Intelligent recognition method of space target components based on deep learning internal and external dual feedback

Abstract

The present invention discloses a method for intelligent recognition of space target components based on deep learning internal and external dual feedback, including 1) an internal feedback network constructed based on space target intelligent recognition and loss function; 2) an external feedback network constructed based on principal component analysis and BP neural network for space target image quality evaluation and image quality improvement. Among them, the internal feedback intelligent recognition is based on a 21-layer deep convolutional neural network built on the basis of the VGG16 neural network, and the neural network receptive field is further expanded in the network to improve the accuracy of target recognition. In the external feedback link, the image quality evaluation adopts 18 non-reference image quality evaluation indicators in four categories commonly used for space images, and a relationship model between the image quality of space targets and the accuracy of component recognition is constructed through the BP neural network. Image quality improvement is carried out for sensitive items that affect the accuracy of component recognition, so as to achieve the improvement of the accuracy of target component recognition under the elimination of the inherent "low-quality" influencing factors of the space environment.

Description

Technical Field

[0001] The present invention relates to the technical field of deep learning target recognition, and in particular to a method for intelligently recognizing space target components based on deep learning internal and external dual feedback. Background Art

[0002] With the rapid development of aerospace technology, traditional manual ground interpretation and command control can no longer meet the development trend of diversity, massive information, and real-time response of spacecraft on-orbit service missions. Space cameras and other detection payloads have changed from traditional ground-based manual remote control to providing information service applications such as autonomous identification, interpretation, and decision-making on board. As the primary link in the autonomous processing of information for control-type aircraft, intelligent identification of space target components has become the key to determining the success or failure of high-real-time control service missions on orbit. The accuracy of space target recognition is largely affected by the quality of the target image. During the on-orbit operation of the camera, the volatilization of lens materials, structural deformation, camera defocusing, high and low frequency vibration of the platform, and the rapid changes in space lighting conditions will cause the camera imaging quality to decline to varying degrees, resulting in a decrease in the accuracy of space target component recognition, or even the inability to complete component recognition and interpretation, which in turn affects the completion of subsequent on-orbit missions. There are relatively few studies in China on improving the image quality of space target images by combining environmental factors affecting space targets to adapt to the intelligent recognition algorithm and build an external feedback mechanism for intelligent recognition. Most of the research on intelligent recognition of space targets builds an internal feedback mechanism based on a loss function within the intelligent recognition algorithm, accepts "low-quality" images as input, and on this basis obtains as much target feature information as possible by increasing the sensitivity of target feature extraction. This process does not eliminate the "low-quality" influence unique to space, resulting in poor improvement in the recognition accuracy of space targets.

[0003] In order to provide accurate information support for subsequent on-orbit mission planning and verification in complex space environments and improve the efficiency of on-orbit space services, it is necessary to conduct research on target component recognition methods that are adapted to space imaging environments. Aiming at the actual engineering application needs of intelligent recognition of space targets, this paper studies a method for intelligent recognition of space target components based on deep learning internal and external dual feedback. Currently, no description or report of related technologies similar to this invention has been found, and no similar materials at home and abroad have been collected. Summary of the invention

[0004] The purpose of the present invention is to provide a method for intelligent recognition of space target components based on deep learning internal and external dual feedback, so as to solve the problem of decreased accuracy of intelligent recognition of space target components caused by complex lighting conditions and poor imaging quality of space cameras.

[0005] In order to solve the above technical problems, the technical solution of the present invention is to provide a method for intelligent recognition of space target components based on deep learning internal and external dual feedback, comprising the following steps:

[0006] S1. Construct an internal feedback network based on convolutional neural network and loss function to achieve preliminary detection of space target components and obtain the recognition accuracy of space target components;

[0007] S2. Construct an external feedback network for space target image quality evaluation and image quality improvement based on principal component analysis and BP neural network, and send the image with improved image quality back to the internal feedback network based on convolutional neural network and loss function in step S1 for training and recognition, so as to improve the recognition accuracy of space target components and broaden the environmental adaptability and recognition robustness of the on-orbit intelligent recognition algorithm.

[0008] Furthermore, the step S1 includes:

[0009] S1-1, using an image scaler at the input end of the neural network to scale the training image size to a uniform size, and normalizing and adjusting the corresponding component label parameters to achieve the adaptability of the neural network algorithm to any input training sample;

[0010] S1-2. Build a 21-layer deep convolutional neural network based on the space target sample training library, which includes 9 convolutional layers, 9 pooling layers, and 3 fully connected layers;

[0011] In each convolutional layer, the input image is convolved with a nonlinear filter, the preset bias weights are superimposed, and the feature map of the layer is obtained through a nonlinear activation function; each convolutional layer is followed by a pooling layer for further use; every three convolutional layers and three pooling layers are followed by a fully connected layer for target feature classification;

[0012] S1-3. Construct a loss function for the characteristics of space target components, including: target positioning offset loss L loc (l, g), target confidence loss L conf (o, c) and target classification loss L cla (o, c):

[0013] L(O,o,C,c,l,g)=λ1L conf (o,c)+λ2L cla (O,C)+λ3L loc (l, g) (1)

[0014] Among them, λ1, λ2, and λ3 are weight coefficients. By constructing an internal feedback network based on a convolutional neural network and a loss function through the above steps, the preliminary detection of space target components can be achieved and the recognition accuracy of space target components can be obtained.

