CN109657679B - A kind of identification method of application satellite function type - Google Patents

Abstract

The invention provides an application satellite function type identification method, comprising: acquiring a target satellite space image, and adjusting the resolution of the acquired target satellite space image to obtain the target image; and performing data analysis on the target image based on a ResNet neural network model Process to determine the function type corresponding to the target satellite. The application satellite function type identification method provided by the present invention obtains an image that can be recognized by the ResNet neural network model by adjusting the resolution of the target satellite space image, and independently identifies the application satellite function type through the ResNet neural network model on-orbit. Relying on manual interpretation and judgment on the ground, the recognition efficiency is improved, and the need for real-time service and operation of non-cooperative space targets lacking prior information is met.

Description

A kind of identification method of application satellite function type

technical field

The invention relates to an application satellite function type identification method, which belongs to the technical field of satellite identification.

Background technique

Applied satellites are artificial satellites that directly serve the national economy, military activities, and cultural education. Among all kinds of artificial satellites, applied satellites have the largest number of launches and the largest variety. Various application satellites can be roughly divided into three categories according to the basic characteristics of their work, namely earth observation, radio relay and navigation and positioning reference. Applied satellites play an important role in military and civilian fields such as communications, navigation, and remote sensing.

On-orbit service for application satellites can prolong the life of satellites and improve the ability of mission execution, which is one of the current research hotspots at home and abroad. During the on-orbit service of the application satellite, operations such as auxiliary orbit change, fuel supply, attitude control, satellite takeover, and fault repair can be performed on the on-orbit satellite as needed. When performing in-orbit maintenance on a failed or faulty satellite, it is first necessary to safely approach the serviced satellite. For non-cooperative targets, it is difficult to obtain their star catalog features, key loads, and motion states in advance. During the approaching process, it is necessary to accurately recognize the target's functional type, motion state, and operating position to determine whether to approach docking or control. strategy and avoid collisions.

The existing space non-cooperative target type cognition usually obtains the space image of the target aircraft through the service aircraft and transmits the image to the ground command and control center. The ground command center uses image recognition methods such as edge detection and feature fitting to manually determine non-cooperative targets Target aircraft type, and send the confirmed type to the serving aircraft. The existing methods have the following problems: the satellite-ground loop has a large delay, which cannot meet the needs of real-time service and operation of non-cooperative space targets lacking prior information.

SUMMARY OF THE INVENTION

In view of the problems existing in the prior art, the present invention provides a method for identifying function types of application satellites. The method autonomously processes and classifies visible light images of space targets generated in orbit. The classification speed of single target function type is less than 100ms, and the accuracy rate is to 90%.

The technical solution of the present invention is:

An application satellite function type identification method, comprising:

Obtain the target satellite space image, and adjust the resolution of the obtained target satellite space image to obtain the target image;

Data processing is performed on the target image based on the ResNet neural network model, and the function type corresponding to the target satellite is determined.

In an optional embodiment, the ResNet neural network model includes an initial convolution layer, three residual learning modules and a fully connected layer, each of the residual learning modules includes two residual learning units, and the initial The convolutional layer outputs a feature atlas to the first residual learning module, the first residual learning module outputs a new feature atlas to the second residual learning module, and the second residual learning module The module outputs a new feature atlas to the third residual learning module, and the third residual module outputs a new feature atlas to the fully connected layer, where:

The initial convolutional layer is used for:

Perform a two-dimensional convolution on the target image to obtain a feature atlas;

The first residual learning unit of the residual learning module is used for:

Perform a convolution operation on the feature map set input to this unit to obtain the residual feature map set of this unit; perform a normalization operation, an activation operation and a convolution operation on the feature map set input to this unit in turn to obtain a volume The feature atlas after product operation, perform normalization operation, activation operation and one convolution operation on the feature atlas after the first convolution operation in turn, to obtain the feature atlas after the second convolution; The residual feature map set and the feature map set after the second convolution of this unit determine the feature map set output by this unit, and output the feature map set output by this unit to the first step of the residual learning module. Two residual learning units;

The second residual learning unit of the residual learning module is used for:

Perform the normalization operation, activation operation and one convolution operation on the feature map set input to this unit in turn to obtain the feature map set after one convolution, and perform the normalization operation, activation operation and A convolution operation is performed to obtain the feature map set after the second convolution; the output feature map set of the unit is determined according to the feature map set input to the unit and the feature map set after the second convolution of the unit;

The fully connected layer is used for:

Perform average pooling on the feature atlas output by the third residual learning module, extract feature vectors used for satellite type identification, and perform a full connection operation, and determine the feature accumulation vector corresponding to each satellite type according to the feature vectors , and perform classification probability statistics through the classifier, so as to determine the function type corresponding to the target satellite.

In an optional embodiment, the resolution of the target image is not lower than 256×256.

