CN111680552B - Feature part intelligent recognition method - Google Patents

Abstract

The invention is an intelligent identification method of characteristic parts, which is suitable for the field of identification of local typical parts of space failed satellites. Traditional identification of typical target parts based on analytical algorithms has problems such as large edge point identification errors. The present invention designs an intelligent identification method for local typical feature parts based on a convolutional neural network. First, for the task of identifying local typical parts of failed satellites, a database of local typical parts of satellites containing rich information was created, the components of the typical parts were annotated, and a training data set and a test data set were constructed. Then build a deep convolutional network and use the training data set to train the network parameters. After the training is completed, the network can intelligently identify typical parts from the input image.

Description

An intelligent identification method for characteristic parts

Technical field

The invention belongs to the field of spatial target feature recognition, and specifically relates to an intelligent recognition method for feature parts.

Background technique

The protection and on-orbit services of important space targets, mainly satellites, have become an important development direction of aerospace technology in various countries around the world, and the identification technology of satellite characteristic components (such as antennas, sails, and engines) is a key link. With the increasing maturity of aircraft approach technology and optical imaging technology, it is possible to use space-based platforms to perform high-resolution optical imaging of space targets. This also puts forward higher requirements for satellite target identification, especially the accurate identification of satellite characteristic components. needs.

Spatial target recognition mainly uses spatial target characteristic data to effectively judge and identify its identity, posture, status and other attributes. At present, the source of target characteristic data is mainly ground-based detection, including optical and radar equipment. However, the detection data of ground-based equipment is related to many factors such as observation angle, target characteristics, solar angle and atmosphere, making the detection results highly uncertain. Although there are also studies on space-based target detection and recognition at home and abroad, most of them focus on point target detection at long distances and recognition of on-orbit motion states. The recognition methods used rely on known online feature points, and the features are Points have strong constraints. Once the known feature points change, it is difficult for the existing recognition methods to accurately identify the feature parts. In recent years, with the development of deep learning technology, the effect of image target recognition has been greatly improved, which has brought new methods and means to the field of spatial target recognition.

Contents of the invention

The technical problem solved by this invention is to overcome the shortcomings of existing recognition technology's over-reliance on known feature points, and propose an intelligent recognition method for feature parts, which can realize intelligent recognition of typical target parts and tactical features in complex spatial environments. , providing technical support for in-depth recognition and understanding of on-orbit scenes.

The technical solution adopted by the present invention is: an intelligent identification method of characteristic parts. The steps are as follows:

(1) Based on the attitude rotation method, construct a spacecraft image database in space. The spacecraft image database stores three-dimensional geometric images of the spacecraft under different three-axis attitude angle combinations;

(2) Perform a weighted average of the pixel values of the three color channels of red (R), green (G), and blue (B) for each three-dimensional geometric image in the spacecraft image database to obtain the grayscale of the three-dimensional geometric image. Image; based on the grayscale image, adding measurement noise can obtain a noisy grayscale image;

(3) Use the cubic interpolation method to reduce the grayscale image of each three-dimensional geometric image in the spacecraft image database to obtain the corresponding interpolated image;

(4) Rotate the grayscale image of each three-dimensional geometric image in the spacecraft image database in step (2) to obtain the rotated grayscale image;

(5) Interpolate the grayscale image of each three-dimensional geometric image in the spacecraft image database in step (2) and the grayscale image with noise, and the grayscale image of each three-dimensional geometric image in step (3), and step (4) After rotating the grayscale image of each three-dimensional geometric image, the contours of the characteristic parts of the spacecraft are marked respectively; and the names of the marked contours of the characteristic parts of the spacecraft are set respectively to obtain the labels of the characteristic parts and complete the characteristic parts of the target spacecraft. label;

(6) Construct a convolutional neural network, and use most of the images with the contours of the spacecraft's characteristic parts in the spacecraft image database to train the constructed convolutional neural network to obtain the trained convolutional neural network; the remaining The image with the outline of the spacecraft's characteristic parts in the spacecraft image database is input into the trained convolutional neural network to identify the spacecraft's characteristic parts, and the recognition results are output to realize the intelligent identification of the spacecraft's characteristic parts.