[0015] Furthermore, the step S2 includes:

[0016] S2-1. Construct a space target image quality evaluation parameter library, extract 18 non-reference image quality evaluation parameters as the image quality evaluation index library, which are grayscale parameters, texture parameters, edge parameters, and other parameters in order; use the image quality evaluation parameters to calculate the image quality of the space target sample set participating in deep learning training and the test set images participating in intelligent recognition as the input of image quality correlation analysis;

[0017] S2-2, using the principal component analysis method to perform decorrelation processing on the space target image quality evaluation parameters to achieve parameter optimization and data dimension reduction, and obtain the non-correlated evaluation polynomial containing the 18 image quality evaluation parameters in step S2-1:

[0018] Let X1, X2, ..., X p is the space target image quality evaluation parameter value, recorded as X = (X1, X2, ..., X p ), the covariance matrix can be expressed as:

[0019] ∑=(σ ij ) p =E[(XE(X))(XE(X)) T ] (2)

[0020] The above formula is a p-order non-negative matrix; for x1, x2, ..., x p Perform linear transformation to obtain uncorrelated variables Y1, Y2, ..., Y that can meet the above requirements p , whose expression is as follows:

[0021]

[0022] In the formula, I i =(I i1 , I i2 , ..., I ip )(i=1,2,...p)——constant vector;

[0023] S2-3, using BP neural network to construct a description of the relationship between target image quality parameters and target intelligent recognition accuracy, and determine the sensitive items in image quality that affect component recognition accuracy;

[0024] S2-4. For sensitive items in image quality that affect the accuracy of component recognition, the modulation transfer function inverse convolution and grayscale nonlinear stretching enhancement method generated by the full-link frequency domain simulation model of the space optical camera are used to improve the high-frequency information and spatial contrast in the target frequency domain.

[0025] Furthermore, in step S2-1, the grayscale parameters include grayscale mean, mean square error, radiation accuracy steepness, and grayscale entropy; the texture parameters include angular second-order moment, texture entropy, contrast, autocorrelation, and inverse difference; the edge parameters include edge energy and detail energy; and the other parameters include small gradient advantage, large gradient advantage, gradient distribution unevenness, gradient average, and gradient variance.

[0026] Furthermore, in step S2-1, the grayscale mean reflects the brightness of the image as a whole, and is a quantitative reflection of the actual radiation energy of the scene objects. Its basic expression is:

[0027]

[0028] Where m, n are the height and width of the image, and p(i, j) is the grayscale value of a point in the image;

[0029] The radiation accuracy steepness reflects the richness of the grayscale levels of the image, and its calculation formula is:

[0030] K=∑(i-μ) 4 P(i) / σ 8 (6)

[0031] In the formula, i is the pixel gray value, μ is the mean, σ 2 is the variance, P(i) is the distribution probability of gray level i in the image;

[0032] Entropy is a measure of the richness of image information from the perspective of information theory, which can be expressed as:

[0033]

[0034] Where L is the maximum gray level of the image, P i is the distribution probability of the gray value of the pixel i on the image;

[0035] The angular second-order moment is an indicator to measure the consistency of texture. Its basic expression is:

[0036]

[0037] Where P(i, j) is the statistical number of grayscale pairs;

[0038] The contrast of an image is the ratio or logarithmic difference between the density of the brightest and darkest parts of the image. Its basic expression is:

[0039]

[0040] Where |ij|=n;

[0041] Image autocorrelation is a measure of the linearity of image grayscale, which can reflect the spatial arrangement of objects. Its basic expression is:

[0042]

[0043] Where:

[0044]

[0045] The inverse moment describes the local homogeneity of the image and reflects the contrast of the local texture direction:

[0046]

[0047] The average gradient can sensitively reflect the image's ability to express the contrast of tiny details. Its calculation formula is:

[0048]

[0049] In the formula, and are the grayscale of the image points and their gradients in the row and column directions respectively;

[0050] Detail energy describes the richness of the detail edges of an image from the local part of the image, and its calculation formula is:

[0051]

[0052] in

[0053]

[0054] In the formula, m f (x, y) is the average brightness of the image;

[0055] Suppose the original image is f(x, y), x = 1, 2, ..., N, and its gray level is L; discretize the gradient image into gray levels, and set the number of gray levels to L g , the new grayscale is in The grayscale-gradient co-occurrence matrix is ​​{H ij ,i=0,1,...L-1,j=0,1,...L g -1}, where H ij is defined as the number of elements in the set {(x, y)|f(x, y)=i, G(x, y)=j); After normalization, we get

