Asteroid point cloud intelligent reconstruction method based on single-view radar image
By constructing an intelligent reconstruction network for asteroid point clouds using convolutional neural networks, the problems of numerous parameters, long time consumption, and reliance on manual adjustment in traditional methods are solved, enabling rapid and accurate reconstruction of the three-dimensional shape of asteroids.
Patent Information
- Application Number
- CN202511150906.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-18
- Publication Date
- 2025-12-09
AI Technical Summary
Existing technologies for reconstructing the three-dimensional shape of asteroids suffer from numerous parameters, long processing times, susceptibility to local optima, and reliance on manual adjustments, leading to shape inversion distortion and making rapid and accurate reconstruction difficult.
A convolutional neural network is used to construct an intelligent reconstruction network for asteroid point clouds. The encoder and generator modules learn shape features from the two-dimensional images of asteroids, and the network is trained using gradient descent. A loss function is defined to optimize the reconstruction process and achieve automated reconstruction.
It achieves rapid and accurate 3D shape reconstruction, reduces human intervention, and improves reconstruction efficiency and accuracy, which is superior to traditional methods.
Smart Images
Figure CN121095437A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of inverse synthetic aperture radar technology, and particularly relates to an intelligent reconstruction method for asteroid point clouds based on single-view radar images. Background Technology
[0002] Asteroids are an important component of the solar system, and studying them is crucial for understanding Earth's evolution and human development. Asteroids are considered "fossils" of early solar system material, preserving the original components of the solar system before multiple collisions or evolutions. Furthermore, asteroids have undergone long periods of collisions, mergers, and fragmentation. Analyzing their shapes and surface structures helps us understand the high-energy collision environment of the early solar system and provides technological support for asteroid formation and interactions, as well as planetary defense.
[0003] Current methods for reconstructing the three-dimensional shape of asteroids primarily utilize shape inversion methods that search for the spherical harmonic series of asteroids using optical and radar data. Optical telescopes are the main equipment for acquiring light curve data, but they are greatly affected by weather conditions and limited by the size of the asteroid, essentially providing only a single "bright spot." Ground-based radar is the main means of acquiring radar data for asteroids and is currently the most powerful method for asteroid detection, offering advantages such as all-weather, all-day visibility and the ability to achieve high-precision two-dimensional range-Doppler images, with the highest resolution reaching the meter level. A representative method of this approach is the SHAPE software, which can combine light curves, radar images, and other data to construct an initial fitting model for simulation imaging. It then uses methods such as least squares to fit the model to the actual observation data, finding the best-matching set of parameters to determine the asteroid's shape.
[0004] Current traditional methods mainly rely on optimizing the errors between optical, radar, and simulation data. Parametric methods require many parameters and are time-consuming. With so many parameters, it is easy to get trapped in local optima, leading to distortion of the retrieved asteroid's 3D shape. At the same time, due to the complexity of the inversion process of traditional methods, not only is it necessary to effectively set the initial values of the model, but it also requires a lot of experience to manually adjust the parameters, resulting in a large amount of computation and long processing time, which is not conducive to obtaining the asteroid's shape quickly and accurately. Summary of the Invention
[0005] In view of this, the present invention provides an intelligent reconstruction method for asteroid point clouds based on single-view radar images. The present invention uses a convolutional neural network to build an intelligent reconstruction network for asteroid 3D point clouds, which can learn the morphological features from massive asteroid 2D imaging data, train a pre-trained model, and quickly and effectively reconstruct the 3D shape of asteroids.
[0006] The technical solution for implementing the present invention is as follows:
[0007] A method for intelligent reconstruction of asteroid point clouds based on single-view radar images, the specific process of which is as follows:
[0008] Network Construction: The intelligent reconstruction network for asteroid point clouds includes an encoder module and a generator module. Each convolutional layer and hidden layer of the encoder module is followed by a dropout layer and a batch normalization layer. The generator module consists of fully connected layers, with each fully connected layer followed by a dropout layer and a batch normalization layer.