Preferably, (1) based on the attitude rotation method, construct a spacecraft image database in space. The spacecraft image database stores three-dimensional geometric images of the spacecraft under different combinations of three-axis attitude angles, specifically:

(1) Based on the attitude rotation method, construct a spacecraft image database in space, that is, set the three-axis attitude angle of the target spacecraft as the roll angle Pitch angle θ, yaw angle ψ. The transformation range of the three-axis attitude angle is [0, 360°], and the value is taken every N° to form multiple sets of three-axis attitude angle combinations of the spacecraft. Import the simulation modeling software to perform the simulation of the spacecraft at different three-axis attitude angles. Combining the acquisition of three-dimensional geometric images under the circumstances, a spacecraft image database is obtained.

Preferably, (2) perform a weighted average of the pixel values of the three color channels of red (R), green (G), and blue (B) of each three-dimensional geometric image in the spacecraft image database to obtain the three-dimensional geometric image. Grayscale image, specifically:

G(i,j)＝0.299×R(i,j)+0.578×G(i,j)+0.114×B(i,j)

In the formula: i, j are the horizontal and vertical coordinates of the grayscale image, i is greater than or equal to 1, j is greater than or equal to 1, R(i,j), G(i,j), and B(i,j) are grayscale images respectively. The pixel values corresponding to the red (R), green (G), and blue (B) pixels in row i and column j. G(i,j) is the pixel value of row i and column j in the grayscale image.

Preferably, (3) use cubic interpolation method to reduce the grayscale image of each three-dimensional geometric image in the spacecraft image database to obtain the corresponding interpolated image; specifically:

f(i+u,j+v)＝ABC

In the formula: f(i+u,j+v) is the pixel value of row i and column j of the interpolated image, u, v are the interpolation intervals, A, B, C are coefficient matrices, and their form is as follows:

A＝[S(1+u) S(u) S(1-u) S(2-u)]

C＝[S(1+v) S(v) S(1-v) S(2-v)] ^T

In the formula: S is the interpolation correlation function, specifically:

Through the above optimized rotation scheme and parameter requirements, the effect of image target recognition is greatly improved.

Preferably, (4) rotate the grayscale image of each three-dimensional geometric image in the spacecraft image database in step (2) to obtain the rotated grayscale image; specifically:

Assume that the coordinates of a certain pixel in the original image (that is, the grayscale image of the three-dimensional geometric image) are (i, j), and the coordinates in the rotated grayscale image after rotation are (i ₁ , j ₁ ), then the rotation transformation The matrix expression is:

The inverse transformation is:

In the formula, θ represents the angle of rotation around the center point of the original image (that is, the grayscale image of the three-dimensional geometric image) when changing from the original image to the rotated grayscale image in the plane where the original image is located;

a, b are the coordinates of the rotation center when the image is not rotated, c, d are the coordinates of the center point after the image is rotated, as:

In the formula, w ₀ is the width of the original image (before rotation), h ₀ is the length of the original image (before rotation); w ₁ is the width of the grayscale image after rotation, h ₁ is the length of the grayscale image after rotation; is Really simulates the constraint influence of the size (length, width) of the optical camera imaging. The size of the image of the rotated spacecraft in the optical camera changes. Therefore, the length h ₁ and width w ₁ of the rotated image are the same as the original image. h ₀ and width w ₀ change.

Through the above optimized rotation scheme and parameter requirements, the effect of image target recognition is greatly improved.

Preferably, (1) based on the attitude rotation method, construct a spacecraft image database in space. The spacecraft image database stores three-dimensional geometric images of the spacecraft under different combinations of three-axis attitude angles, specifically:

The target spacecraft is M types of spacecraft in space. The three-axis attitude angle combinations of multiple groups of target spacecraft are imported into the STK software. In the STK software, the three-dimensional geometric images of the M types of spacecraft under different three-axis attitude angle combinations are calculated. Obtain and store in spacecraft image database.

Preferably, in order to simulate the noise generated when using real shooting equipment to shoot the spacecraft, after the grayscale image is obtained in step (2), noise is added as the final grayscale image obtained in step (2); on the basis of the grayscale image On the above, adding measurement noise can obtain a noisy grayscale image, specifically: adding Gaussian noise for blurring, which further improves the accuracy of recognition.

Add Gaussian noise for blurring, specifically:

On the basis of the grayscale image of the three-dimensional geometric image, Gaussian noise is added for blurring to obtain a grayscale image with noise, specifically:

G _noise (i,j)＝G(i,j)+G _GB (i,j)

In the formula: i,j are the horizontal and vertical coordinates of the grayscale image, and G _noise (i,j) is the pixel value of row i and column j in the grayscale image with noise. σ is the standard deviation of the Gaussian noise normal distribution; r is the blur radius of the noisy grayscale image.