In an optional embodiment, the described application satellite function type identification method further includes:

establishing a satellite space image sample library, the sample library includes a plurality of satellite types and image sample sets corresponding to each of the satellite types;

Based on the satellite space image sample library, the initial ResNet neural network model is trained and tested to obtain the ResNet neural network model.

In an optional embodiment, the building a satellite space image sample library includes:

Establish three-dimensional models of different types of satellites, simulate the space environment, image the established three-dimensional models, obtain a certain number of simulated space image samples, establish a correspondence between satellite types and the simulated space image samples, and generate satellite space images sample library.

In an optional embodiment, the building three-dimensional models of different types of satellites includes:

According to the structural characteristics of various types of satellites, establish three-dimensional models of the structures of different types of satellites;

According to the surface texture information of the various types of satellites, the three-dimensional models of the structures are rendered to obtain three-dimensional models of different types of satellites.

In an optional embodiment, the simulated space environment includes:

The simulated light source is parallel light, the density of atmospheric molecules is 0-0.01 times the density of atmospheric molecules on the ground, the light intensity index is 2-3 times the daily light intensity on the ground, and the incident direction of the light source is randomly generated in the 4π space of the three-dimensional model.

In an optional embodiment, the imaging of the established three-dimensional model includes:

The established three-dimensional model is randomly imaged within the range of 60° cone angle except the angle between the three-dimensional model and the normal to the sky, and during imaging, the angle between the light source beam direction and the camera imaging axis is less than 45°.

In an optional embodiment, the angle between the beam direction of the light source and the imaging axis of the camera is determined according to the following formula:

Where α<45°, is the angle between the light source beam direction and the camera imaging axis;

R1 is the distance from the light source to the origin of the 3D model of the satellite, R2 is the distance from the camera to the origin of the 3D model of the satellite, x1 is the x-axis coordinate of the light source, X2 is the x-axis coordinate of the camera, y1 is the y-axis coordinate of the light source, and y2 is the y-axis of the camera Coordinates, z1 is the z-axis coordinate of the light source, and z2 is the z-axis coordinate of the camera.

In an optional embodiment, after obtaining a certain number of simulated space image samples, the method further includes:

performing data enhancement on each of the simulated space image samples to obtain a simulated space image sample set with an expanded number of samples;

Correspondingly, establishing the correspondence between the satellite type and the simulated space image includes:

establishing a correspondence between the satellite type and the simulated space samples in the expanded simulated space image sample set.

In an optional embodiment, the initial ResNet neural network model is trained and tested based on the satellite space image sample library, including:

Each sample in the satellite space image sample library is processed from a three-channel color image into a one-channel grayscale image, and then the initial ResNet neural network model is trained and tested.

The beneficial effects of the present invention compared with the prior art are:

(1) The application satellite function type identification method provided by the present invention obtains an image that can be identified by the ResNet neural network model by adjusting the resolution of the target satellite space image, and autonomously carries out the application satellite function type through the ResNet neural network model on-orbit. Recognition does not depend on manual interpretation and judgment on the ground, which improves the recognition efficiency and meets the needs of real-time service and operation of non-cooperative space targets lacking prior information.

(2) The present invention does not need to use a series of image processing algorithms such as traditional image segmentation, recognition and classification when the application satellite function type is independently identified in orbit. The operation is simple, the identification speed is fast, and the accuracy rate is high, which can improve the accuracy of application satellite functions. Type of on-orbit autonomous identification speed.

(3) The present invention can realize autonomous processing and classification and identification of visible light images of space targets generated by spacecraft on-orbit, can simultaneously adapt to three-channel color images and one-channel grayscale images, and can adapt to image data of different lighting conditions and shooting angles, Strong adaptability to data.

(4) The present invention simulates the visible light imaging environment and the reflection characteristics of star catalogue materials under the condition of space vacuum, and generates image samples with real visible light reflection characteristics under different illumination and angles, which greatly enriches the deep learning sample library, and is used for classification and identification of space targets. The neural network training provided by the neural network provides sufficient training and testing material, which improves the reliability of the resulting model.

(5) The present invention can realize autonomous processing and classification and identification of visible light images of space targets generated on orbit by space vehicles, the classification speed of target function types is less than 100ms, the accuracy rate can reach 90%, and the intelligent cognition level of space targets can be greatly improved. Enhance the autonomous capability of the aircraft in orbit.

Description of drawings

1 is a flowchart of a method for identifying a function type of an application satellite provided by an embodiment of the present invention;

FIG. 2 is a schematic diagram of imaging the established three-dimensional model by simulating a space environment according to an embodiment of the present invention.

