Multi-mode fusion sea surface target identification method of high-dynamic unmanned platform

Abstract

The invention relates to the technical field of sea surface target multi-mode fusion recognition, in particular to a multi-mode fusion sea surface target recognition method of a high-dynamic unmanned platform. According to the method, a multi-modal fusion recognition framework based on a multi-kernel learning method is provided, and a small-sample multi-modal fusion target recognition algorithm for a sea-surface target is designed on the basis of the multi-modal fusion recognition framework. Target detection and feature extraction are respectively performed according to data characteristics of different sources of visible light images, SAR images and active and passive radar one-dimensional range profile signals, multi-mode heterogeneous features are mapped by using a multi-scale kernel method, weighted fusion of a kernel matrix is performed, global training optimization is performed, and finally a classification result of each sample can be output. The algorithm not only considers the performance of a data-driven target recognition algorithm, but also considers platform hardware computing power limitation and small sample conditions, so that the balance among the performance, computing power and samples can be realized.

Description

Technical Field

[0001] The invention relates to the technical field of multi-mode fusion recognition of sea surface targets, and in particular to a multi-mode fusion sea surface target recognition method for a high-dynamic unmanned platform. Background Art

[0002] The sea surface targets that the present invention is concerned with are slow-moving targets on the sea surface, such as ships, boats, etc., which have a certain radar reflection cross-sectional area or are equipped with radar equipment and radiate electromagnetic signals outward, but other moving targets, such as cruise missiles flying over the sea, are not included in this list. Slow-moving targets on the sea surface are the focus of sea observation by various sensors. Compared with stationary targets such as sea surface buildings and islands, their echo signals in typical radar systems, such as airborne active radar systems, will fall into the main lobe clutter area. The targets are mixed with the main clutter, resulting in defocus or displacement in the radar image, thereby affecting the identification of the targets, making it difficult to detect and locate the targets. These targets are often targets of military interest, such as sea surface ships and boats.

[0003] The high-dynamic unmanned platform described in the present invention refers to an aircraft platform that can achieve rapid acceleration, deceleration or change direction in a short period of time, including missile-borne and airborne platforms such as missiles and high-speed drones with high-speed movement capabilities. Such platforms often have high mobility and flexibility, and are suitable for occasions that require rapid intelligent perception, rapid decision-making response, high mobility and real-time performance.

[0004] There are many ways to obtain information about sea surface targets, including: UAV-borne, missile-borne visible light, infrared, synthetic aperture radar (SAR), active radar, passive radar, etc.

[0005] Based on this, in the prior art, there are several methods for detecting and identifying sea surface targets:

[0006] 1. Sea surface target detection and recognition based on visible light images

[0007] Sea surface target detection based on satellite-borne and airborne visible light remote sensing images has always attracted much attention. It mainly includes three technical systems: target detection based on traditional methods, target detection based on visual saliency, and target detection and recognition based on deep learning.

[0008] Object detection and recognition based on traditional methods can usually be divided into four steps: ① Select candidate target regions from the input image; ② Extract features from the candidate regions; ③ Use a classifier for classification; ④ Use non-maximum suppression to delete intersection regions with low scores.

[0009] Target detection and recognition based on visual saliency mainly focuses on visual attention mechanism and digital image processing technology. Since there is a large texture difference between ships and the sea surface, the texture information of the image will be an important feature for detecting ships on the sea surface, while the classic Itti model only extracts the brightness, color and direction features of the image. Therefore, in order to further improve the target detection effect, based on the traditional Itti model, the color, direction and brightness features of the remote sensing image can be extracted, and the texture features of the image can be obtained by using the gray level co-occurrence matrix. The distribution range of the ship target is estimated according to the resolution of the remote sensing image and the shape features of the ship-like target, and finally the regional filtering is performed through threshold segmentation to obtain the ship target. This method has a high recognition accuracy for a single ship target, but the accurate segmentation effect for multiple ships at a close distance is poor.

[0010] In recent years, with the rapid development of deep learning algorithms, object detection and recognition based on deep learning have been rapidly promoted and applied. For image recognition problems, feature extraction is an important part of the detection process, and the quality of features will directly affect the accuracy of detection. Traditional machine learning algorithms require manual design of feature sets for complex scenes, which is time-consuming and requires a high level of technical skills from researchers. The emergence of deep learning algorithms effectively solves the shortcomings of manually designed features, realizes automatic feature extraction, and can automatically combine simple features to form more complex features.

[0011] 2. Sea surface target detection and recognition based on infrared images

[0012] Sea surface target detection and recognition based on infrared images generally include three categories. The first is the infrared small target detection algorithm based on traditional methods, such as mathematical morphology methods, which can effectively eliminate the interference of high-frequency noise and high-frequency signals of background edges, and mainly detect targets of corresponding shapes in images through structural elements with certain forms and basic operations; the second is the method based on the human visual system. This type of algorithm uses the significant changes in local texture caused by the human eye to sea surface targets based on the perception characteristics of the human eye, such as using the brightness feature map, color feature map and direction feature map of the image, combining feature maps of different scales to obtain brightness, color and direction saliency maps, and fusion to obtain the final visual saliency map. The third is the infrared sea surface target detection algorithm based on neural networks. Deep learning technology provides a new perspective for sea surface target detection. Compared with traditional methods, neural networks have the advantages of fitting any complex function, strong fitting ability, and can complete more complex nonlinear mappings. In the case of large samples, better target detection and recognition results can be obtained.

[0013] 3. Sea surface target detection and recognition based on SAR images

[0014] The constant false alarm rate (CFAR) target detection algorithm is one of the most commonly used methods in radar image detection. It mainly detects targets in images through a predetermined background clutter distribution model and false alarm rate. The algorithm has the characteristics of low computational complexity and adaptive threshold to maintain a constant false alarm rate. However, complex sea conditions can lead to high non-uniformity of sea clutter in SAR images, thus affecting the effect of constant false alarm detection. Therefore, many scholars have improved and developed the CFAR detection algorithm.

[0015] In addition to the CFAR detection algorithm, there are also template-based detection algorithms. This type of algorithm not only considers the characteristics of the target, but also the characteristics of background clutter. Among them, the most famous is the SUMO detector; detection algorithms based on SAR image target ship features, such as the shape, texture and other features of the ship. A typical example is the high-resolution SAR image target ship detection algorithm based on feature analysis, which mainly uses the target ship's kernel density estimation, aspect ratio and other features to identify the ship. In addition, there is a method of detecting ships using target clustering. This method is based on a high-pass filter in nonlinear space. First, the filter is roughly adjusted to produce the maximum response to structures whose size is less than or equal to the expected object. After filtering, histogram and connected area processing are performed, and then the relevant theoretical knowledge of shape and cluster deployment mode is used to eliminate false alarms.

[0016] 4. Sea surface target detection and recognition based on active radar

[0017] Sea surface target detection based on active radar mainly utilizes the one-dimensional range image data formed by radar echo and uses target detection algorithm to detect targets. It is widely used on airborne and missile-borne platforms.