After adding noise, the grayscale image with noise of the spacecraft can be obtained, which can more realistically simulate the image of the spacecraft in the space environment, provide accurate images for neural network training and feature part identification, and improve the accuracy of feature part identification. Robustness and accuracy.

Preferably, the specific training of the constructed convolutional neural network is as follows: the convolutional neural network receives image data input, outputs the label of the feature part of the image as a prediction label, and inputs the prediction label and the label of the feature part of step (5) into the device together. A certain label loss function is used to calculate the training error through the loss function. According to the training error, the neural network parameters are dynamically adjusted to achieve error convergence. When the error converges to the set requirements, the trained convolutional neural network is obtained.

Preferably, in step (1), in different three-axis attitude angle combinations, the transformation range of the three-axis attitude angle is [0, 360°], and the value is taken every N°, and N is the common divisor of 360; a total of (360/ N) The combination of ^three groups of three-axis attitude angles of the target spacecraft.

Preferably, the three-dimensional geometric images of the spacecraft stored in the spacecraft image database in (1) under different three-axis attitude angle combinations are specifically:

Import the three-axis attitude angle combinations of multiple groups of target spacecraft into the simulation modeling software. The simulation modeling software is preferably Satellite Toll Kit (STK) software, which can build a three-dimensional model of the spacecraft, thereby obtaining the spacecraft Three-dimensional geometric images under different combinations of three-axis attitude angles.

Compared with the prior art, the beneficial effects of the present invention are:

(1) The present invention proposes an intelligent recognition method for feature parts, which integrates feature extraction and feature recognition, avoids manual design of features, has stronger versatility, and greatly improves the recognition of feature parts of space targets. Efficiency and precision.

(2) This invention is based on satellite image data from multiple sources, including two-dimensional images projected by satellite three-dimensional models, etc., taking into account the actual operating environment of the satellite in orbit, and simulates and renders the satellite images, so that not only can a more realistic satellite image be obtained, Image data can also increase the amount of satellite image data, which is more conducive to the training of convolutional networks.

(3) The present invention proposes an intelligent identification method of characteristic parts, which does not rely on known characteristic points in orbit. It still has strong identification ability for spacecraft without characteristic points, and has stronger versatility and recognition. The robust capability greatly improves the identification efficiency and accuracy of characteristic parts of the spacecraft, such as sail panels.

Description of the drawings

Figure 1 is a schematic diagram of the method of the present invention;

Figure 2 is a three-dimensional model of the spacecraft simulated by the method of the present invention;

Figure 3 is a graph of the loss function during the training process of the method of the present invention;

Figure 4 shows the recognition results obtained by the method of the present invention.

Detailed ways

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

The invention is an intelligent identification method of characteristic parts, which is suitable for the field of identification of local typical parts of failed space satellites. Traditional identification of typical target parts based on analytical algorithms has problems such as large edge point identification errors. The present invention designs an intelligent identification method for local typical feature parts based on a convolutional neural network. First, for the task of identifying local typical parts of failed satellites, a database of local typical parts of satellites containing rich information was created, the components of the typical parts were annotated, and a training data set and a test data set were constructed. Then build a deep convolutional network and use the training data set to train the network parameters. After the training is completed, the network can intelligently identify typical parts from the input image.

Tasks such as capturing failed spacecraft in space and rendezvous and docking between spacecrafts require precise acquisition of information such as the relative attitude of the spacecraft. The recognition of spacecraft characteristic parts is an important technical means to estimate relative attitude and other information. Traditional feature part identification methods based on analytical algorithms have problems such as large edge point identification errors and over-reliance on known feature points manually marked. In view of this, the present invention designs an intelligent identification method of local typical feature parts based on a convolutional neural network, which solves the problem of over-reliance on manually identified known feature points in the identification method, and achieves blurred images with low light and jitter. Identification of typical local characteristic parts of spacecraft (such as solar wings) in complex space environments.