Detailed ways

The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

An embodiment of the present invention provides a method for identifying a function type of an application satellite, comprising the following steps:

Step 101: acquiring a target satellite space image, and adjusting the resolution of the acquired target satellite space image to obtain a target image;

Step 102: Perform data processing on the target image based on the ResNet neural network model, and determine the function type corresponding to the target satellite.

The application satellite function type identification method provided by the present invention obtains an image that can be identified by the ResNet neural network model by adjusting the resolution of the target satellite space image, and independently identifies the application satellite function type through the ResNet neural network model on-orbit. Relying on manual interpretation and judgment on the ground, the recognition efficiency is improved, and the need for real-time service and operation of non-cooperative space targets lacking prior information is met.

Table 1 Specific structural parameters of the ResNet neural network model

Among them, Conv0 is the initial convolutional layer, ConV1_x is the first residual learning module, ConV2_x is the second residual learning module, ConV3_x is the third residual learning module, and Full connection is the fully connected layer.

As shown in Table 1, in an optional embodiment, the ResNet neural network model includes an initial convolution layer, three residual learning modules and a fully connected layer, and each residual learning module includes two residuals A learning unit, the initial convolution layer outputs a feature atlas to the first residual learning module, the first residual learning module outputs a new feature atlas to the second residual learning module, the second Each of the residual learning modules outputs a new feature atlas to the third residual learning module, and the third residual module outputs a new feature atlas to the fully connected layer, where:

The initial convolutional layer is used for:

Perform a two-dimensional convolution on the target image to obtain a first feature atlas, wherein the preferred number of convolution kernels is 16, the size of convolution kernels is 3×3, and the convolution operation step size is 1;

The first residual learning unit of the first residual learning module is used for:

Perform a convolution operation on the first feature atlas to obtain a second feature atlas, wherein the number of convolution kernels is preferably 16, the size of the convolution kernels is 1×1, and the convolution operation step size is 1; The feature atlas is subjected to a normalization (normalization) operation, an activation operation, and a convolution operation in turn to obtain a third feature atlas. Among them, the number of convolution kernels is preferably 16, the size of the convolution kernel is 3×3, and the step size of the convolution operation is is 1; perform normalization operation, activation operation and one convolution operation on the third feature map set in turn to obtain a fourth feature map set, wherein the number of convolution kernels is preferably 16, the size of convolution kernels is 3×3, and the convolution kernel size is 3×3. The operation step is 1; the fifth feature atlas is determined according to the second feature atlas and the fourth feature atlas. Specifically, the corresponding features of each image in the second and fourth feature atlases are set. The values are added to obtain the images in the fifth feature atlas;

The second residual learning unit of the first residual learning module is used to:

Perform a normalization operation, an activation operation, and a convolution operation on the fifth feature map set in turn to obtain a sixth feature map set, wherein the preferred number of convolution kernels is 16, the size of the convolution kernel is 3×3, and the step size of the convolution operation is is 1; perform normalization operation, activation operation and one convolution operation on the sixth feature map set in turn to obtain the seventh feature map set, wherein the number of convolution kernels is preferably 16, the size of convolution kernels is 3×3, and the convolution kernel size is 3×3. The operation step is 1; the eighth feature atlas is determined according to the fifth and seventh feature atlases;

The first residual learning unit of the second residual learning module is used for:

Perform a convolution operation on the eighth feature atlas to obtain the ninth feature atlas;

Perform the normalization operation, activation operation and one convolution operation on the eighth feature atlas in turn to obtain the tenth feature atlas, and perform the normalization operation, activation operation and one convolution operation on the tenth feature atlas in turn to obtain the eleventh feature atlas, determining the twelfth feature atlas according to the eleventh feature atlas and the ninth feature atlas;

The second residual learning unit of the second residual learning module is used to:

Perform the normalization operation, activation operation, and a convolution operation on the twelfth feature atlas in turn to obtain the thirteenth feature atlas, and perform the normalization operation, activation operation, and a convolution operation on the thirteenth feature atlas in turn to obtain the thirteenth feature atlas. Fourteen feature atlas; determine the fifteenth feature atlas according to the fourteenth and twelfth feature atlas;

The first residual learning unit of the third residual learning module is used for:

Perform a convolution operation on the fifteenth feature atlas to obtain the sixteenth feature atlas;

Perform the normalization operation, activation operation and one convolution operation on the fifteenth feature atlas in turn to obtain the seventeenth feature atlas, and perform the normalization operation, activation operation and one convolution operation on the seventeenth feature atlas in turn to obtain the first Eighteen feature atlases; determine the nineteenth feature atlas according to the eighteenth and fifteenth feature atlases;

The second residual learning unit of the third residual learning module is used for:

Perform a normalization operation, an activation operation and a convolution operation on the nineteenth feature atlas in turn to obtain the twentieth feature atlas, and perform a normalization operation, an activation operation and a convolution operation on the twentieth feature atlas in turn to obtain the first Twenty-one feature atlases; the twenty-second feature atlas is determined according to the twenty-first feature atlas and the nineteenth feature atlas.