[0018] 5. Sea surface target detection and recognition based on passive radar

[0019] Passive radar detection is also called passive detection. It is generally used on ground fixed platforms to detect and track non-cooperative "low, slow and small" aerial targets. It can also be used on larger electronic reconnaissance aircraft, electronic reconnaissance ships or some high-dynamic platforms (such as missile-borne platforms) to perceive electromagnetic radiation signals of specific categories of targets. Since it is difficult to obtain passive radar signals of non-cooperative targets, there are few literatures on the processing of passive radar signals, and most of them use simulation to obtain signal samples.

[0020] 6. Sea surface target detection and recognition based on multimodal fusion

[0021] Due to the limitations of platform payload and computing device performance, in the application of multi-modal information fusion, multi-modal target detection and recognition are mostly used in large manned electronic reconnaissance aircraft, integrated reconnaissance and strike drones, reconnaissance ships, reconnaissance vehicles, high-orbit satellites and other systems. With the gradual reduction in the size and power consumption of electronic devices and the rapid improvement of AI computing performance, multi-modal fusion recognition has also begun to be promoted and applied in some smaller drones and missile-borne platforms. At present, in the field of sea surface target detection and recognition, various modal applications based on optical imaging such as visible light and infrared, high-resolution two-dimensional imaging based on SAR, and one-dimensional range images based on active and passive radar signals have been widely used, but in the multi-modal fusion of optical images, radar images and radar signals, especially in some high-dynamic small unmanned platform applications, there is no relatively complete practical framework. In addition, since it is difficult to obtain samples of typical sea surface target detection by missile-borne or small unmanned aerial vehicle platforms, how to apply the current popular data-driven target detection and recognition algorithm in the case of small samples remains to be explored in depth.

[0022] The present invention proposes a multimodal fusion recognition framework based on a multi-kernel learning method, and designs a small sample multimodal fusion target recognition algorithm for sea surface targets on this basis. The algorithm extracts multimodal heterogeneous features of visible light images, SAR images, and active and passive radar one-dimensional range image signals, maps them using a multi-scale kernel method, and then performs weighted fusion of the kernel matrix, and then performs global training optimization based on structural risk minimization, and finally outputs the classification results of each sample. The algorithm takes into account the performance of the data-driven target recognition algorithm, as well as the platform hardware computing power limitations and small sample conditions, and can achieve a balance between performance, computing power, and samples. Summary of the invention

[0023] The purpose of the present invention is to avoid the shortcomings of the prior art and provide a multi-mode fusion sea surface target recognition method for a high-dynamic unmanned platform. It designs a sea surface target recognition system based on a multi-mode heterogeneous feature fusion multi-core classifier to solve the problems raised in the above background technology.

[0024] To achieve the above purpose, the technical solution adopted by the present invention is: the multi-mode fusion sea surface target recognition method of a high-dynamic unmanned platform is characterized by comprising:

[0025] Step 1: Preprocessing of sea surface target data.

[0026] The image data is acquired through the high-dynamic unmanned platform sensor module, including image filtering preprocessing of optical images and SAR images from sea surface targets, and signal sampling preprocessing of active radar signals and passive radar signals from sea surface targets;

[0027] Step 2: Extraction of sea surface target features.

[0028] The four types of data, namely, optical image, SAR image, active radar signal and passive radar signal, are subjected to feature extraction respectively. For optical image and SAR image data, the fast sampling filtering method is used after preprocessing, and the PCA algorithm is used for feature extraction to convert image samples into feature vectors. For active radar signal and passive radar signal data, the feature extraction after preprocessing mainly completes the calculation of 10 feature quantities, namely, enhanced average absolute value, average amplitude change, wavelength, maximum fractal length, root mean square, zero crossing, slope sign change, time moment, Teager-Kaiser energy operator and skewness, and then combines them into a 10-dimensional feature vector.

[0029] Step 3: Sea surface target identification,

[0030] The modal features of the four data features of sea surface targets, namely optical images, SAR images, active radar signals and passive radar signals, are extracted and identified through a sea surface target recognition method based on multi-mode heterogeneous features fusion and multi-core machine learning, i.e., the determination of target categories. It mainly completes the functions of model storage, multi-feature kernel matrix operation, fusion of multi-mode heterogeneous features, recognition test and storage of recognition results.

[0031] Furthermore, the optical image of the sea surface target in step 1 is subjected to image filtering preprocessing, which is to use a general sea surface target detection method based on image threshold segmentation and morphology to achieve preliminary extraction and processing of the sea surface target, which specifically includes the following steps:

[0032] (1) Using a certain threshold, perform threshold segmentation on the image to obtain a binary image;

[0033] (2) Perform morphological processing on the binary image, including dilation operation, opening operation, and erosion operation, to obtain a binary image of the target;

[0034] (3) Calculate the center of gravity of each target in the binary image and mark it;

[0035] (4) Expand outward from the center of gravity, mark the target area, and obtain the target slice image.

[0036] The determined threshold refers to converting the original image from a color image to a grayscale image, first performing histogram equalization on the image, and then using a global threshold method to maximize the difference in variance between the foreground and background parts of the image.

[0037] The dilation operation is implemented by using a structural element on a binary image. The element is a binary matrix, usually a small rectangle, circle or ellipse. In the dilation operation, for each pixel, if there is a white pixel overlapping with the structural element, the pixel is also set to white. The effect of the operation is to make the target object larger, the connectivity stronger, fill holes, and increase the perimeter of the object; the opening operation is a combination of corrosion and then dilation. It can be used to remove small holes, slender objects and noise points, make the edges of objects smooth, and will not destroy the connectivity of the object; the corrosion operation is the process of reducing the shape of the target object in the image, which is implemented by using a structural element on the binary image. In the corrosion and expansion operation, for each pixel, if the structural element is completely contained in the black pixel, the pixel is set to black. The effect of the operation is to make the target object smaller, separate connected objects, remove noise points outside the object and reduce the perimeter.

[0038] The centroid of each target in the binary image is determined by calculating the total moment of the target pixel set. The specific calculation method is as follows:

[0039] 1) Extract the target pixel set in the binary image, assuming that the target pixel set is S;

[0040] 2) Calculate the total moment of the target pixel set S, denoted as M00. The total moment can be calculated by the secondary moment in the horizontal and vertical directions: Where I(x,y) is the pixel value in the binary image. If (x,y)∈S, then I(x,y)=1, otherwise I(x,y)=0;

[0041] 3) Calculate the numerator of the total moment, which represents the secondary moment of the target in the horizontal and vertical directions, denoted as M10 and M01 respectively.

[0042] 4) Calculate the coordinates of the center of gravity (xc, yc), x c =M 10 M 00 , the coordinates of the centroid (xc, yc) represent the center position of the target in the binary image. In actual programming, these calculations can be simplified using functions provided by libraries such as NumPy.