As shown in Figure 1, a method for intelligent identification of characteristic parts of the present invention includes the following steps:

(1) Construct a spacecraft image database in space based on the attitude rotation method, that is, set the three-axis attitude angle of the target spacecraft as the roll angle Pitch angle θ, yaw angle ψ. Preferably, the transformation range of the three-axis attitude angle is [0, 360°], and the value is taken every N° (the N is the common divisor of 360, preferably 90°), forming a total of (360/N) ³ A combination of three-axis attitude angles of a group of target spacecraft;), forming a combination of three-axis attitude angles of multiple groups of target spacecraft, and importing it into the simulation modeling software to calculate the three-dimensional geometric image of the spacecraft under different combinations of three-axis attitude angles. Obtain and obtain the spacecraft image database, and the further optimized plan is as follows:

(1-1) Set the three-axis attitude angle and roll angle of the spacecraft The pitch angle θ and yaw angle ψ are in [0, 360°], and the values are taken every N° during the traversal, and the roll angle is traversed/> All combinations of pitch angle θ and yaw angle ψ form a total of (360/N) ³ combinations.

(1-2) Select the i-th spacecraft in the simulation modeling software "Satellite Toll Kit (STK)", and set the three-axis attitude of the spacecraft to all traversal combinations obtained in (1-1) Attitude, intercept the geometric picture of the spacecraft under each attitude combination.

(1-3) Select the i+1 spacecraft in the simulation modeling software "Satellite Toll Kit (STK)", and set the three-axis attitude of the spacecraft to all traversals obtained by (1-1) Combined attitude, intercept the geometric picture of the spacecraft under each combination of attitude. Assuming that the simulation modeling software "Satellite TollKit" (STK) contains a total of M spacecraft, then M×(360/N) ³ three-dimensional geometric images can be obtained to obtain a spacecraft image database, which includes It collects multi-attitude, multi-feature parts and other information, provides a rich database for accurate identification of spacecraft feature parts, and further provides basic data for improving the recognition effect.

(2) Perform a weighted average of the pixel values of the three color channels of red (R), green (G), and blue (B) for each three-dimensional geometric image in the spacecraft image database to obtain the grayscale of the three-dimensional geometric image. image, (marked as subset 1) and the grayscale image of the noisy three-dimensional geometric image (marked as subset 2), the preferred solution is specifically:

(2-1) As shown in Figure 2, a weighted average of the pixel values of the three color channels of red (R), green (G), and blue (B) for each three-dimensional geometric image in the spacecraft image database can be obtained. Obtain the grayscale image of the three-dimensional geometric image; the preferred formula is as follows:

G(i,j)＝0.299×R(i,j)+0.578×G(i,j)+0.114×B(i,j)

In the formula: i, j are the horizontal and vertical coordinates of the image, i is greater than or equal to 1, j is greater than or equal to 1, R(i,j), G(i,j), B(i,j) are the i row j in the image respectively. The pixel values corresponding to the column pixels red (R), green (G), and blue (B). G(i,j) is the pixel value of row i and column j in the image. Through the combination of the Chinese coefficients of the above formula, the grayscale image obtained further improves the recognition effect.

Through the grayscale processing in the above step (2-1), M × (360/N) ³ grayscale images of the three-dimensional geometric image of the spacecraft can be obtained, that is, all the images in subset 1 can be obtained.

(2-2) Considering factors such as noise, jitter and blur that exist when imaging spacecraft in the space environment, in step (2-1), based on the grayscale image of the three-dimensional geometric image, add Gaussian noise for blurring, The grayscale image with noise is obtained, that is, subset 2. The preferred expression is specifically:

G _noise (i,j)＝G(i,j)+G _GB (i,j)

In the formula: i,j are the horizontal and vertical coordinates of the grayscale image, and G _noise (i,j) is the pixel value of row i and column j in the grayscale image with noise. σ is the standard deviation of the Gaussian noise normal distribution; r is the blur radius of the noisy grayscale image.

After adding noise, the noisy grayscale image of the spacecraft is obtained, which can more realistically simulate the image of the spacecraft in the space environment, provide accurate images for neural network training and feature part identification, and further improve the identification of feature parts. robustness and accuracy.