[0009] Training dataset creation: ISAR imaging simulation was performed using a known asteroid point cloud model to generate two-dimensional images of the asteroid;
[0010] Network training: Set the network loss function Loss, input the training dataset into the asteroid point cloud intelligent reconstruction network, and use the gradient descent method to update the network weights until the Loss value decreases to within the set range, thus obtaining the trained asteroid point cloud intelligent reconstruction network.
[0011] Intelligent reconstruction of asteroid point clouds: Input the two-dimensional image data of the asteroid to be reconstructed into the network to obtain the prediction results of the three-dimensional point cloud of the asteroid.
[0012] Optionally, the loss function Loss described in this invention includes chamfer loss Loss. CD Wasserstein distance loss EMD and regularization loss Laplacian .
[0013] Optionally, the chamfer loss described in this invention CD for:
[0014]
[0015] Where x represents a point in the point cloud S1 obtained in the previous training iteration, and y represents a point in the point cloud S2 obtained in the current training iteration. It represents the 2-norm.
[0016] Optionally, the Wasserstein distance loss described in this invention... EMD for:
[0017]
[0018] Where f(x,y) represents the weight mapping from point x in point cloud S1 to point y in point cloud S2, and d(x,y) is the distance between the two points.
[0019] Optionally, the regularization loss described in this invention... Laplacian
[0020]
[0021] Among them, v iLet N(i) be the coordinates of the i-th vertex in the point cloud, and let N(i) be the coordinates of vertex v. i The set of 10 nearest neighbors. It is the average position of the vertices in the nearest neighbor region of a vertex.
[0022] Optionally, the encoder module of the present invention sequentially includes a 2D convolutional layer, a residual block, a Flatten layer, and a hidden layer. The number of convolutional kernel channels in the 2D convolutional layer is set to 64, the kernel size is set to 3×3, the spatial dimension is preserved, and ReLU activation is applied. The residual block is set to 3 layers, each residual block consists of a main path and a short circuit. Each residual layer calculates the output of the main path (i.e., convolution, normalization, activation, etc.) and adds it to the short circuit connection to realize the residual connection, and ReLU activation is applied. The Flatten layer is used to flatten the multidimensional feature map into a one-dimensional vector. The hidden layer is set to 6 layers, and the feature dimensions are set to 512, 256, 128, 64, 32, and 16 respectively, compressing the features to low-dimensional features.
[0023] Optionally, the generator module of the present invention outputs a three-dimensional point cloud model containing 1024 points, and the output layer uses the Sigmoid function.
[0024] Optionally, the present invention uses gradient descent to update the weights of the network until the loss value drops to less than 10%, thus obtaining a trained asteroid point cloud intelligent reconstruction network.
[0025] Beneficial effects:
[0026] This invention utilizes convolutional neural networks to automatically extract features from two-dimensional images of asteroids, thus providing a method for rapidly reconstructing the shape of asteroids. A comparison of this invention with the traditional SHAPE method, which reconstructs the shape of asteroids using two-dimensional images, is as follows:
[0027] 1) Traditional methods, which rely on least squares to fit parameters, are greatly affected by the quality of asteroid radar images. When the image quality is poor, it is difficult to invert and fit the shape of the asteroid. The quality of asteroid radar images is affected by the physical characteristics of the asteroid and observation conditions. The signal-to-noise ratio of radar images varies when observing each asteroid. For image data with low signal-to-noise ratio, convolutional neural networks can effectively extract features from the two-dimensional images of asteroids.
[0028] 2) Compared with traditional methods, the convolutional neural network-based method can learn the shape features of asteroids from radar images using neural networks without human intervention. The pre-trained neural network model can be directly applied to asteroid 3D reconstruction tasks, quickly predicting the 3D shape of asteroids and improving reconstruction efficiency.
[0029] This invention extracts the shape features of asteroids from single-view radar images by constructing a convolutional neural network, which can efficiently and quickly obtain the three-dimensional shape information of asteroids. Therefore, this invention has a better ability to reconstruct asteroid point clouds than the traditional SHAPE method. Attached Figure Description
[0030] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0031] Figure 1 This is a schematic diagram of the process of the present invention;
[0032] Figure 2 This is a schematic diagram of the encoder module of the present invention;
[0033] Figure 3 This is a schematic diagram of the residual block structure in the encoder of the present invention;
[0034] Figure 4 This is the generator module of the present invention;
[0035] Figure 5 For the specific asteroid reconstruction results of this invention, (a) input ISAR image data, (b) reconstructed point cloud. Detailed Implementation
[0036] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
[0037] It should be noted that, in the absence of conflict, the following embodiments and features can be combined with each other; and, based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.