In the embodiment of the present invention, the preferred number of convolution kernels for convolution operation in the second residual learning module is 32, and the convolution step size is 2; the preferred number of convolution kernels for convolution operation in the third residual learning module is 64 , with a convolution stride of 2.

The fully connected layer is used for:

Perform average pooling on the twenty-second feature atlas, extract feature vectors used for satellite type identification, and perform a full connection operation, determine the feature accumulation vector corresponding to each satellite type according to the feature vectors, and use the classifier to classify the probability Statistics, so as to determine the function type corresponding to the target satellite.

Specifically, the fully connected layer determines the feature accumulation vector corresponding to each satellite type according to the following formula, and determines the category through the softmax multi-classifier:

where a1, a2,...aT are the feature accumulation vectors corresponding to each satellite type output by the fully connected layer, W is the feature weight matrix, x is the input of the fully connected layer feature vector, b is the fully connected layer bias parameter, and T is the target The number of categories, N is the number of feature vectors input to the fully connected layer.

The fully connected layer completes the transformation from distributed feature representation to the sample label space, which is a key step for target classification; the fully connected layer integrates the local feature vectors obtained in the previous convolution and pooling layers through the weight matrix, which can be compared The large retention of the representation ability of the model is beneficial to the subsequent model fine-tuning and transfer learning.

The model of the ResNet deep neural network of the present invention is based on the deep learning theory, which can simulate the neural connection structure of the human brain. Its image data processing flow conforms to the cognitive law of primate visual system, that is, it first detects edges and initial shapes, and then gradually forms more complex visual shapes. The applied satellite function type cognitive neural network model can form a more abstract high-level representation, attribute category or feature by combining low-level features, and finally give a hierarchical feature representation of image data.

The applied satellite function type cognitive neural network model can transform the feature representation of the sample in the original space into a new feature space by performing layer-by-layer feature transformation on the original signal, and automatically learn to obtain a hierarchical feature representation, which is more conducive to classification or visualization of features. The model is hierarchical, with many parameters and large enough capacity, so it can better represent the data features. For the difficult problem of image recognition, good results can be achieved on the basis of a large amount of training data.

The front-end input of the convolutional neural network used in this model adopts several convolution kernels to extract image information, and fully considers the translation, rotation and scaling invariance of the image target in space. The convolution kernels have the same structure and weight. Value sharing enables the neural network to maintain a large front-end scale while having fewer variable adjustment parameters, which greatly reduces the computational load and the burden of parameter optimization. The model maintains a relatively low amount of computation while the scale of the input image signal remains unchanged. Compared with the traditional artificially set image preprocessing filtering and convolution processes, the front-end processing process has been optimized for performance, and features are automatically extracted. It is specific to the image content, so the performance is better than the manual preprocessing process.

Compared with traditional image classification algorithms, this model uses relatively less preprocessing, does not need to rely on prior knowledge, and can avoid the difficulty of manual feature design in traditional image classification algorithms to the greatest extent, and uses filters for automatic feature extraction. It can quickly and accurately classify and identify unknown targets.

The ResNet neural network structure used in this model allows the original input information to be retained in the feature extraction results, effectively protecting the integrity of the information, and eliminating the phenomenon that the error in the training set increases with the deepening of the number of layers in the neural network, which can be extremely Quickly accelerate the training of ultra-deep neural networks, the accuracy of the model is also greatly improved, and it has better portability.

In an optional embodiment, the resolution of the target image is not lower than 256×256.

In an optional embodiment, the described application satellite function type identification method further includes:

establishing a satellite space image sample library, the sample library includes a plurality of satellite types and image sample sets corresponding to each of the satellite types;

Based on the satellite space image sample library, the initial ResNet neural network model is trained and tested to obtain the ResNet neural network model.

In an optional embodiment, the building a satellite space image sample library includes:

Establish three-dimensional models of different types of satellites, simulate the space environment, image the established three-dimensional models, obtain a certain number of simulated space image samples, establish a correspondence between satellite types and the simulated space image samples, and generate satellite space images sample library.

In an optional embodiment, the building three-dimensional models of different types of satellites includes:

According to the structural characteristics of various types of satellites, establish three-dimensional models of the structures of different types of satellites;

According to the surface texture information of the various types of satellites, the three-dimensional models of the structures are rendered to obtain three-dimensional models of different types of satellites.

In an optional embodiment, the simulated space environment includes:

The simulated light source is parallel light, the density of atmospheric molecules is 0-0.01 times the density of atmospheric molecules on the ground, the light intensity index is 2-3 times the daily light intensity on the ground, and the incident direction of the light source is randomly generated in the 4π space of the three-dimensional model.