(3) Use the cubic interpolation method to reduce the grayscale image of each three-dimensional geometric image in the spacecraft image database to obtain the corresponding interpolated image; the preferred solution is as follows:

f(i+u,j+v)＝ABC

In the formula: f(i+u,j+v) is the pixel value of row i and column j of the interpolated image, u, v are the interpolation intervals, A, B, C are coefficient matrices. The preferred solution is as follows:

A＝[S(1+u) S(u) S(1-u) S(2-u)]

C＝[S(1+v) S(v) S(1-v) S(2-v)] ^T

In the formula: S(·) is the interpolation correlation function, specifically:

The interpolated image obtained through step (3) can realistically simulate images of a spacecraft imaged without distance in a space environment, providing a more accurate training data set for neural network recognition.

(4) Rotate the grayscale image of each three-dimensional geometric image in subset 1 in the spacecraft image database in step (2) to obtain the rotated grayscale image; the preferred solution is specifically:

Assume that the coordinates of a certain pixel in the original image (that is, the grayscale image of the three-dimensional geometric image) are (i, j), and the coordinates in the rotated grayscale image after rotation are (i ₁ , j ₁ ), then the optimal rotation The matrix expression of the transformation is:

The preferred matrix expression of the inverse transformation is:

In the formula, θ represents the angle of rotation around the center point of the original image (that is, the grayscale image of the three-dimensional geometric image) when changing from the original image to the rotated grayscale image in the plane where the original image is located;

a and b are the coordinates of the rotation center when the image is not rotated, which is the geometric center point of the spacecraft's position in the image. c and d are the coordinates of the center point after the image is rotated. There is both rotation and translation in the rotated center point and the rotated center point. Further preferred solutions are as follows:

In the formula, w ₀ =600 is the width of the original image, h ₀ =1000 is the length of the original image; w ₁ =600 is the width of the rotated grayscale image, h ₁ =360 is the length of the rotated grayscale image; is Really simulates the constraint influence of the size (length and width) of the image of the optical camera. The size of the image of the rotated spacecraft in the optical camera changes. Therefore, the length h ₁ and width w ₁ of the rotated image have the same length h as the original image. ₀ and the width w ₀ changes.

Through the above optimized formula and parameter requirements of this step, the accuracy of identification of characteristic parts can be further improved.

(5) Interpolate the grayscale images of subset 1 and subset 2 of each three-dimensional geometric image in the spacecraft image database in step (2), the grayscale image of each three-dimensional geometric image in step (3), and the interpolated image in step (4) After rotating the grayscale image of each three-dimensional geometric image, the contours of the characteristic parts of the spacecraft are marked respectively; and the names of the marked contours of the characteristic parts of the spacecraft are set respectively to obtain the labels of the characteristic parts and complete the characteristic parts of the target spacecraft. annotation; taking the spacecraft solar panel as an example, the preferred solution is as follows:

(5-1) Eliminate images in which the location of the solar panel’s characteristic components cannot be determined;

(5-2) Use a matrix box to delineate and mark the outline of the spacecraft solar panel, and record the area delineated by the rectangular box as the s1 label.

(5-3) Label the pictures where the spacecraft solar panel is blocked, use polylines to delineate and mark the outline of the solar panel, and record the delineated part as the s1 label.

Through the image specifically processed in this step above, the accuracy of identifying the characteristic parts can be further improved.

(6) Construct a convolutional neural network, and use most of the images with contour annotations of the characteristic parts of the spacecraft in the spacecraft image database to train the constructed convolutional neural network; obtain the trained convolutional neural network; combine the rest In the spacecraft image database, the image with the outline of the characteristic parts of the spacecraft marked is input into the trained convolutional neural network to identify the characteristic parts of the spacecraft, and the recognition results are output to realize the intelligent identification of the characteristic parts of the spacecraft. The preferred solution is as follows:

(6-1) The loss function for edge recognition of feature parts is constructed as L ₂ , which is preferably expressed as

Among them, y is the area of the area marked by each feature part of the image in the spacecraft image database. It is the area of the characteristic parts of the spacecraft identified by the neural network. M ₂ is the weight coefficient (the M ₂ is selected as 1).

(6-2) Construct an optimal neural convolution network. Design the neural convolution network to have a total of K channels. The weight coefficient of each channel is W _j , 1≤j≤K (the K is preferably 10000). The weight coefficient of each channel is initially W _j (0) (the W _j (0) is preferably 0.001).

(6-3) Use most of the images with contours of characteristic parts of the spacecraft in the spacecraft image database to train the neural network. Output all annotated images in the spacecraft image database to the constructed neural convolution network. The area y of the area where the spacecraft characteristic part is located in the annotated image can be obtained through the label. The area of the spacecraft’s characteristic parts is calculated through the weight coefficient W _j (0) of each channel of the neural network.