[0056] The small gradient advantage is defined as The large gradient advantage is defined as The gradient distribution inhomogeneity is defined as The gradient average is defined as The gradient variance is defined as

[0057] Furthermore, in step S2-2:

[0058] For typical space target simulation image sources, the principal component analysis method is used to optimize the 18 image quality evaluation parameters, and the first, second, and third principal components with cumulative contribution rates greater than 90% are selected as effective inputs for subsequent analysis of the relationship between image quality and tracking accuracy. The expressions of the first, second, and third principal components of the image are obtained as follows:

[0059] Y1=0.89X1-0.17X2-0.50X3+0.66X4+0.97X5-0.66X6-0.99X7+0.99X8+0.50X9-0.79X 10 -0.97X 11 +0.86X 12 -0.99X 13 +0.98X 14 -0.99X 15 +0.98X 16 +0.96X 17 +0.96X 18 (18)

[0060] Y2=-0.71X1-0.88X2-0.52X3-0.06X4-0.37X5-0.21X6+0.33X7-0.52X8-0.11X9-0.13X 10 +0.1X 11 -0.16X 12 +0.83X 13 -0.04X 14 +0.06X 15 -0.04X 16 +0.08X 17 -0.07X 18 (19)

[0061] Y3=0.39X1-0.33X2-0.27X3-0.66X4-0.51X5+0.23X6+0.009X7-0.012X8-0.14X9+0.06X 10 +0.0095X 11 -0.03X 12 +0.01X 13 -0.09X 14 +0.016X 15 -0.09X 16 -0.21X17 +0.05X 18 (20)

[0062] Furthermore, in step S2-4:

[0063] The phase plane translation caused by the change in the distance between the fixed-focus optical camera and the target is

[0064]

[0065] Where f is the focal length, s1 is the object distance before the target distance changes, s2 is the object distance after the target distance changes, s1 is the image distance before the target distance changes, s2 is the image distance after the target distance changes;

[0066] The image shift modulation transfer function of linear motion is expressed as:

[0067]

[0068] Where S linear =vTf / h, V is the speed of the camera relative to the target; T is the camera integration time; f is the focal length of the optical system; h is the altitude of the satellite, f v is the spatial frequency;

[0069] The modulation transfer function of the effect of high-frequency vibration of the satellite platform on image quality is expressed as:

[0070] MTF hf =J0(2πf v A) (23)

[0071] Where f v is the spatial frequency; J0 is the zero-order Bessel function; A is the maximum image amplitude; θ is the amplitude angle; f is the focal length of the camera, A = fθ;

[0072] The modulation transfer function of the random jitter of the satellite platform on the image quality is expressed as:

[0073] MTF gauss =exp(-2π 2 σ 2 f v 2 ) (twenty four)

[0074] Where σ is the standard deviation of displacement; f v is the spatial frequency;

[0075] The modulation transfer function of the effect of the receiving surface of the area array CCD deviating from the focal plane of the optical system on the image quality can be expressed as:

[0076]

[0077] Where J1 is the first-order Bessel function; Δf is the axial defocus; NA = D / (2f) = 1 / (2F); D is the incident pupil diameter; f v is the spatial frequency;

[0078] The modulation transfer function of the effect of CCD pixel geometry on image quality can be expressed as:

[0079]

[0080] Where d is the pixel size; f k is the spatial frequency of the output signal;

[0081] The MTF expression due to charge transfer loss is:

[0082]

[0083] Where n is the number of CCD pixels; f is the spatial frequency of the output signal; f N is the Nyquist frequency; ε is the charge transfer loss rate, which is the ratio of the charge transferred to the next potential well to the charge in the original potential well.

[0084] The beneficial effects of the space target component intelligent recognition method based on deep learning internal and external dual feedback provided by the present invention are:

[0085] By constructing an internal feedback network based on convolutional neural networks and loss functions, the preliminary detection of space target components can be achieved and the recognition accuracy of space target components can be obtained. By constructing an external feedback network for space target image quality evaluation and image quality improvement based on principal component analysis and BP neural network, and sending the image with improved image quality back to the internal feedback network for training and recognition, the recognition of optical cameras can be further achieved, and the recognition accuracy of typical small-size components of satellites (radar antennas, optical lenses) can be improved by more than 20%, effectively broadening the environmental adaptability and recognition robustness of the on-orbit intelligent recognition algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] The invention will be further described below in conjunction with the accompanying drawings:

[0087] Figure 1 It is a schematic diagram of intelligent recognition of space target components based on double feedback neural network;

[0088] Figure 2 It is the workflow diagram of BP neural network;

[0089] Figure 3 This is a schematic diagram of the simulation based on optical full-link imaging;

[0090] Figure 4 This is a diagram showing the influence of the imaging distance of the optical system on the imaging results. DETAILED DESCRIPTION

[0091] The following is a further detailed description of the space target component intelligent recognition method based on deep learning internal and external dual feedback proposed by the present invention in conjunction with the accompanying drawings and specific embodiments. The advantages and features of the present invention will become clearer according to the following description and claims. It should be noted that the accompanying drawings are all in a very simplified form and use non-precise ratios, which are only used to conveniently and clearly assist in explaining the purpose of the embodiments of the present invention.