[0038] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this disclosure, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.
[0039] This application provides an intelligent reconstruction method for asteroid point clouds based on single-view radar images. By constructing an encoder-decoder neural network, shape features are learned from single-view asteroid image data, and the three-dimensional point cloud of the asteroid is reconstructed. For example... Figure 1 As shown, the present invention specifically includes the following steps:
[0040] Step 1: Constructing an intelligent reconstruction network for asteroid point clouds
[0041] In this embodiment, the intelligent reconstruction network for asteroid point clouds consists of an encoder module and a generator module connected together. Wherein:
[0042] The encoder module is used to extract asteroid shape features from 2D ISAR data and includes convolutional layers, hidden layers, and residual blocks. For example... Figure 2 As shown, in an embodiment of the present invention, it has a total of 11 layers, and the structure is as follows: 2D convolutional layer → first residual block → second residual block → third residual block → Flatten layer → first hidden layer → second hidden layer → third hidden layer → fourth hidden layer → fifth hidden layer → sixth hidden layer.
[0043] The parameters for each layer are set as follows: In the 2D convolutional layer, the number of kernel channels is set to 64, the kernel size is set to 3×3, spatial dimensions are preserved, and ReLU activation is used; the residual blocks are set to 3 layers, and the main path and short-circuit connections are calculated in each layer, and ReLU activation is also used. Figure 3 As shown; the Flatten layer flattens the multidimensional feature map into a one-dimensional vector; the feature dimensions in the first to sixth hidden layers are set to 512, 256, 128, 64, 32, and 16 respectively, compressing the features to low-dimensional features.
[0044] In this method, the encoder employs a convolutional layer structure. To extract higher-level features, residual blocks are also introduced to obtain deeper features. Random dropout layers and batch normalization are applied after each convolutional and hidden layer to reduce learning difficulty, improve the learning speed of the neural network model, and enhance generalization ability.
[0045] The generator module consists of fully connected layers that transform the latent space feature vectors output by the encoder into an output with a specific shape. For example... Figure 4 As shown, in an embodiment of the present invention, a three-dimensional point cloud model containing 1024 points is set. The generator gradually recovers the point cloud model through fully connected layers. After each fully connected layer, a dropout layer and a normalization layer are added to stabilize the training and prevent overfitting. The output layer uses the Sigmoid function to make the data conform to the requirements of the neural network.
[0046] Finally, connecting the encoder module and the generator module forms the asteroid point cloud intelligent reconstruction network of this invention.
[0047] Step 2: Create the training dataset:
[0048] In this invention, the dataset is generated by ISAR imaging simulation using a known asteroid point cloud model to produce a two-dimensional image of the asteroid. The original three-dimensional model data comes from the Asteroid Radar Research Database of NASA's Jet Propulsion Laboratory (JPL) and the DAMIT database.
[0049] Step 3: Training the intelligent reconstruction network for asteroid point clouds:
[0050] The training set is input into the asteroid point cloud intelligent reconstruction network, and the weights of the network are updated using the gradient descent method until the loss value drops to less than 10%, thus obtaining the trained asteroid point cloud intelligent reconstruction network.
[0051] This invention defines the Loss function as:
[0052] Loss Total =Loss CD +Loss EMD +Loss Laplacian
[0053] Among them, Loss CD For chamfering loss, a distance metric method used for point cloud comparison, Loss EMD For Wasserstein distance loss, consider the point-to-point correspondence between point clouds; Loss Laplacian The regularization loss constrains the consistency of adjacent vertices.
[0054] Loss CD The formula is as follows:
[0055]
[0056] Where x represents a point in point cloud S1, and y represents a point in point cloud S2.
[0057]
[0058] Where f(x,y) represents the weight mapping from point x in point cloud S1 to point y in point cloud S2, and d(x,y) is the distance between the two points.