(6-4) Calculate the loss function for edge recognition of feature parts in (6-1) as L ₂ .

(6-5) Determine whether L ₂ ≤ L _2min (the L _2min is preferably 0.16) is established. If established, proceed to (6-7); otherwise, proceed to (6-6);

(6-6) Update the weight coefficient of each channel of the neural convolution network: W _j (l+1)=W _j (l)*dt. (dt is the design coefficient. l is the number of steps for neural network training, satisfying 1≤l≤l _max ). Return to step (6-3) for iterative calculation. As shown in Figure 3, through multiple iterations of training, the edge loss function can be converged to within 0.16.

(6-7) Solidify the weight coefficient W _j (l) of each channel of the neural network parameters to obtain the trained neural network.

(6-8) Input the remaining images with contours of the spacecraft's characteristic parts in the spacecraft image database to obtain the trained neural network, calculate the area where the spacecraft's characteristic parts are located, and output the recognition results. As shown in Figure 4, through the identification of the above preferred convolutional neural network, the location of the spacecraft sailboard can be further accurately identified, and the intelligent identification of the characteristic parts of the spacecraft can be realized.

The present invention integrates feature extraction and feature recognition, avoids manual design of features, has stronger versatility, and greatly improves the recognition efficiency and accuracy of spatial target feature components. This invention is based on satellite image data from multiple sources, including two-dimensional images projected by satellite three-dimensional models, etc., and considers the actual operating environment of the satellite in orbit to simulate and render the satellite image. In this way, not only can more realistic satellite image data be obtained, At the same time, it can increase the amount of satellite image data, which is more conducive to the training of convolutional networks.

Moreover, the present invention does not rely on known feature points in orbit. It still has strong identification ability for spacecraft without feature points, has stronger versatility and robust identification capabilities, and greatly improves the characteristics of spacecraft. Recognition efficiency and accuracy of parts, such as windsurfing boards.

Contents not described in detail in the specification of the present invention are well-known technologies to those skilled in the art.

Claims (10)

1. The intelligent characteristic part identifying method is characterized by comprising the following steps:

(1) Based on an attitude rotation method, constructing a spacecraft image database in space, and storing three-dimensional geometric images of the spacecraft under the condition of different three-axis attitude angle combinations in the spacecraft image database;

(2) The pixel values of three color channels of red R, green G and blue B of each three-dimensional geometric image in the spacecraft image database are weighted and averaged, so that a gray image of the three-dimensional geometric image can be obtained; adding measurement noise energy on the basis of the gray image to obtain a noisy gray image;

(3) Performing reduction processing on the gray level image of each three-dimensional geometric image in the spacecraft image database by adopting a cubic interpolation method to obtain a corresponding interpolated image;

(4) Carrying out rotation change on the gray level image of each three-dimensional geometric image in the spacecraft image database in the step (2) to obtain a rotated gray level image;

(5) Respectively labeling the gray level images of all three-dimensional geometric images in the spacecraft image database in the step (2), the gray level images with noise, the gray level images of all three-dimensional geometric images in the step (3), and the gray level images of all three-dimensional geometric images in the step (4) after the gray level images of all three-dimensional geometric images are rotated; respectively performing name setting on the marked spacecraft characteristic part outlines to obtain labels of characteristic parts, and finishing marking the target spacecraft characteristic parts;

(6) Constructing a convolutional neural network, and training the constructed convolutional neural network by utilizing most of images marked by the outline of the characteristic part of the spacecraft in the spacecraft image database to obtain a trained convolutional neural network; and inputting the images with the marked outlines of the characteristic parts of the spacecraft in the rest spacecraft image database, inputting the trained convolutional neural network to identify the characteristic parts of the spacecraft, outputting the identified results, and realizing the intelligent identification of the characteristic parts of the spacecraft.

2. The intelligent feature recognition method according to claim 1, wherein: (1) Based on an attitude rotation method, constructing a spacecraft image database in space, and storing three-dimensional geometric images of the spacecraft under the condition of different three-axis attitude angle combinations in the spacecraft image database, wherein the three-dimensional geometric images specifically comprise:

(1) Based on the gesture rotation method, a spacecraft image database in the space is constructed, namely, a three-axis gesture angle of the target spacecraft is set as a rolling anglePitch angle θ, yaw angle ψ; the transformation range of the three-axis attitude angle is [0, 360 ]]And taking values every N degrees to form three-axis attitude angle combinations of a plurality of groups of spacecraft, importing simulation modeling software, and acquiring three-dimensional geometric images of the spacecraft under the condition of different three-axis attitude angle combinations to obtain a spacecraft image database.