In an optional embodiment, the imaging of the established three-dimensional model includes:

The established three-dimensional model is randomly imaged within the range of 60° cone angle except the angle between the three-dimensional model and the normal to the sky, and during imaging, the angle between the light source beam direction and the camera imaging axis is less than 45°.

In an optional embodiment, the angle between the beam direction of the light source and the imaging axis of the camera is determined according to the following formula:

Where α<45°, is the angle between the light source beam direction and the camera imaging axis;

R1 is the distance from the light source to the origin of the 3D model of the satellite, R2 is the distance from the camera to the origin of the 3D model of the satellite, x1 is the x-axis coordinate of the light source, X2 is the x-axis coordinate of the camera, y1 is the y-axis coordinate of the light source, and y2 is the y-axis of the camera Coordinates, z1 is the z-axis coordinate of the light source, and z2 is the z-axis coordinate of the camera.

In an optional embodiment, after obtaining a certain number of simulated space image samples, the method further includes:

performing data enhancement on each of the simulated space image samples to obtain a simulated space image sample set with an expanded number of samples;

Correspondingly, establishing the correspondence between the satellite type and the simulated space image includes:

establishing a correspondence between the satellite type and the simulated space samples in the expanded simulated space image sample set.

In an optional embodiment, the initial ResNet neural network model is trained and tested based on the satellite space image sample library, including:

Each sample in the satellite space image sample library is processed from a three-channel color image into a one-channel grayscale image, and then the initial ResNet neural network model is trained and tested.

The following is a specific embodiment of the present invention:

(1) Establish a sample library of satellite air images:

Collect and organize satellite images of various types of public information, and select three types of satellite images that can fully show the shape of the satellite and at least partially reflect the characteristics of the respective star catalogs of the three types of application satellites (such as the camera load of earth observation satellites, and the communication antenna of radio relay satellites). satellite images;

Firstly, according to the satellite external information displayed in each satellite image, the geometric ratio data such as satellite outline and height are calculated, and according to the calculated ratio, the origin of the three-dimensional coordinate system is used as the centroid of the satellite body to construct a three-dimensional white model of the satellite (without surface texture data);

According to the characteristics of the star catalog displayed in each satellite image, the solar panels, camera lenses, sensors, thrusters, star-arrow docking rings, measurement and control antennas, data transmission antennas and other equipment and components are added to the three-dimensional white model of the satellite in proportion to obtain the satellite. 3D model of the structure (without surface texture data);

According to the actual material reflection characteristics of the star catalog, the surface texture information of the three types of satellites is determined. In this embodiment, the visible light reflection characteristics of the materials of the main parts of the star catalog meet the requirements of the real material reflection characteristics, and the main parts include the front of the solar panel (covering the solar cells). , the back of the solar panel, aluminized thermal control multi-layer, gold-plated thermal control multi-layer, secondary surface mirror and white paint, and other parts of the star catalog, etc.; the three-dimensional model of the satellite structure is subsequently rendered according to the surface texture information, to obtain three-dimensional models of three types of satellites;

Referring to Figure 2, with the origin of the coordinates of the three-dimensional model of the satellite as the center, a spherical surface with a radius of R is established, and the spherical surface is divided into (N+1)×(N+1) points by N longitude lines and N latitude lines, then each The X coordinate of the point is x(i,j), the Y coordinate is y(i,j), and the Z coordinate is z(i,j), where i and j are the numbers of the latitude and longitude, respectively, and their range is 1～ N. Let the distance between the parallel light source of the simulated sunlight and the origin of the three-dimensional satellite model be R1, then the parallel light source is on a spherical surface with a radius of R1, and the latitude i=a and the meridian j=b are randomly selected. At this time, the coordinates of the light source are (x1(a , b), y1(a, b), z1(a, b));

In the same way, set the distance between the camera imaging the 3D satellite model and the origin of the satellite 3D model as R2, then the imaging camera randomly selects the latitude line number i=c and the longitude line number j=d value on a spherical surface with a radius of R2. At this time, the camera coordinates is (x2(c, d), y2(c, d), z2(c, d)).

Since most of the application loads of the application satellites are mainly on the earth, in order to reflect the key loads and characteristic parts of the satellite surface as much as possible in the simulation imaging process, to avoid monotonous imaging of the top surface of the satellite, and to increase the diversity and richness of satellite samples as much as possible , the viewing angle direction of the satellite into each satellite optical imaging image is randomly generated within the range of 60° cone angle excluding the normal angle between the satellite and the sky. Therefore, when selecting the latitude line number i, avoid selecting the latitude line within the range of 60° cone angle, that is, i>(60°/180°)×N.