[0043] Further, the SAR image of the sea surface target in step 1 is subjected to image filtering preprocessing, which is to realize the preliminary extraction and processing of the sea surface target by using an image segmentation algorithm based on a constant false alarm rate (CFAR), which specifically includes the following steps:

[0044] (1) Preprocessing the image to suppress noise, including performing a logarithmic transformation on the image to convert the multiplicative speckle noise into additive noise; using a general template filtering method to perform a convolution operation on the image and the template. Common template filtering methods include Gaussian filtering and mean filtering; performing an exponential operation on the image obtained after template filtering to transform it back to the image in the original image space;

[0045] (2) The target area is obtained by CFAR detection of the preprocessed segmentation, that is, the region of interest (ROI). The key to CFAR detection technology is to determine the adaptive threshold. Assume that p(x) is the probability density function of the radar background clutter distribution model. Let Obviously, F(x) is an increasing function on [0,+∞). By solving the equation,

[0046]

[0047] Get the threshold I, where p fa The traditional CFAR method is mainly a method for processing 1D radar signals. When used for SAR images, a sliding window method is usually used to traverse the image: the central pixel of the sliding window is used as the test point (Cell Under Test, CUT), and the pixels around the window are used to estimate the clutter distribution, completing target detection at a fixed constant false alarm rate;

[0048] Furthermore, the feature extraction of the optical image and SAR image data in step 2 adopts a fast and robust feature extraction method for sea surface target images, including a method for extracting local multi-resolution features, a method for designing and implementing an image filter kernel, and a method for removing redundant features of image targets;

[0049] The method for extracting local multi-resolution features is specifically implemented in the following steps:

[0050] (1) Use a non-uniform sampling filter kernel to perform non-uniform filtering sampling on the target in each band. The sampling is uniform within the fovea, and the maximum sampling rate is 1, which is the original resolution of the image. In the peripheral area, the sampling rate gradually decreases with the distance from the center point. In this way, the input information is greatly compressed, which greatly reduces the difficulty of storage and processing.

[0051] (2) After extracting feature maps in different scale spaces, the central-peripheral difference strategy is imitated and the interpolation difference method is used to perform peripheral difference transformation on the multi-scale space to filter out background noise and amplify the contour information of the detected target;

[0052] (3) The contrast scale space feature map is stretched row by row to obtain a one-dimensional feature vector, and the PCA algorithm is used to reduce its dimensionality to obtain the final multi-scale feature;

[0053] The design and implementation method of the image filter kernel is based on the "Center-Surround Antagonism, CSA" mechanism of the biological visual system. By simulating the change of sampling rate with distance in the visual attention mechanism, the present invention constructs a non-uniform sampling filtering method, in which uniform sampling with a sampling rate of 1 is performed within the fovea, that is, the image is not compressed; outside the fovea, the sampling rate decreases with the increase of the distance from the sampling point to the center point, and the farther away from the fovea, the smaller the sampling rate; in summary, the filter kernel is set as:

[0054]

[0055] The image size is N×N, where N is usually an integer power of 2, and δ represents the radius of the fovea. When it is within δ, it is uniformly sampled.

[0056] Based on the discrete wavelet transform (DWT) method of discretizing the scale according to the power series on the scale-shift phase plane, the points on the phase plane are non-uniform. This aspect also uses the DWT idea to set the scale parameter.

[0057] For a fixed scale parameter m, the corresponding sampling rate is γα -m ,make:

[0058] h(x)=γα -m (3)

[0059] Since m is an integer power of 2,

[0060] m=-g(log 2 h(x)) (4)

[0061] Among them, g(y) is the rounding down of y. For the filter kernel, the scale parameter m is the same within the same concentric circle. In the central area, the scale parameter m=0; outside the central area, the scale parameter increases with the increase of distance;

[0062] The method for removing redundant features of image targets refers to extracting features of different scales, removing redundant information between different scales, so as to more effectively detect stable key points and strengthen the spatial texture features of the image, that is, performing differential processing on feature images obtained at different scales, and outputting the layered results as features; wherein the fine-scale feature map represents the output features of the light-giving center cells, and the coarse-scale feature map represents the output features of the light-removing center cells. By performing differential processing on the feature maps between different scales, the texture features of the image can be strengthened, and the image noise can be filtered, so as to fully utilize the image scale information;

[0063] In order to improve the classification speed and reduce the feature dimension, the correlation between the feature vectors must be eliminated. Use principal component analysis (PCA) to whiten the data, calculate the eigenvalues ​​and eigenvectors of the sample's correlation coefficient matrix, sort the eigenvalues ​​in descending order, and adjust the eigenvectors accordingly; calculate the contribution rate of each eigenvalue, extract the principal component eigenvector according to the given extraction rate, and project the sample on the extracted eigenvector to extract the features;

[0064] Furthermore, the feature extraction of active radar signal and passive radar signal data in step 2 adopts a general feature extraction method of active and passive radar one-dimensional range profiles.

[0065] For active radar, the characteristics of target range profile length, number of scattering centers, range profile symmetry, dispersion, and range profile entropy can reflect the characteristics of the target to a certain extent; while for passive radar, the characteristic parameters that can be used for target identification include pulse width, pulse arrival time, pulse repetition period or repetition frequency, pulse arrival angle, pulse repetition interval (PRI) variation, carrier frequency variation, etc. Affected by environmental factors, it is not possible to extract all this information in each frame of active and passive radar signals, or this characteristic information may not be able to accurately reflect the type of target, but for both sensors, the target can be described in the form of a one-dimensional range profile signal.

[0066] In order to simplify the feature extraction process of active and passive radar signals, the radar one-dimensional range profile signal is treated as a conventional digital signal. Its most basic features include wavelength and mean absolute value (MAV). Based on this, 10 simple and effective signal feature parameters are extracted. Finally, these 10 feature parameters constitute the feature vector of the target, which has a small amount of calculation and is more general. Let x i represents the i-th component value of the signal x, L represents the total length of the signal, and its mathematical definition is as follows:

[0067] 1) Enhanced Mean absolute value (EMAV)

[0068]

[0069] 2) Average Amplitude Change (AAC)

[0070]

[0071] 3) Waveform Length (WL)

[0072]

[0073] 4) Maximum Fractal Length (MFL)

[0074]

[0075] 5) Root Mean Square (RMS)

[0076]

[0077] 6) Zero Crossing (ZC)

[0078]

[0079] 7) Slope Sign Change (SSC)

[0080]

[0081] 8) Temporal Moment (TM)

[0082]

[0083] 9) Teager–Kaiser Energy Operator (Log Teager Kaiser Energy Operator, LTKEO)

[0084]

[0085] 10) Skew

[0086]

[0087] Furthermore, the sea surface target recognition method of multi-mode heterogeneous feature fusion multi-core machine learning in step 3 includes designing a typical sea surface target multi-model sample library, designing a general sea surface target recognition framework integrating multi-mode features and multi-core learning methods, a target recognition algorithm based on multi-mode heterogeneous feature fusion and multi-core machine learning, and evaluation of experimental results of multi-mode heterogeneous feature fusion multi-core machine learning;

[0088] The design of the typical sea surface target multi-sample library includes obtaining the target's visible light image, SAR image, active radar one-dimensional range image, and passive radar one-dimensional range image to construct the multi-sample library. The visible light image sample is a target slice intercepted from the public Google earth high-definition image, with a size of 512×512; the SAR target image corresponding to the visible light image is simulated and generated by the image style transfer method based on the Generative Adversarial Networks (GANs), the source domain target image is the collected visible light image, the training image is a public sea surface SAR image data set, and the target domain image is the generated SAR image target slice, which has the same size as the visible light image, also 512×512; the active radar one-dimensional range image data is simulated and generated by a linear frequency modulation signal with a bandwidth of 30MHz. The effective working frequency range of the passive radar is composed of four frequency bands of 0.85 / 0.94GHz, 0.95 / 1.2GHz, 2.9 / 3.1GHz, and 5.4 / 5.8GHz.