[0092] The core idea of ​​the present invention is that the portable feeler gauge automatic online calibration device provided by the present invention is simple in equipment and highly practical, which simplifies the workload of operators in traditional measurements and improves measurement efficiency and measurement quality; using the portable feeler gauge automatic online calibration device provided by the present invention, the entire calibration process is easy to operate and the data is accurate.

[0093] In order to solve the problem of decreased accuracy of intelligent recognition of space target components caused by complex lighting conditions and poor imaging quality of space cameras, the purpose of the present invention is to propose a method for intelligent recognition of space target components based on deep learning internal and external dual feedback. The method of the present invention has two parts, namely 1) an internal feedback network constructed based on intelligent recognition of space targets and loss function, and 2) an external feedback network constructed based on principal component analysis, BP neural network for space target image quality evaluation and image quality improvement. The specific steps are as follows:

[0094] The problem of decreased accuracy in intelligent recognition of space target components caused by decreased image quality due to overexposure, underexposure, defocusing, high-frequency vibration, low-frequency vibration, and linear motion smearing of space cameras under complex lighting conditions solved by the present invention is a space target dataset constructed based on space optical full-link imaging simulation. The dataset contains 150 space targets and 1.5 million labeled sample sets. The annotated satellite parts include five types of components: satellite body, solar sail panel, docking ring, engine nozzle, and radar antenna. The images in the dataset contain a series of degraded images of different sizes caused by overexposure, underexposure, defocusing, high-frequency vibration, low-frequency vibration, and linear motion smearing of optical cameras.

[0095] Step 1: Use the improved nearest neighbor algorithm image resizer at the input end of the neural network to scale the training image size to a uniform size, with optional 512×512, 1024×1024, and 2048×2048. Normalize and adjust the corresponding component label parameters to achieve the adaptability of the neural network algorithm to any input training sample.

[0096] Step 2: Based on the classic VGG16 network, a 21-layer deep convolutional neural network is constructed, which includes 9 convolutional layers, 9 pooling layers, and 3 fully connected layers. In each convolutional layer, the input image is convolved with a nonlinear filter, the preset bias weights are superimposed, and the feature map of the layer is obtained through a nonlinear activation function. Each convolutional layer is followed by a pooling layer for the next step. After every three convolutional layers and three pooling layers, a fully connected layer is used to classify the target features.

[0097] Step 3: Construct a loss function for the characteristics of space target components, including: target positioning offset loss L loc (l, g), target confidence loss L conf (o, c) and target classification loss L cla (O, C).

[0098] L(O,o,C,c,l,g)=λ1L conf (o,c)+λ2L cla (O,C)+λ3L loc (l, g) (4)

[0099] Among them, λ1, λ2, and λ3 are weight coefficients, and the weight ratios of the three are 0.2, 0.5, and 0.3 respectively. By constructing an internal feedback network based on a convolutional neural network and a loss function through the above steps, the preliminary detection of space target components can be achieved, and the recognition accuracy of space target components can be obtained.

[0100] Step 4: Construct a space target image quality evaluation parameter library, and extract 18 non-reference image quality evaluation parameters as the image quality evaluation index library, which are grayscale parameters (grayscale mean, mean square error, radiation accuracy steepness, grayscale entropy), texture parameters (angular second-order moment, texture entropy, contrast, autocorrelation, inverse difference), edge parameters (edge ​​energy, detail energy), and other parameters (small gradient advantage, large gradient advantage, gradient distribution unevenness, gradient average, gradient variance).

[0101] The grayscale mean reflects the brightness of the image as a whole and is a quantitative reflection of the actual radiation energy of the scene objects. Its basic expression is:

[0102]

[0103] Where m and n are the height and width of the image, and p(i, j) is the gray value of a point in the image. The radiation accuracy steepness reflects the richness of the gray level of the image. The calculation formula is:

[0104] K=∑(i-μ) 4 P(i) / σ 8 (6)

[0105] In the formula, i is the pixel gray value, μ is the mean, σ 2 is the variance, and P(i) is the distribution probability of gray level i in the image.

[0106] Entropy is a measure of the richness of image information from the perspective of information theory, which can be expressed as:

[0107]

[0108] Where L is the maximum gray level of the image, P i is the distribution probability of the gray value of pixel i in the image.

[0109] The angular second-order moment is an indicator to measure the consistency (local stability) of the texture. Its basic expression is:

[0110]

[0111] Where P(i, j) is the statistical number of grayscale pairs.

[0112] The contrast of an image is the ratio or logarithmic difference between the density of the brightest and darkest parts of the image. Its basic expression is:

[0113]

[0114] In the formula, |ij|=n.