3. The intelligent feature recognition method according to claim 1, wherein: (2) The pixel values of three color channels of red R, green G and blue B of each three-dimensional geometric image in the spacecraft image database are weighted and averaged, so that a gray image of the three-dimensional geometric image can be obtained, specifically:

G(i,j)＝0.299×R(i,j)+0.578×G(i,j)+0.114×B(i,j)

wherein: i, j is the abscissa of the gray level image, i is greater than or equal to 1, j is greater than or equal to 1, R (i, j), G (i, j) and B (i, j) are pixel values corresponding to the red R, green G and blue B pixels of the i rows and j columns in the gray level image respectively; g (i, j) is the pixel value of row i and column j in the gray scale image.

4. The intelligent feature recognition method according to claim 1, wherein: (3) Performing reduction processing on the gray level image of each three-dimensional geometric image in the spacecraft image database by adopting a cubic interpolation method to obtain a corresponding interpolated image; the method comprises the following steps:

f(i+u,j+v)＝ABC

wherein: f (i+u, j+v) is the pixel value of the i row and j column of the interpolated image, u, v is the interpolation interval, A, B and C is the coefficient matrix, and the form is as follows:

A＝[S(1+u) S(u) S(1-u) S(2-u)]

C＝[S(1+v) S(v) S(1-v) S(2-v)] ^T

wherein: g (i, j) is the pixel value of i rows and j columns in the gray scale image, S is the interpolation correlation function, specifically

5. The intelligent feature recognition method according to claim 1, wherein: (1) Based on an attitude rotation method, constructing a spacecraft image database in space, and storing three-dimensional geometric images of the spacecraft under the condition of different three-axis attitude angle combinations in the spacecraft image database, wherein the three-dimensional geometric images specifically comprise:

the target spacecraft is M kinds of spacecraft in space, three-axis attitude angle combinations of a plurality of groups of target spacecraft are imported into STK software, three-dimensional geometric images of the M kinds of spacecraft under different three-axis attitude angle combinations are acquired in the STK software, and the three-dimensional geometric images are stored in a spacecraft image database in space.

6. The method for intelligently identifying the characteristic parts according to claim 1, wherein noise is added after the gray level image of the three-dimensional geometric image is obtained in the step (2) to simulate noise generated when the spacecraft is shot by adopting real shooting equipment, and the noise is used as the gray level image finally obtained in the step (2); on the basis of the gray level image of the three-dimensional geometric image, adding measurement noise to obtain the gray level image with noise, specifically: gaussian noise is added for blurring.

7. The intelligent feature recognition method according to claim 1, wherein: training the constructed convolutional neural network specifically comprises the following steps: the convolutional neural network receives the image data and inputs the image data, outputs the label of the characteristic part of the image, inputs the predicted label and the label of the characteristic part of the step (5) into a set label loss function together as a predicted label, calculates a training error through the loss function, dynamically adjusts the parameters of the neural network according to the training error, realizes error convergence, and obtains the trained convolutional neural network after the error converges to the set requirement.

8. The intelligent feature recognition method according to claim 1, wherein: in the step (1), in the combination of different three-axis attitude angles, the transformation range of the three-axis attitude angles is [0, 360 degrees ], the value is taken every N degrees, and N is a common divisor of 360.

9. The intelligent feature recognition method according to claim 8, wherein: step (1) combining different three-axis attitude angles to form (360/N) ³ And combining three-axis attitude angles of the group target spacecraft.

10. The intelligent feature recognition method according to claim 1, wherein: (1) The three-dimensional geometric images of the spacecraft stored in the spacecraft image database under the combination condition of different three-axis attitude angles are specifically as follows:

the three-axis attitude angle combinations of the plurality of groups of target spacecrafts are imported into simulation modeling software, wherein the simulation modeling software is satellite system tool kit software, and can construct a three-dimensional model of the spacecrafts, so that three-dimensional geometric images of the spacecrafts under the condition of different three-axis attitude angle combinations are obtained.