[0089] The sea surface target recognition framework that integrates the multi-modal features and multi-core learning methods is designed in a universal way. It extracts multi-scale features of the target based on a variety of missile-borne sensors, organically integrates the multi-modal features with the multi-core in the multi-core learning framework, obtains the fusion kernel matrix by mapping the samples to the regenerated Hilbert space, and then optimizes the multi-core combination coefficients through iterative training, and finally obtains the trained multi-scale multi-core classifier. By inputting the test samples, the target is finally recognized.

[0090] In order to capture the scale variability of complex targets, the target information features obtained by different sensors are extracted and visible light image kernels, SAR image kernels, active radar data kernels and passive radar data kernels at different scales are constructed, and weighted combinations are performed on them. The specific form is as follows:

[0091] Visible light image:

[0092] SAR imagery:

[0093] Active Radar:

[0094] Passive radar:

[0095] in represents the k-th scale feature extracted from the i-th target, δ k Represents the bandwidth of the Gaussian kernel function. Then, the above four base kernels are substituted into the multi-scale multi-kernel learning framework, and then the multi-scale kernel classifier is trained based on this to realize the fusion classification of multi-source heterogeneous features.

[0096] The target recognition algorithm based on multi-mode heterogeneous feature fusion and multi-core machine learning is based on the principle of multi-mode heterogeneous feature fusion multi-scale kernel classifier target recognition framework, and specifically includes the following steps:

[0097] (1) Preprocessing: Slice, segment, translate, rotate, scale, and normalize visible light images and SAR images, and divide the samples into training sets and test sets;

[0098] (2) Feature extraction: Based on the extraction of multi-resolution robust features of visible light images and SAR images, and the extraction of feature parameters of active and passive radar data, feature vectors representing various types of targets are obtained, and a multi-source feature library of various types of targets is established;

[0099] (3) Feature fusion: The base kernel matrix is ​​calculated on each feature library according to the kernel functions of different scales, and the synthetic kernel matrix is ​​calculated by linear weighting;

[0100] (4) Learning combination weights: Use the multi-core learning algorithm to iteratively train the SVM classifier and update the combination weights until the iteration stop condition is met;

[0101] (5) Target identification: import test set features and use the trained classifier for identification;

[0102] The target recognition algorithm based on multi-mode heterogeneous feature fusion and multi-core machine learning uses the SimpleMKL algorithm of the simple multi-core learning method as the multi-core classifier. By adding constraints to the sparse kernel combination weights, a weighted 2-norm regularization formula is designed to solve the MKL problem. The classifier has a simple principle and is efficient to implement. In the SimpleMKL algorithm, the synthetic kernel model that satisfies Mercer's theorem is a weighted convex sum of several simple basic kernel functions. The dot product operation in the complex feature space is replaced by the synthetic kernel function;

[0103]

[0104] In the formula, K m is a positive definite kernel function, and a definition of <·,·> m The reproducing kernel Hilbert space H of the dot product operation mIt can be simply a classical kernel with different parameters, such as linear kernel, polynomial kernel, Gaussian kernel, etc. M is the number of basis kernels, d m is the weight coefficient of the corresponding base kernel function. Then, by adding constraints to the weights of the sparse kernel combination, a weighted l2-norm regularized convex and differentiable primitive objective formula is designed to solve the MKL problem, which can effectively obtain an approximate primitive solution, as shown in Equation (16).

[0105]

[0106] in,

[0107]

[0108] For J(d), the derivative of the Lagrangian function is set to zero according to the original variable, and its dual problem is obtained as follows:

[0109]

[0110] Note that, let K(x i ,x j )replace The above formula is actually the dual form of the standard support vector machine, which can be calculated by any standard support vector machine solver; therefore, under this objective function framework, the MKL problem is transformed into learning the kernel function weight coefficient d through a single optimization problem in a supervised environment m and the kernel coefficient α i ;

[0111] SimpleMKL uses a two-step method to iterate the standard SVM model. In each iteration, the Lagrangian operator α of the SVM is first fixed and the kernel weight d is updated. Then the updated kernel weight d is fixed and the Lagrangian operator α of the SVM is updated at the same time until the stopping criterion is met. In this way, the hot-start-based two-step training method can reduce the complexity of the standard support vector machine solver. The stopping criterion can be based on the duality gap, the KKT condition, the rate of change of d between two iterations, or simply based on the maximum number of iterations.

[0112] The multi-core construction method used by SimpleMKL is also a simple linear combination of classic kernels with different parameters. In order to give full play to the effectiveness of multi-resolution features of image targets, the present invention integrates multi-core learning and scale analysis methods, and designs an automatic target recognition method based on the organic fusion of local multi-resolution feature extraction and multi-scale kernels. This method is based on the idea of ​​integrating multi-level features, uses an SVM model based on a multi-scale kernel method, and uses kernel functions of different scales to map image features of different levels separately, and then synthesizes them, which can greatly improve the discriminability of features and improve classification accuracy.

[0113]

[0114] The calculation of the above formula does not require the knowledge of the input space D = {(x i ,y i )}(i=1,2,L,N) and feature mapping function Φ(·), the feature set S is implicitly mapped to the multi-scale multi-kernel space through the kernel technique, which reduces the amount of calculation; before feature fusion, the basic kernel function K in the multi-scale feature fusion process is used m (·,·), usually a basis kernel function with multi-scale representation capability is selected, such as the Gaussian radial basis kernel, as shown below:

[0115]

[0116] where σ 1 <···<σ M When σ is small, the classification effect is good for samples with drastic changes, while when σ is large, the classification effect is good for samples with gentle changes, and better generalization performance can be obtained. In specific implementation, the value of σ can refer to the law of scale change in wavelet transform, and σ can be defined by the following formula:

[0117] σ i =2 i σ,i=0,1,2L (21)

[0118] The evaluation of the experimental results of multi-mode heterogeneous feature fusion multi-core machine learning uses three indicators as evaluation criteria: overall classification accuracy (OA), average classification accuracy (AA) and Kappa coefficient. The overall classification accuracy refers to the ratio of the number of correctly classified samples to the total number of samples. The average classification accuracy refers to the average classification accuracy of each type of sample. Let C be the number of categories, N be the number of samples, and N i is the number of samples in the i-th category, M is the confusion matrix obtained from the experiment, and the calculation methods of OA, AA and Kappa coefficients are as follows:

[0119]

[0120] The beneficial effects of the present invention are as follows: the present invention provides a multi-mode fusion sea surface target recognition method for a high-dynamic unmanned platform, which proposes a multi-mode fusion recognition framework based on a multi-kernel learning method, and designs a small-sample multi-mode fusion target recognition algorithm for sea surface targets on this basis. By targeting the data characteristics of different sources of visible light images, SAR images, and active and passive radar one-dimensional range image signals, target detection and feature extraction are performed respectively, and multi-mode heterogeneous features are mapped using a multi-scale kernel method, and weighted fusion of the kernel matrix is ​​performed. After global training optimization, the classification results of each sample can be output in the end. The algorithm takes into account the performance of the data-driven target recognition algorithm, as well as the platform hardware computing power limitations and small sample conditions, and can achieve a balance between performance, computing power, and samples. BRIEF DESCRIPTION OF THE DRAWINGS

[0121] Figure 1 It is a schematic diagram of the process of the present invention;

[0122] Figure 2 It is a schematic diagram of an extraction method for extracting local multi-resolution features in the implementation of the present invention;

[0123] Figure 3 It is a schematic diagram of a typical visible light image target multi-scale sampling feature extraction method in the implementation of the present invention;

[0124] Figure 4 is a schematic diagram of multi-scale filtering results of a typical visible light image target in the implementation of the present invention;

[0125] Figure 5 It is a schematic diagram of multi-scale filtering results of a typical SAR image target in the implementation of the present invention;

[0126] Figure 6 Schematic diagram of multi-scale comparison results of central-peripheral differences of visible light images in the implementation of the present invention;

[0127] Figure 7 Schematic diagram of multi-scale comparison results of central-peripheral differences of SAR images in the implementation of the present invention;

[0128] Figure 8 It is a schematic diagram of the two-dimensional characteristic distribution of active radar signals extracted in the implementation of the present invention;

[0129] Fig. 9 It is a schematic diagram of the two-dimensional characteristic distribution of passive radar signals extracted in the implementation of the present invention;

[0130] Fig.10 It is a flow chart of the sea surface target recognition algorithm based on multi-source heterogeneous feature fusion in the implementation of the present invention;

[0131] Fig.11 The present invention is implemented Figure 1 The corresponding schematic diagram. DETAILED DESCRIPTION

[0132] The principles and features of the present invention are described below in conjunction with the accompanying drawings. The examples given are only used to explain the present invention and are not used to limit the scope of the present invention.

[0133] In order to achieve the above object, the present invention provides the following specific implementation methods: Figure 1 and 11 As shown, the multi-mode fusion sea surface target recognition method for a high-dynamic unmanned platform is characterized by:

[0134] Step 1: Preprocessing of sea surface target data.

[0135] The image data is acquired through the high-dynamic unmanned platform sensor module, including image filtering preprocessing of optical images and SAR images from sea surface targets, and signal sampling preprocessing of active radar signals and passive radar signals from sea surface targets;

[0136] The image filtering preprocessing of optical images mainly realizes noise filtering, distortion correction and target segmentation preprocessing; the image filtering preprocessing module of SAR images mainly realizes image coherent speckle noise suppression and CFAR target detection; the signal sampling preprocessing of active radar signals mainly realizes discrete sampling of one-dimensional range image signals of active radars to obtain one-dimensional discrete signals of equal length; the signal sampling preprocessing of passive radar signals mainly realizes discrete sampling of one-dimensional range image signals of passive radars to obtain one-dimensional discrete signals of equal length;

[0137] Step 2: Extraction of sea surface target features.

[0138] The four types of data, namely, optical image, SAR image, active radar signal and passive radar signal, are subjected to feature extraction respectively. For optical image and SAR image data, the fast sampling filtering method is used after preprocessing, and the PCA algorithm is used for feature extraction to convert image samples into feature vectors. For active radar signal and passive radar signal data, the feature extraction after preprocessing mainly completes the calculation of 10 feature quantities, namely, enhanced average absolute value, average amplitude change, wavelength, maximum fractal length, root mean square, zero crossing, slope sign change, time moment, Teager-Kaiser energy operator and skewness, and then combines them into a 10-dimensional feature vector.

[0139] Step 3: Sea surface target identification,

[0140] The modal features of the four data features of sea surface targets, namely optical images, SAR images, active radar signals and passive radar signals, are extracted and identified through a sea surface target recognition method based on multi-mode heterogeneous features fusion and multi-core machine learning, i.e., the determination of target categories. It mainly completes the functions of model storage, multi-feature kernel matrix operation, fusion of multi-mode heterogeneous features, recognition test and storage of recognition results.

[0141] Furthermore, the optical image of the sea surface target in step 1 is subjected to image filtering preprocessing, which is to use a general sea surface target detection method based on image threshold segmentation and morphology to achieve preliminary extraction and processing of the sea surface target, such as Figure 2 As shown, the specific steps include:

[0142] (1) Using a certain threshold, perform threshold segmentation on the image to obtain a binary image;

[0143] (2) Perform morphological processing on the binary image, including dilation operation, opening operation, and erosion operation, to obtain a binary image of the target;

[0144] (3) Calculate the center of gravity of each target in the binary image and mark it;

[0145] (4) Expand outward from the center of gravity, mark the target area, and obtain the target slice image.

[0146] The determined threshold refers to converting the original image from a color image to a grayscale image, first performing histogram equalization on the image, and then using a global threshold method to maximize the difference in variance between the foreground and background parts of the image.

[0147] The dilation operation is implemented by using a structural element on a binary image. The element is a binary matrix, usually a small rectangle, circle or ellipse. In the dilation operation, for each pixel, if there is a white pixel overlapping with the structural element, the pixel is also set to white. The effect of the operation is to make the target object larger, the connectivity stronger, fill holes, and increase the perimeter of the object; the opening operation is a combination of corrosion and then dilation. It can be used to remove small holes, slender objects and noise points, make the edges of objects smooth, and will not destroy the connectivity of the object; the corrosion operation is the process of reducing the shape of the target object in the image, which is implemented by using a structural element on the binary image. In the corrosion and expansion operation, for each pixel, if the structural element is completely contained in the black pixel, the pixel is set to black. The effect of the operation is to make the target object smaller, separate connected objects, remove noise points outside the object and reduce the perimeter.

[0148] The centroid of each target in the binary image is determined by calculating the total moment of the target pixel set. The specific calculation method is as follows:

[0149] 1) Extract the target pixel set in the binary image, assuming that the target pixel set is S;

[0150] 2) Calculate the total moment of the target pixel set S, denoted as M00. The total moment can be calculated by the secondary moment in the horizontal and vertical directions: Where I(x,y) is the pixel value in the binary image. If (x,y)∈S, then I(x,y)=1, otherwise I(x,y)=0;

[0151] 3) Calculate the numerator of the total moment, which represents the secondary moment of the target in the horizontal and vertical directions, denoted as M10 and M01 respectively.