[0115] Image autocorrelation is a measure of the linearity of image grayscale, which can reflect the spatial arrangement of objects. Its basic expression is:

[0116]

[0117] Where:

[0118]

[0119] The inverse moment describes the local homogeneity of the image (reflecting the contrast of the local direction of the texture)

[0120]

[0121] The average gradient can sensitively reflect the image's ability to express the contrast of tiny details. Its calculation formula is:

[0122]

[0123] In the formula, and They are the grayscale of the image points and their gradients in the row and column directions respectively.

[0124] Detail energy describes the richness of the detail edges of an image from the local part of the image. Its calculation formula is:

[0125]

[0126] in

[0127]

[0128] In the formula, m f (x, y) is the average brightness of the image.

[0129] Let the original image be f(x, y), x = 1, 2, ..., N, and its gray level be L. Discretize the gradient image into gray levels, and let the number of gray levels be L g , the new grayscale is in The grayscale-gradient co-occurrence matrix is ​​{H ij ,i=0,1,...L-1,j=0,1,...L g -1}, where H ij is defined as the number of elements in the set {(x, y)|f(x, y)=i, G(x, y)=j}. After normalization, we get

[0130] The small gradient advantage is defined as The large gradient advantage is defined as The gradient distribution inhomogeneity is defined as The gradient average is defined as The gradient variance is defined as

[0131] Image quality evaluation parameters are used to calculate the image quality of the spatial target sample set participating in deep learning training and the test set images participating in intelligent recognition as the input of image quality correlation analysis.

[0132] Step 5: Use the principal component analysis method to decorrelate the space target image quality evaluation parameters to achieve parameter optimization and data dimensionality reduction, and obtain an uncorrelated evaluation polynomial containing the 18 image quality evaluation parameters in step 4.

[0133] Let X1, X2, ..., X p is the space target image quality evaluation parameter value, recorded as X = (X1, X2, ..., X p ), the covariance matrix can be expressed as:

[0134] ∑=(σ ij ) p =E[(XE(X))(XE(X)) T ] (16)

[0135] The above formula is a p-order non-negative matrix. p Perform linear transformation to obtain uncorrelated variables Y1, Y2, ..., Y that can meet the above requirements p , whose expression is as follows:

[0136]

[0137] In the formula, I i =(I i1 , I i2 , ..., I ip )(i=1,2,…p)——constant vector.

[0138] For typical space target simulation image sources, the principal component analysis method is used to optimize the 18 image quality evaluation parameters in the previous section, and the first, second, and third principal components with cumulative contribution rates greater than 90% are selected as effective inputs for subsequent analysis of the relationship between image quality and tracking accuracy. The expressions of the first, second, and third principal components of the image are obtained as follows:

[0139] Y1=0.89X1-0.17X2-0.50X3+0.66X4+0.97X5-0.66X6-0.99X7+0.99X8+0.50X9-0.79X 10 -0.97X 11 +0.86X 12 -0.99X 13 +0.98X 14 -0.99X 15 +0.98X 16 +0.96X 17 +0.96X 18 (18)

[0140] Y2=-0.71X1-0.88X2-0.52X3-0.06X4-0.37X5-0.21X6+0.33X7-0.52X8-0.11X9-0.13X 10 +0.1X 11 -0.16X 12 +0.83X 13 -0.04X 14 +0.06X 15 -0.04X 16 +0.08X 17 -0.07X 18 (19)

[0141] Y3=0.39X1-0.33X2-0.27X3-0.66X4-0.51X5+0.23X6+0.009X7-0.012X8-0.14X9+0.06X 10 +0.0095X 11 -0.03X 12 +0.01X 13 -0.09X 14 +0.016X 15 -0.09X 16 -0.21X 17 +0.05X 18 (20)

[0142] Step 6: Use BP neural network to construct the relationship description between target image quality parameters and target intelligent recognition accuracy, and determine the sensitive items in image quality that affect component recognition accuracy. Take the target image quality parameters as input and the target intelligent recognition accuracy as output, calculate the data of each unit in the hidden layer and output layer, compare the square difference between the target value and the actual value and the loss function, and adjust the weight values ​​of the hidden layer and output layer if the requirements are not met until the requirements are met.

[0143] Table 1 The influence of various image quality evaluation parameters on target recognition accuracy

[0144] Image quality evaluation parameters Impact Factor Image quality evaluation parameters Impact Factor Mean square error 0.044 Contrast 0.099 Peak signal-to-noise ratio 0.078 Autocorrelation 0.052 Normalized cross-correlation coefficient 0.041 Reverse Gap 0.040 Edge Energy 0.066 Grayscale entropy 0.042 Detail Energy 0.049 Small gradient advantage 0.040 Radiation accuracy steepness 0.053 Large gradient advantage 0.047 Angular second moment 0.12 Gradient distribution unevenness 0.039 Texture Entropy 0.040 Gradient Averaging 0.047 Grayscale mean 0.033 Gradient Variance 0.032

[0145] Step 7: For sensitive items in image quality that affect the accuracy of component recognition, the modulation transfer function inverse convolution and grayscale nonlinear stretching enhancement method generated by the full-link frequency domain simulation model of the space optical camera are used to improve the high-frequency information and spatial contrast in the target frequency domain.