[0059]
[0060] Among them, v i Here are the coordinates of the i-th vertex, and N(i) is the coordinate of vertex v. i The set of 10 nearest neighbors. It is the average position of the vertices in the nearest neighbor region of a vertex.
[0061] Step 4: Perform 3D point cloud reconstruction on the asteroid to be reconstructed:
[0062] Two-dimensional image data of asteroids are input into a trained intelligent reconstruction network for asteroid point clouds, and the network outputs prediction results of three-dimensional point clouds of asteroids.
[0063] Figure 5 This is the reconstruction result for the specific asteroid Vesta.
[0064] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for intelligent reconstruction of asteroid point clouds based on single-view radar images, characterized in that, The specific process is as follows: Network Construction: The intelligent reconstruction network for asteroid point clouds includes an encoder module and a generator module. Each convolutional layer and hidden layer of the encoder module is followed by a dropout layer and a batch normalization layer. The generator module consists of fully connected layers, with each fully connected layer followed by a dropout layer and a batch normalization layer. Training dataset creation: ISAR imaging simulation was performed using a known asteroid point cloud model to generate two-dimensional images of the asteroid; Network training: Set the network loss function Loss, input the training dataset into the asteroid point cloud intelligent reconstruction network, and use the gradient descent method to update the network weights until the Loss value decreases to within the set range, thus obtaining the trained asteroid point cloud intelligent reconstruction network. Intelligent reconstruction of asteroid point clouds: Input the two-dimensional image data of the asteroid to be reconstructed into the network to obtain the prediction results of the three-dimensional point cloud of the asteroid.
2. The intelligent reconstruction method for asteroid point clouds based on single-view radar images according to claim 1, characterized in that, The loss function Loss includes chamfer loss Loss. CD Wasserstein distance loss EMD and regularization loss Laplacian .
3. The intelligent reconstruction method for asteroid point clouds based on single-view radar images according to claim 2, characterized in that, The chamfering loss Loss CD for: Where x represents a point in the point cloud S1 obtained in the previous training iteration, and y represents a point in the point cloud S2 obtained in the current training iteration. It represents the 2-norm.
4. The intelligent reconstruction method for asteroid point clouds based on single-view radar images according to claim 2, characterized in that, The Wasserstein distance loss Loss EMD for: Where f(x,y) represents the weight mapping from point x in point cloud S1 to point y in point cloud S2, and d(x,y) is the distance between the two points.
5. The intelligent reconstruction method for asteroid point clouds based on single-view radar images according to claim 2, characterized in that, The regularization loss Loss Laplacian Among them, v i Let N(i) be the coordinates of the i-th vertex in the point cloud, and let N(i) be the coordinates of vertex v. i The set of 10 nearest neighbors. It is the average position of the vertices in the nearest neighbor region of a vertex.
6. The intelligent reconstruction method for asteroid point clouds based on single-view radar images according to claim 2, characterized in that, The encoder module sequentially includes 2D convolutional layers, residual blocks, flattened layers, and hidden layers. The number of convolutional kernel channels in the 2D convolutional layers is set to 64, the kernel size is set to 3×3, the spatial dimension is preserved, and ReLU activation is applied. The residual blocks are set to 3 layers, each consisting of a main path and a short circuit. Each residual layer calculates the main path output and adds it to the short circuit connection to realize the residual connection, and ReLU activation is applied. The flattened layers are used to flatten the multidimensional feature map into a one-dimensional vector. The hidden layers are set to 6 layers, with feature dimensions set to 512, 256, 128, 64, 32, and 16 respectively, compressing the features to low-dimensional features.
7. The intelligent reconstruction method for asteroid point clouds based on single-view radar images according to claim 2, characterized in that, The generator module outputs a 3D point cloud model containing 1024 points, and the output layer uses the Sigmoid function.
8. The intelligent reconstruction method for asteroid point clouds based on single-view radar images according to claim 2, characterized in that, The weights of the network are updated using gradient descent until the loss value drops to less than 10%, resulting in a well-trained intelligent reconstruction network for asteroid point clouds.