In addition, because the space vacuum environment has no influence of air scattering factors, the imaging contrast between the star table part illuminated by the light and the occluded star table part is very large, so when the imaging angle of the visible light camera and the parallel light source are too large, it will cause most of the occlusion. It can be seen that the imaging effect is poor, it cannot normally reflect the characteristics of the star catalog and the shape of the satellite, and it cannot be used as an effective sample for neural network training. Therefore, the angle between the light source beam direction and the camera imaging axis should be less than 45°, and the angle between the light source and the camera is α. The calculation formula is as follows:

And α<45°.

When the position of the light source and the position of the camera meet the above requirements, the position coordinates of the light source and the camera are brought into the three-dimensional modeling software (3D MAX), and the three-dimensional model is imaged and rendered. Multiple sets of camera and light source positions can be set as needed to enrich the diversity of samples, increase the number of samples, and enhance the learning ability of the neural network.

In the model building software, set the imaging size to 1920×1080 to render the imaging effect. The total number of samples is not less than 5,000, covering three types of application satellites evenly.

Count the number of satellite samples of each type, and expand the number of satellite samples of each type in proportion. Image processing methods such as rotation, translation, and flipping are mainly used to enhance the number of samples. The rotation is a random angle from 0 to 360 degrees, and the translation distance does not exceed 1/32 of the length or width of the image, so as to avoid translating the satellite body outside the image;

According to the expanded samples, three types of satellite space image sample libraries are generated. The three types of satellite samples in the space image sample library are equal in number. 1/3 of the images are randomly selected from each type of satellite samples as test samples, and the rest of the images are used as training samples.

(2) Tensorflow platform is used for deep convolutional neural network learning and training. Convert the training sample set and the test sample set to TFRecord format files accepted by the Tensorflow platform, with the suffix of ".tfrecords". During the generation of the format file, the size of all sample images is adjusted to 360×640, and all three-channel color image data are converted into single-channel grayscale image data;

Set the number of batch images of deep convolutional network Batch_size=10, all training samples are learned for one epoch at a time, and the total training time is 250 epochs. Complete an epoch to complete an evaluation using test samples. The initial learning rate learning_rate is 0.1, adjusted to 0.01 after 100 epochs, 0.001 after 150 epochs, and 0.0001 after 200 epochs. The number of recognition categories is 3;

According to the deep learning typical convolutional neural network ResNet residual network model, the deep convolutional neural network model shown in Table 1 is established. The neural network consists of 14 layers, consisting of 1 convolutional layer, 3 residual learning modules (blocks), and 1 fully connected layer. Each residual learning module includes two residual learning units (bottleneck), each The residual learning unit adopts a 2-layer structure, and each layer structure includes a convolution operation. The number of filters in the first block is 16, the number of filters in the second block is 32, and the number of filters in the third block is 64. During the training process, the neural network model is stored for subsequent test evaluation. The network model uses the Relu activation function, uses the full connection for output after average pooling, uses the softmax multi-classifier for prediction, and the optimization algorithm uses the Momentum algorithm;

Use the spatial image sample library generated in step (1) and the deep convolutional neural network model shown in Table 1 to perform training and testing of autonomous identification of satellite function types. The number of batch image training in the training process batch_size is 10, and 10 images from the training sample library are randomly selected each time and sent to the neural network for processing until all samples are trained, which is an epoch. A total of 250 epochs were completed in the autonomous identification training of three types of satellite function types. A test is performed every 10 epochs, that is, all test samples are brought into the neural network for recognition, and the recognition accuracy is obtained after the recognition classification results are compared with the labels. The initial learning rate for neural network training is set to 0.1.

After the neural network training is completed, the trained network model is used to carry out the autonomous identification test of the application satellite function type, and the test accuracy rate is counted. The self-identification test of the application satellite function type can be divided into two types: unlabeled sample test and labeled (labeled as the application satellite type) sample test.

When the unlabeled sample test is performed, that is, when there is no prior sample type information, the trained neural network model is used to identify the sample image to obtain three types of application satellite recognition probabilities. Statistics of recognition accuracy.

When testing labeled samples, that is, when you know in advance which type of satellite the sample to be tested is, use the trained neural network model to identify the sample image, and then automatically compare the satellite category with the highest probability with the sample label to determine whether the recognition is correct. . When a large number of labeled samples are tested, the statistical results of the recognition accuracy can be obtained.

When the recognition accuracy rate reaches 90%, the trained model is considered as the final model. During the on-orbit test, the space image of the target satellite is obtained, and the resolution of the obtained space image of the target satellite is adjusted to obtain the target image. The model performs data processing on the target image to determine the function type corresponding to the target satellite.

The above is only the best specific embodiment of the present invention, but the protection scope of the present invention is not limited to this. Substitutions should be covered within the protection scope of the present invention.

The parts not described in detail in the present invention belong to the common knowledge of those skilled in the art.