[0146] Among them, the phase plane translation caused by the change in the distance between the fixed-focus optical camera and the target is

[0147]

[0148] Wherein, f is the focal length, s1 is the object distance before the target distance is changed, s2 is the object distance after the target distance is changed, s1 is the image distance before the target distance is changed, and s2 is the image distance after the target distance is changed.

[0149] The image shift modulation transfer function of linear motion can be expressed as:

[0150]

[0151] Where S linear =vTf / h, v is the speed of the camera relative to the target; T is the camera integration time; f is the focal length of the optical system; h is the altitude of the satellite, f v is the spatial frequency.

[0152] The modulation transfer function of the effect of high-frequency vibration of the satellite platform on image quality can be expressed as:

[0153] MTF hf =J0(2πf v A) (23)

[0154] Where f v is the spatial frequency; J0 is the zero-order Bessel function; A is the maximum image amplitude; θ is the amplitude angle; f is the focal length of the camera, A=fθ.

[0155] The modulation transfer function of the random jitter of the satellite platform on the image quality can be expressed as:

[0156] MTF gauss =exp(-2π 2 σ 2 f v 2 ) (twenty four)

[0157] Where σ is the standard deviation of displacement; f v is the spatial frequency.

[0158] The modulation transfer function of the effect of the receiving surface of the area array CCD deviating from the focal plane of the optical system on the image quality can be expressed as:

[0159]

[0160] Where J1 is the first-order Bessel function; Δf is the axial defocus; NA = D / (2f) = 1 / (2F); D is the incident pupil diameter; f v is the spatial frequency.

[0161] The modulation transfer function of the effect of CCD pixel geometry on image quality can be expressed as:

[0162]

[0163] Where d is the pixel size; f k is the spatial frequency of the output signal.

[0164] The MTF expression due to charge transfer loss is:

[0165]

[0166] Where n is the number of CCD pixels; f is the spatial frequency of the output signal; f N is the Nyquist frequency; ε is the charge transfer loss rate, which is the ratio of the charge transferred to the next potential well to the charge in the original potential well.

[0167] Through the above steps, an external feedback network for space target image quality evaluation and image quality improvement based on principal component analysis and BP neural network is constructed, and the image with improved image quality is sent back to the internal feedback network for training and recognition. The recognition of optical cameras can be further realized, and the recognition accuracy of typical small-size components of satellites (radar antennas, optical lenses) is improved by more than 20%, effectively broadening the environmental adaptability and recognition robustness of the on-orbit intelligent recognition algorithm.

[0168] Table 2 Component recognition accuracy results before and after image quality improvement

[0169]

[0170]

[0171] The contents not described in detail in this specification belong to the prior art known to those skilled in the art. It is obvious to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. Therefore, no matter from which point of view, the embodiments should be regarded as exemplary and non-restrictive, and the scope of the present invention is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention.

Claims

1. A space target component intelligent recognition method based on deep learning internal and external dual feedback, characterized in that: The steps include: S1. Construct an internal feedback network based on a convolutional neural network and a loss function, implement preliminary detection of space target components for space target images, and obtain the recognition accuracy of space target components; S2, constructing an external feedback network for space target image quality evaluation and image quality improvement based on principal component analysis and BP neural network, and sending the image with improved image quality back to the internal feedback network based on convolutional neural network and loss function in step S1 for training and recognition, so as to improve the recognition accuracy of space target components and broaden the environmental adaptability and recognition robustness of the on-orbit intelligent recognition algorithm; The step S2 comprises: S2-1. Construct a space target image quality evaluation parameter library, extract 18 non-reference image quality evaluation parameters as the image quality evaluation index library, which are grayscale parameters, texture parameters, edge parameters, and other parameters in order; use the image quality evaluation parameters to calculate the image quality of the space target sample set participating in deep learning training and the test set images participating in intelligent recognition as the input of image quality correlation analysis; S2-2, using the principal component analysis method to perform decorrelation processing on the space target image quality evaluation parameters to achieve parameter optimization and data dimension reduction, and obtain an uncorrelated evaluation polynomial containing the 18 image quality evaluation parameters in step S2-1; S2-3, using BP neural network to construct a description of the relationship between target image quality parameters and target intelligent recognition accuracy, and determine the sensitive items in image quality that affect component recognition accuracy; S2-4. For sensitive items in image quality that affect the accuracy of component recognition, the modulation transfer function inverse convolution and grayscale nonlinear stretching enhancement method generated by the full-link frequency domain simulation model of the space optical camera are used to improve the high-frequency information and spatial contrast in the target frequency domain.