[0152] 4) Calculate the coordinates of the center of gravity (xc, yc), x c =M 10 M 00 , the coordinates of the centroid (xc, yc) represent the center position of the target in the binary image. In actual programming, these calculations can be simplified using functions provided by libraries such as NumPy.

[0153] The center of gravity extracted in the above process can also be used as the starting position for multi-resolution decomposition and multi-resolution feature extraction of the slice image, which is convenient for subsequent multi-target feature acquisition and classification testing; the target detection result image area can also be directly used as a target sample as a training or test sample for the deep neural network recognizer.

[0154] Further, the SAR image of the sea surface target in step 1 is subjected to image filtering preprocessing, which is to realize the preliminary extraction and processing of the sea surface target by using an image segmentation algorithm based on a constant false alarm rate (CFAR), which specifically includes the following steps:

[0155] (1) Preprocessing the image to suppress noise, including performing a logarithmic transformation on the image to convert the multiplicative speckle noise into additive noise; using a general template filtering method to perform a convolution operation on the image and the template. Common template filtering methods include Gaussian filtering and mean filtering; performing an exponential operation on the image obtained after template filtering to transform it back to the image in the original image space;

[0156] (2) The target area is obtained by CFAR detection of the preprocessed segmentation, that is, the region of interest (ROI). The key to CFAR detection technology is to determine the adaptive threshold. Assume that p(x) is the probability density function of the radar background clutter distribution model. Let Obviously, F(x) is an increasing function on [0,+∞). By solving the equation,

[0157]

[0158] Get the threshold I, where p fa The traditional CFAR method is mainly a method for processing 1D radar signals. When used for SAR images, a sliding window method is usually used to traverse the image: the central pixel of the sliding window is used as the test point (Cell Under Test, CUT), and the pixels around the window are used to estimate the clutter distribution, completing target detection at a fixed constant false alarm rate;

[0159] The detection method based on CFAR uses statistical characteristics for detection, and there is no need to divide the training set and the test set. The detection of ship targets is mainly to obtain the position, so as to cut the image and provide samples for target recognition. The present invention uses the Gaussian model as the clutter statistical model to detect the SAR image ship model, so that the possible position of the target in the SAR image can be quickly detected. Figure 3-Figure 5 The figure shows the CFAR segmentation detection result of the sea surface target in the SAR image example. From the detection results of the image samples, it can be found that the CFAR-based method can detect ship targets well, and can detect ship-type targets well when the background is relatively simple; when the SAR image background becomes complex, such as including islands or ports, the detection accuracy decreases, and non-ship-type targets on the sea surface may be detected, that is, false alarms. Therefore, the segmentation detection result of the method of the present invention can still be used as a preprocessing result to obtain a rough target position, and then obtain a target slice for further target recognition.

[0160] The signal sampling preprocessing method of the active radar signal and the passive radar signal of the sea surface target in step 1 belongs to the function of the active and passive radar system. When this patent is implemented, the system obtains the preprocessed data.

[0161] Furthermore, the feature extraction of the optical image and SAR image data in step 2 adopts a fast and robust feature extraction method for sea surface target images, including a method for extracting local multi-resolution features, a method for designing and implementing an image filter kernel, and a method for removing redundant features of image targets;

[0162] The method for extracting local multi-resolution features is as follows: Figure 3 As shown, the specific implementation steps are as follows:

[0163] (1) Use a non-uniform sampling filter kernel to perform non-uniform filtering sampling on the target in each band. The sampling is uniform within the fovea, and the maximum sampling rate is 1, which is the original resolution of the image. In the peripheral area, the sampling rate gradually decreases with the distance from the center point. In this way, the input information is greatly compressed, which greatly reduces the difficulty of storage and processing.

[0164] (2) After extracting feature maps in different scale spaces, the central-peripheral difference strategy is imitated and the interpolation difference method is used to perform peripheral difference transformation on the multi-scale space to filter out background noise and amplify the contour information of the detected target;

[0165] (3) The contrast scale space feature map is stretched row by row to obtain a one-dimensional feature vector, and the PCA algorithm is used to reduce its dimensionality to obtain the final multi-scale feature;

[0166] This method constructs a multi-scale non-uniform sampling space, extracts features of the target in different scale spaces, uses small-scale kernels to detect areas with significant feature changes, and large-scale kernels to detect areas with relatively gentle changes, fully mining the spatial scale information of remote sensing images to improve classification accuracy.

[0167] The design and implementation method of the image filter kernel is based on the "Center-Surround Antagonism, CSA" mechanism of the biological visual system. By simulating the change of sampling rate with distance in the visual attention mechanism, the present invention constructs a non-uniform sampling filtering method, in which uniform sampling with a sampling rate of 1 is performed within the fovea, that is, the image is not compressed; outside the fovea, the sampling rate decreases with the increase of the distance from the sampling point to the center point, and the farther away from the fovea, the smaller the sampling rate; in summary, the filter kernel is set as:

[0168]

[0169] The image size is N×N, where N is usually an integer power of 2, and δ represents the radius of the fovea. When it is within δ, it is uniformly sampled.

[0170] Based on the discrete wavelet transform (DWT) method of discretizing the scale according to the power series on the scale-shift phase plane, the points on the phase plane are non-uniform. This aspect also uses the DWT idea to set the scale parameter.

[0171] For a fixed scale parameter m, the corresponding sampling rate is γα -m ,make:

[0172] h(x)=γα -m (3)

[0173] Since m is an integer power of 2,

[0174] m=-g(log 2 h(x)) (4)

[0175] Among them, g(y) is the rounding down of y. For the filter kernel, the scale parameter m is the same within the same concentric circle. In the central area, the scale parameter m=0; outside the central area, the scale parameter increases with the increase of distance; Figure 4 and Figure 5 These are the sampling and filtering results of typical visible light images and SAR images at five different scales.

[0176] The method for removing redundant features of image targets refers to removing redundant information between different scales of extracted features, so as to more effectively detect stable key points and strengthen the spatial texture features of the image, that is, to perform differential analysis on feature images obtained at different scales and output the hierarchical results as features; Figure 6 and Figure 7 The feature maps are obtained after spatial differentiation at five different scales. Among them, the fine-scale feature map represents the output features of the light-adding center cells, and the coarse-scale feature map represents the output features of the light-removing center cells. By performing differential processing on the feature maps at different scales, the texture features of the image can be enhanced, and the image noise can be filtered, so as to make full use of the image scale information.

[0177] In order to improve the classification speed and reduce the feature dimension, the correlation between the feature vectors must be eliminated. Use principal component analysis (PCA) to whiten the data, calculate the eigenvalues ​​and eigenvectors of the sample's correlation coefficient matrix, sort the eigenvalues ​​in descending order, and adjust the eigenvectors accordingly; calculate the contribution rate of each eigenvalue, extract the principal component eigenvector according to the given extraction rate, and project the sample on the extracted eigenvector to extract the features;

[0178] Furthermore, the feature extraction of active radar signal and passive radar signal data in step 2 adopts a general feature extraction method of active and passive radar one-dimensional range profiles.