Claims (5)

1. An application satellite function type identification method is characterized by comprising the following steps:

acquiring a target satellite space image, and adjusting the resolution of the acquired target satellite space image to obtain a target image;

performing data processing on the target image based on a ResNet neural network model, and determining a function type corresponding to the target satellite;

the deep convolutional neural network model adopts a ResNet residual error network structure and comprises an initial convolutional layer, three residual error learning modules and a full-connection layer, wherein each residual error learning module comprises two residual error learning units, and each residual error learning unit comprises two convolutional operations;

the initial convolutional layer for:

performing one-time two-dimensional convolution on the target image to obtain a first feature map set, wherein the number of convolution kernels is 16, the size of the convolution kernels is 3 multiplied by 3, and the step length of convolution operation is 1;

a first residual learning unit of the first residual learning module, configured to:

performing convolution operation on the first feature map set to obtain a second feature map set, wherein the number of convolution kernels is 16, the size of the convolution kernels is 1 multiplied by 1, and the step length of the convolution operation is 1; sequentially carrying out standardization operation, activation operation and one convolution operation on the first feature map set to obtain a third feature map set, wherein the number of convolution kernels is 16, the size of the convolution kernels is 3 multiplied by 3, and the step length of the convolution operation is 1; sequentially carrying out standardization operation, activation operation and one convolution operation on the third feature map set to obtain a fourth feature map set, wherein the number of convolution kernels is 16, the size of the convolution kernels is 3 multiplied by 3, and the step length of the convolution operation is 1; adding the corresponding characteristic values of the images in the second characteristic image set and the fourth characteristic image set to obtain an image in a fifth characteristic image set;

a second residual learning unit of the first residual learning module, configured to:

sequentially carrying out standardization operation, activation operation and one convolution operation on the fifth feature map set to obtain a sixth feature map set, wherein the number of convolution kernels is 16, the size of the convolution kernels is 3 multiplied by 3, and the step length of the convolution operation is 1; sequentially carrying out standardization operation, activation operation and one convolution operation on the sixth feature map set to obtain a seventh feature map set, wherein the number of convolution kernels is 16, the size of the convolution kernels is 3 multiplied by 3, and the step length of the convolution operation is 1; determining an eighth feature map set according to the fifth feature map set and the seventh feature map set;

a first residual learning unit of the second residual learning module, configured to:

performing convolution operation on the eighth feature map set to obtain a ninth feature map set;

carrying out standardization operation, activation operation and one convolution operation on the eighth feature map set in sequence to obtain a tenth feature map set, carrying out standardization operation, activation operation and one convolution operation on the tenth feature map set in sequence to obtain an eleventh feature map set, and determining a twelfth feature map set according to the eleventh feature map set and the ninth feature map set;

a second residual learning unit of the second residual learning module, configured to:

sequentially carrying out standardization operation, activation operation and one convolution operation on the twelfth feature atlas to obtain a thirteenth feature atlas, and sequentially carrying out standardization operation, activation operation and one convolution operation on the thirteenth feature atlas to obtain a fourteenth feature atlas; determining a fifteenth feature atlas according to the fourteenth feature atlas and the twelfth feature atlas;

a first residual learning unit of the third residual learning module, configured to:

performing convolution operation on the fifteenth feature map set to obtain a sixteenth feature map set;

sequentially carrying out standardization operation, activation operation and one convolution operation on the fifteenth feature atlas to obtain a seventeenth feature atlas, and sequentially carrying out standardization operation, activation operation and one convolution operation on the seventeenth feature atlas to obtain an eighteenth feature atlas; determining a nineteenth feature map set according to the eighteenth feature map set and the fifteenth feature map set;

a second residual learning unit of the third residual learning module, configured to:

sequentially carrying out standardization operation, activation operation and one convolution operation on the nineteenth feature map set to obtain a twentieth feature map set, and sequentially carrying out standardization operation, activation operation and one convolution operation on the twentieth feature map set to obtain a twenty-first feature map set; determining a twenty-second feature map set according to the twenty-first feature map set and the nineteenth feature map set;

the number of convolution kernels for convolution operation in the second residual learning module is 32, and the convolution step length is 2; the convolution kernel number of convolution operation in the third residual learning module is 64, and the convolution step length is 2;

the full connection layer is used for:

performing average pooling on the twenty-second feature map set, extracting feature vectors for satellite type identification, performing full-connection operation, determining feature accumulation vectors corresponding to each satellite type according to the feature vectors, and performing classification probability statistics through a classifier so as to determine the function type corresponding to the target satellite;

the number of the images processed in batches by the deep convolutional network is 10, the total training period is 250, the initial learning rate is 0.1, the initial learning rate is adjusted to 0.01 after 100 periods, the initial learning rate is adjusted to 0.001 after 150 periods, and the initial learning rate is adjusted to 0.0001 after 200 periods;