2. The space target component intelligent recognition method based on deep learning internal and external dual feedback as claimed in claim 1 is characterized in that: The step S1 comprises: S1-1, using an image scaler at the input end of the neural network to scale the training image size to a uniform size, and normalizing and adjusting the corresponding component label parameters to achieve the adaptability of the neural network algorithm to any input training sample; S1-2. Build a 21-layer deep convolutional neural network based on the space target sample training library, which includes 9 convolutional layers, 9 pooling layers, and 3 fully connected layers; In each convolutional layer, the input image is convolved with a nonlinear filter, the preset bias weights are superimposed, and the feature map of the layer is obtained through a nonlinear activation function; each convolutional layer is followed by a pooling layer for further use; every three convolutional layers and three pooling layers are followed by a fully connected layer for target feature classification; S1-3. Construct a loss function for the characteristics of space target components, including: target positioning offset loss L loc (l,g), target confidence loss L conf (o,c) and target classification loss L cla (O,C): L(O,o,C,c,l,g)=λ1L conf (o,c)+λ2L cla (O,C)+λ3L loc (l,g) (1) Among them, λ1, λ2, λ3 are weight coefficients; by constructing an internal feedback network based on a convolutional neural network and a loss function through the above steps, the preliminary detection of space target components can be achieved and the recognition accuracy of space target components can be obtained.

3. The space target component intelligent recognition method based on deep learning internal and external dual feedback as claimed in claim 2 is characterized in that: In the step S2-2: Let X1,X2,...,X p is the space target image quality evaluation parameter value, recorded as X = (X1, X2, ..., X p ), the covariance matrix can be expressed as: ∑=(σ ij ) p =E[(XE(X))(XE(X)) T ] (2) The above formula is a p-order non-negative matrix; for x1,x2,...,x p Perform linear transformation to obtain uncorrelated variables Y1, Y2, ..., Y that can meet the above requirements p , whose expression is as follows: In the formula, l i =(l i1 ,l i2 ,...,l ip )(i=1,2,...p)——constant vector.

4. The space target component intelligent recognition method based on deep learning internal and external dual feedback as claimed in claim 3 is characterized in that: In step S2-1, the grayscale parameters include grayscale mean, mean square error, radiation accuracy steepness, and grayscale entropy; the texture parameters include angular second-order moment, texture entropy, contrast, autocorrelation, and inverse distance; the edge parameters include edge energy and detail energy; and the other parameters include small gradient advantage, large gradient advantage, gradient distribution unevenness, gradient average, and gradient variance.

