-view infrared radiation intensity sequence model of the space target is established.

4. The spatial target parameter inversion method based on multi-angle sequence images according to claim 1, characterized in that, In step S3, the typical space targets include one or more of a cone, a frustum of a cone, a sphere, or rocket debris.

5. The spatial target parameter inversion method based on multi-angle sequence images according to claim 1, wherein In step S4, the high-dimensional feature parameters of the target are used to label the target infrared sequence images corresponding to the scenes, and a two-dimensional feature map is generated based on the attention model using the infrared sequence images; based on the reversible deep auto-encoding network, a bidirectional mapping relationship is constructed using the high-dimensional feature parameters and the two-dimensional feature map; finally, the radiation intensity image sequence is input, the two-dimensional feature map is generated by the trained attention model, and then the high-dimensional feature parameters are generated from the two-dimensional feature map by the trained reversible deep auto-encoding network.

Images