establishing a satellite space image sample library, wherein the sample library comprises a plurality of satellite types and image sample sets corresponding to the satellite types;

training and testing an initial ResNet neural network model based on the satellite space image sample library to obtain a ResNet neural network model;

the establishing of the satellite space image sample library comprises the following steps:

establishing three-dimensional models of different types of satellites, simulating a space environment, imaging the established three-dimensional models to obtain a certain number of simulated space image samples, establishing a corresponding relation between the types of the satellites and the simulated space image samples, and generating a satellite space image sample library;

the establishment of the three-dimensional models of different types of satellites comprises the following steps:

carrying out satellite contour and height measurement and calculation according to satellite external information displayed by each satellite picture, and constructing a satellite three-dimensional white mold by taking the origin of a three-dimensional coordinate system as the centroid of a satellite body according to the measurement and calculation proportion;

adding a solar panel, a camera lens, a sensor, a thruster, a satellite and arrow docking ring, a measurement and control antenna and a data transmission antenna to a satellite three-dimensional white model according to the star surface characteristics displayed by each satellite picture in proportion to obtain structural three-dimensional models of different types of application satellites;

determining surface texture information of a satellite according to the reflection characteristics of the actual materials of the satellite surface, wherein the visible light reflection characteristics of the materials of the main parts of the satellite surface meet the requirements of the reflection characteristics of real materials, and the parts comprise the front surface of a solar panel, the back surface of the solar panel, an aluminum-plated thermal control multilayer, a gold-plated thermal control multilayer, a secondary surface mirror, white paint and other parts of the satellite surface;

rendering the structural three-dimensional model according to the surface texture information of each type of satellite to obtain three-dimensional models of different types of satellites;

the simulated spatial environment comprises:

the simulated light source is parallel light, the atmospheric molecular density is 0-0.01 times of the ground atmospheric molecular density, the illumination intensity index is 2-3 times of the ground daily illumination intensity, and the incident direction of the light source is randomly generated in a 4 pi space of the three-dimensional model;

in the light source and camera position setting, a spherical surface with the radius of R is established by taking the origin of coordinates of a three-dimensional model of a satellite as a center, and the spherical surface is divided into (N +1) × (N +1) points by N warps and N wefts, so that the X coordinate of each point is X (i, j), the Y coordinate is Y (i, j), the Z coordinate is Z (i, j), wherein i and j are respectively the numbers of the wefts and the warps, and the ranges of the i and the j are respectively 1-N; if the distance between the parallel light source simulating the sunlight and the origin of the three-dimensional satellite model is R1, randomly selecting a value of a latitude line i and a value of a longitude line j on a spherical surface with the radius of R1, wherein the coordinates of the light source are (x1(a, b), y1(a, b) and z1(a, b)); similarly, if the distance between the camera imaging the three-dimensional satellite model and the origin of the three-dimensional satellite model is R2, the imaging camera randomly selects the latitude line number i as c, the longitude line number j as d value on the spherical surface with the radius of R2, and the camera coordinates are (x2(c, d), y2(c, d), and z2(c, d)); the visual angle direction of each satellite optical imaging image formed by the satellite is randomly generated in the range of excluding the cone angle of 60 degrees formed by the satellite to the normal of the sky; the weft number i > (60 °/180 °) × N.

2. The method of claim 1, wherein the resolution of the target image is 360 x 640.

3. The method for identifying the type of the satellite function according to claim 1, wherein the angle between the light beam direction of the light source and the imaging axis of the camera is determined according to the following formula:

wherein alpha is less than 45 degrees and is an included angle between the light beam direction of the light source and the imaging axis of the camera;

r1 is the distance from the light source to the origin of the three-dimensional model of the satellite, R2 is the distance from the camera to the origin of the three-dimensional model of the satellite, x1 is the coordinates of the x axis of the light source, x2 is the coordinates of the x axis of the camera, y1 is the coordinates of the y axis of the light source, y2 is the coordinates of the y axis of the camera, z1 is the coordinates of the z axis of the light source, and z2 is the coordinates of the z axis of the camera.

4. The method for identifying the type of function of an application satellite according to claim 1, wherein after obtaining a certain number of simulated aerial image samples, the method further comprises:

performing data enhancement on each analog space image sample to obtain an analog space image sample set with the number of samples expanded;

correspondingly, the establishing of the corresponding relation between the satellite type and the simulated space image comprises:

and establishing a corresponding relation between the satellite type and the simulation space samples in the extended simulation space image sample set.

5. The method for identifying the application satellite function type according to claim 1, wherein the training and testing of the initial ResNet neural network model based on the satellite space image sample library comprises:

and processing each sample in the satellite space image sample library into a channel gray image from a three-channel color image, and then training and testing an initial ResNet neural network model.

Images