5. The space target component intelligent recognition method based on deep learning internal and external dual feedback as claimed in claim 4 is characterized in that: In step S2-1, the grayscale mean reflects the brightness of the image as a whole, and is a quantitative reflection of the actual radiation energy of the scene objects. Its basic expression is: Where m, n are the height and width of the image, and p(i, j) is the grayscale value of a point in the image; The radiation accuracy steepness reflects the richness of the grayscale levels of the image, and its calculation formula is: K=∑(i-μ) 4 P(i) / s 8 (6) In the formula, i is the pixel gray value, μ is the mean, σ 2 is the variance, P(i) is the distribution probability of gray level i in the image; Entropy is a measure of the richness of image information from the perspective of information theory, which can be expressed as: Where L is the maximum gray level of the image, P i is the distribution probability of the gray value of the pixel i on the image; The angular second-order moment is an indicator to measure the consistency of texture. Its basic expression is: Where P(i,j) is the statistical number of grayscale pairs; The contrast of an image is the ratio or logarithmic difference between the density of the brightest and darkest parts of the image. Its basic expression is: Where |ij|=n; Image autocorrelation is a measure of the linearity of image grayscale, which can reflect the spatial arrangement of objects. Its basic expression is: Where: The inverse moment describes the local homogeneity of the image and reflects the contrast of the local texture direction: The average gradient can sensitively reflect the image's ability to express tiny detail contrasts, and its calculation formula is: In the formula, and are the grayscale of the image points and their gradients in the row and column directions respectively; Detail energy describes the richness of the detail edges of an image from the local part of the image, and its calculation formula is: in In the formula, m f (x,y) is the average brightness of the image; Suppose the original image is f(x,y), x=1,2,...,N, and its gray level is L; discretize the gradient image into gray levels, and set the number of gray levels to L g , the new grayscale is in The grayscale-gradient co-occurrence matrix is ​​{H ij ,i=0,1,...L-1,j=0,1,...L g -1}, where H ij is defined as the number of elements in the set {(x,y)|f(x,y)=i,G(x,y)=j}; After normalization, we get The small gradient advantage is defined as The large gradient advantage is defined as The gradient distribution inhomogeneity is defined as The gradient average is defined as The gradient variance is defined as 6. The method for intelligent recognition of space target components based on deep learning and internal and external dual feedback as claimed in claim 3, characterized in that: In the step S2-2: For typical space target simulation image sources, the principal component analysis method is used to optimize the 18 image quality evaluation parameters, and the first, second, and third principal components with cumulative contribution rates greater than 90% are selected as effective inputs for subsequent analysis of the relationship between image quality and tracking accuracy. The expressions of the first, second, and third principal components of the image are obtained as follows: <h2 style=";text-align:left;direction:ltr">Y1:0.89X1-0.17X2-0.50X3+0.66X4+0.97X5-0.66X6-0.99X7+0.99X8+0.50X9-0.79X<h2 style=";text-align:left;direction:ltr"> 10 <h2 style=";text-align:left;direction:ltr"> -0.97X<h2 style=";text-align:left;direction:ltr"> 11 <h2 style=";text-align:left;direction:ltr"> +0.86X<h2 style=";text-align:left;direction:ltr"> 12 <h2 style=";text-align:left;direction:ltr"> -0.99X<h2 style=";text-align:left;direction:ltr"> 13 <h2 style=";text-align:left;direction:ltr"> +0.98X<h2 style=";text-align:left;direction:ltr"> 14 <h2 style=";text-align:left;direction:ltr"> -0.99X<h2 style=";text-align:left;direction:ltr"> 15 <h2 style=";text-align:left;direction:ltr"> +0.98X<h2 style=";text-align:left;direction:ltr"> 16 <h2 style=";text-align:left;direction:ltr"> +0.96X<h2 style=";text-align:left;direction:ltr"> 17 <h2 style=";text-align:left;direction:ltr"> +0.96X<h2 style=";text-align:left;direction:ltr"> 18 <h2 style=";text-align:left;direction:ltr"> (18) Y2=-0.71X1-0.88X2-0.52X3-0.06X4-0.37X5-0.21X6+0.33X7-0.52X8-0.11X9-0.13X 10 +0.1X 11 -0.16X 12 +0.83X 13 -0.04X 14 +0.06X 15 -0.04X 16 +0.08X 17 -0.07X 18 (19) <h2 style=";text-align:left;direction:ltr">Y3=0.39X1-0.33X2-0.27X3-0.66X4-0.51X5+0.23X6+0.009X7-0.012X8-0.14X9+0.06X<h2 style=";text-align:left;direction:ltr"> 10 <h2 style=";text-align:left;direction:ltr"> +0.0095X<h2 style=";text-align:left;direction:ltr"> 11 <h2 style=";text-align:left;direction:ltr"> -0.03X<h2 style=";text-align:left;direction:ltr"> 12 <h2 style=";text-align:left;direction:ltr"> +0.01X<h2 style=";text-align:left;direction:ltr"> 13 <h2 style=";text-align:left;direction:ltr"> -0.09X<h2 style=";text-align:left;direction:ltr"> 14 <h2 style=";text-align:left;direction:ltr"> +0.016X<h2 style=";text-align:left;direction:ltr"> 15 <h2 style=";text-align:left;direction:ltr"> -0.09X<h2 style=";text-align:left;direction:ltr"> 16 <h2 style=";text-align:left;direction:ltr"> -0.21X<h2 style=";text-align:left;direction:ltr"> 17 <h2 style=";text-align:left;direction:ltr"> +0.05X<h2 style=";text-align:left;direction:ltr"> 18 <h2 style=";text-align:left;direction:ltr"> (20) 7. The method for intelligent recognition of space target components based on deep learning and internal and external dual feedback as claimed in claim 3, characterized in that: In the step S2-4: The phase plane translation caused by the change in the distance between the fixed-focus optical camera and the target is Where f is the focal length, s1 is the object distance before the target distance changes, s2 is the object distance after the target distance changes, s′1 is the image distance before the target distance changes, and s′2 is the image distance after the target distance changes; The image shift modulation transfer function of linear motion is expressed as: Where S linear =vTf / h, v is the speed of the camera relative to the target; T is the camera integration time; f is the focal length of the optical system; h is the satellite's flying height, f v is the spatial frequency; The modulation transfer function of the effect of high-frequency vibration of the satellite platform on image quality is expressed as: MTF hf =J0(2πf v A) (23) Where f v is the spatial frequency; J0 is the zero-order Bessel function; A is the maximum image amplitude; θ is the amplitude angle; f is the focal length of the camera, A = fθ; The modulation transfer function of the random jitter of the satellite platform on the image quality is expressed as: MTF gauss =exp(-2π 2 s 2 f v 2 ) (24) Where σ is the standard deviation of displacement; f v is the spatial frequency; The modulation transfer function of the effect of the receiving surface of the area array CCD deviating from the focal plane of the optical system on the image quality can be expressed as: Where J1 is the first-order Bessel function; △f is the axial defocus; NA = D / (2f) = 1 / (2F); D is the incident pupil diameter; f v is the spatial frequency; The modulation transfer function of the effect of CCD pixel geometry on image quality can be expressed as: Where d is the pixel size; f k is the spatial frequency of the output signal; The MTF expression due to charge transfer loss is: Where n is the number of CCD pixels; f is the spatial frequency of the output signal; f N is the Nyquist frequency; ε is the charge transfer loss rate, which is the ratio of the charge transferred to the next potential well to the charge in the original potential well.

Images