[0179] For active radar, the characteristics of target range profile length, number of scattering centers, range profile symmetry, dispersion, and range profile entropy can reflect the characteristics of the target to a certain extent; while for passive radar, the characteristic parameters that can be used for target identification include pulse width, pulse arrival time, pulse repetition period or repetition frequency, pulse arrival angle, pulse repetition interval (PRI) variation, carrier frequency variation, etc. Affected by environmental factors, it is not possible to extract all this information in each frame of active and passive radar signals, or this characteristic information may not be able to accurately reflect the type of target, but for both sensors, the target can be described in the form of a one-dimensional range profile signal.

[0180] In order to simplify the feature extraction process of active and passive radar signals, the radar one-dimensional range profile signal is treated as a conventional digital signal. Its most basic features include wavelength and mean absolute value (MAV). Based on this, 10 simple and effective signal feature parameters are extracted. Finally, these 10 feature parameters constitute the feature vector of the target, which has a small amount of calculation and is more general. Let x i represents the i-th component value of the signal x, L represents the total length of the signal, and its mathematical definition is as follows:

[0181] 1) Enhanced Mean absolute value (EMAV)

[0182]

[0183] 2) Average Amplitude Change (AAC)

[0184]

[0185] 3) Waveform Length (WL)

[0186]

[0187] 4) Maximum Fractal Length (MFL)

[0188]

[0189] 5) Root Mean Square (RMS)

[0190]

[0191] 6) Zero Crossing (ZC)

[0192]

[0193] 7) Slope Sign Change (SSC)

[0194]

[0195] 8) Temporal Moment (TM)

[0196]

[0197] 9) Teager–Kaiser Energy Operator (Log Teager Kaiser Energy Operator, LTKEO)

[0198]

[0199] 10) Skew

[0200]

[0201] The present invention simplifies the active and passive radar signal processing part, ignores the problems of system error correction and data format alignment, and directly extracts the ten-dimensional feature parameters of the acquired active and passive radar signals according to the above definition. Figure 8 and Fig. 9 The typical aircraft carrier, destroyer and shipless sea surface are used as targets, and the features of the acquired active and passive radar signals are extracted respectively. It can be seen from the figure that under different environments, the influence of 10 features on the recognition of different targets is quite different. For active radar signals, the sea surface and destroyers are both aliased within the aircraft carrier feature range, while for the skewness feature, the three types of targets are completely separated. For passive radar signals, the two-dimensional features of the sea surface are quite different, while the two-dimensional features of aircraft carriers and destroyers are aliased together. Therefore, the traditional linear classifier can only distinguish whether it is a sea surface or a non-sea surface target. To further distinguish between destroyers and aircraft carriers, it is necessary to use nonlinear classifiers and the support of other features.

[0202] Furthermore, the sea surface target recognition method of multi-mode heterogeneous feature fusion multi-core machine learning in step 3 includes designing a typical sea surface target multi-model sample library, designing a general sea surface target recognition framework integrating multi-mode features and multi-core learning methods, a target recognition algorithm based on multi-mode heterogeneous feature fusion and multi-core machine learning, and evaluation of experimental results of multi-mode heterogeneous feature fusion multi-core machine learning;

[0203] The design of the typical sea surface target multi-sample library includes obtaining the target's visible light image, SAR image, active radar one-dimensional range image, and passive radar one-dimensional range image to construct the multi-sample library. The visible light image sample is a target slice intercepted from the public Google earth high-definition image, with a size of 512×512; the SAR target image corresponding to the visible light image is simulated and generated by the image style transfer method based on the Generative Adversarial Networks (GANs), the source domain target image is the collected visible light image, the training image is a public sea surface SAR image data set, and the target domain image is the generated SAR image target slice, which has the same size as the visible light image, also 512×512; the active radar one-dimensional range image data is simulated and generated by a linear frequency modulation signal with a bandwidth of 30MHz. The effective working frequency range of the passive radar is composed of four frequency bands of 0.85 / 0.94GHz, 0.95 / 1.2GHz, 2.9 / 3.1GHz, and 5.4 / 5.8GHz.

[0204] The present invention selects typical aircraft carriers, destroyers and targetless sea surfaces as three types of targets to be identified. In the above sample library, there are 1134 visible light image target slices, of which there are 390 aircraft carriers, 390 destroyers, and 354 targetless sea surface slices. There are 2571 SAR image slices, of which there are 857 aircraft carriers, destroyers and sea surface slices. The sample data of the three types of target images of multi-mode sensors in the sample library are detailed in Table 1. In the specific experiment, the data set is divided into training set and test set according to different proportions.

[0205] Table 1 Sample data of three types of target images of multi-mode sensors used in the experiment

[0206]

[0207] In order to preserve the original information of the image as much as possible, the visible light images in the sample library of the present invention are stored in BMP format, and the SAR images are stored in PNG format. In order to enhance the versatility and convenience of sample data storage and reading, the one-dimensional range image signals of active and passive radars in the sample library of the present invention are stored in the discrete time series after sampling in .txt document format, using variable-length UTF-8 encoding, and supporting Chinese characters.

[0208] The sea surface target recognition framework that integrates the multi-modal features and multi-core learning methods is designed in a universal way. It extracts multi-scale features of the target based on a variety of missile-borne sensors, organically integrates the multi-modal features with the multi-core in the multi-core learning framework, obtains the fusion kernel matrix by mapping the samples to the regenerated Hilbert space, and then optimizes the multi-core combination coefficients through iterative training, and finally obtains the trained multi-scale multi-core classifier. By inputting the test samples, the target is finally recognized.

[0209] In order to capture the scale variability of complex targets, the target information features obtained by different sensors are extracted and visible light image kernels, SAR image kernels, active radar data kernels and passive radar data kernels at different scales are constructed, and weighted combinations are performed on them. The specific form is as follows:

[0210] Visible light image:

[0211] SAR imagery:

[0212] Active Radar:

[0213] Passive radar:

[0214] in represents the k-th scale feature extracted from the i-th target, δ k Represents the bandwidth of the Gaussian kernel function. Then, the above four base kernels are substituted into the multi-scale multi-kernel learning framework, and then the multi-scale kernel classifier is trained based on this to realize the fusion classification of multi-source heterogeneous features.

[0215] For the small sample pattern analysis problem that the present invention focuses on, the kernel method represented by the support vector machine (SVM) has achieved great success in solving the nonlinear pattern analysis problem. By introducing the kernel technique and extending it to the nonlinear case, the kernel function operation is used to avoid the complex inner product operation in the feature space, and its excellent performance has been shown in the fields of pattern classification, target recognition, regression analysis, feature extraction, etc. Although the kernel machine is effective and practical in many fields, the performance of the kernel machine varies greatly in different application scenarios, so the selection of the kernel function and its parameters plays a key role in processing learning tasks. The multiple kernel learning method (MKL) uses multiple kernel functions instead of a specific kernel function, and achieves better classification and regression performance in some complex situations. On the one hand, different kernels correspond to different feature spaces, which can not only enhance the interpretability of the decision function but also improve the performance of the kernel machine; on the other hand, data from multiple sources have different characteristics, and different kernels use components from different sources as inputs. The fusion of heterogeneous data sources is achieved through multi-core combination, and then the classifier can be trained based on this to achieve good performance.

[0216] The sea surface target recognition method based on multi-mode heterogeneous feature fusion multi-kernel learning classifier proposed in the present invention makes full use of the complementary characteristics of multi-source sensors such as visible light, SAR, active and passive radar, and performs feature fusion processing on the information extracted by each of the multiple sensors to eliminate the redundancy and contradiction that may exist between the multi-sensor information, eliminate the influence of interference, reduce the uncertainty of target recognition, and thus improve the correct recognition rate.

[0217] The target recognition algorithm based on multi-mode heterogeneous feature fusion and multi-core machine learning is based on the principle of multi-mode heterogeneous feature fusion multi-scale kernel classifier target recognition framework, such as Fig.10 As shown, the specific steps include:

[0218] (1) Preprocessing: Slice, segment, translate, rotate, scale, and normalize visible light images and SAR images, and divide the samples into training sets and test sets;

[0219] (2) Feature extraction: Based on the extraction of multi-resolution robust features of visible light images and SAR images, and the extraction of feature parameters of active and passive radar data, feature vectors representing various types of targets are obtained, and a multi-source feature library of various types of targets is established;

[0220] (3) Feature fusion: The base kernel matrix is ​​calculated on each feature library according to the kernel functions of different scales, and the synthetic kernel matrix is ​​calculated by linear weighting;

[0221] (4) Learning combination weights: Use the multi-core learning algorithm to iteratively train the SVM classifier and update the combination weights until the iteration stop condition is met;

[0222] (5) Target identification: import test set features and use the trained classifier for identification;

[0223] The target recognition algorithm based on multi-mode heterogeneous feature fusion and multi-core machine learning uses the SimpleMKL algorithm of the simple multi-core learning method as the multi-core classifier. By adding constraints to the sparse kernel combination weights, a weighted 2-norm regularization formula is designed to solve the MKL problem. The classifier has a simple principle and is efficient to implement. In the SimpleMKL algorithm, the synthetic kernel model that satisfies Mercer's theorem is a weighted convex sum of several simple basic kernel functions. The dot product operation in the complex feature space is replaced by the synthetic kernel function;

[0224]

[0225] In the formula, K m is a positive definite kernel function, and a definition of <·,·> m The reproducing kernel Hilbert space H of the dot product operation mIt can be simply a classical kernel with different parameters, such as linear kernel, polynomial kernel, Gaussian kernel, etc. M is the number of basis kernels, d m is the weight coefficient of the corresponding base kernel function. Then, by adding constraints to the weights of the sparse kernel combination, a weighted l2-norm regularized convex and differentiable primitive objective formula is designed to solve the MKL problem, which can effectively obtain an approximate primitive solution, as shown in Equation (16).

[0226]

[0227] in,

[0228]

[0229] For J(d), the derivative of the Lagrangian function is set to zero according to the original variable, and its dual problem is obtained as follows:

[0230]

[0231] Note that, let K(x i ,x j )replace The above formula is actually the dual form of the standard support vector machine, which can be calculated by any standard support vector machine solver; therefore, under this objective function framework, the MKL problem is transformed into learning the kernel function weight coefficient d through a single optimization problem in a supervised environment m and the kernel coefficient α i ;

[0232] SimpleMKL uses a two-step method to iterate the standard SVM model. In each iteration, the Lagrangian operator α of the SVM is first fixed and the kernel weight d is updated. Then the updated kernel weight d is fixed and the Lagrangian operator α of the SVM is updated at the same time until the stopping criterion is met. In this way, the hot-start-based two-step training method can reduce the complexity of the standard support vector machine solver. The stopping criterion can be based on the duality gap, the KKT condition, the rate of change of d between two iterations, or simply based on the maximum number of iterations.

[0233] The multi-core construction method used by SimpleMKL is also a simple linear combination of classic kernels with different parameters. In order to give full play to the effectiveness of multi-resolution features of image targets, the present invention integrates multi-core learning and scale analysis methods, and designs an automatic target recognition method based on the organic fusion of local multi-resolution feature extraction and multi-scale kernels. This method is based on the idea of ​​integrating multi-level features, uses an SVM model based on a multi-scale kernel method, and uses kernel functions of different scales to map image features of different levels separately, and then synthesizes them, which can greatly improve the discriminability of features and improve classification accuracy.

[0234]

[0235] The calculation of the above formula does not require the knowledge of the input space D = {(x i ,y i )}(i=1,2,L,N) and feature mapping function Φ(·), the feature set S is implicitly mapped to the multi-scale multi-kernel space through the kernel technique, reducing the amount of calculation. Before feature fusion, the basic kernel function K in the multi-scale feature fusion process is m (·,·), usually a basis kernel function with multi-scale representation capability is selected, such as the Gaussian radial basis kernel, as shown below:

[0236]

[0237] where σ 1 <···<σ M When σ is small, the classification effect is good for samples with drastic changes, while when σ is large, the classification effect is good for samples with gentle changes, and better generalization performance can be obtained. In specific implementation, the value of σ can refer to the law of scale change in wavelet transform, and σ can be defined by the following formula:

[0238] σ i =2 i σ,i=0,1,2L (21)

[0239] The evaluation of the experimental results of multi-mode heterogeneous feature fusion multi-core machine learning uses three indicators as evaluation criteria: overall classification accuracy (OA), average classification accuracy (AA) and Kappa coefficient. The overall classification accuracy refers to the ratio of the number of correctly classified samples to the total number of samples. The average classification accuracy refers to the average classification accuracy of each type of sample. Let C be the number of categories, N be the number of samples, and N i is the number of samples in the i-th category, M is the confusion matrix obtained from the experiment, and the calculation methods of OA, AA and Kappa coefficients are as follows:

[0240]

[0241] In the experiment, support vector machine classifier is used, and multi-scale kernel function is introduced to realize multi-scale kernel classifier. The "One VS All" strategy is adopted to transform multi-class problem into two-class problem, and the class with the largest confidence probability is selected as the classification result. The value range of multi-scale kernel parameter is: σ∈{0.5,1:1:10}, and the value range of penalty coefficient is: C∈{1,10,100,1000}. The experiment will use aircraft carriers, destroyers and sea surface as identification targets, and divide the multi-sensor information fusion scheme into 3 groups, namely SAR / visible light / passive radar, SAR / visible light / active radar one-dimensional range profile and SAR / active radar one-dimensional range profile / passive radar. In order to avoid system self-interference, active radar and passive radar adopt time-sharing working mode.

[0242] First, a multi-mode heterogeneous fusion recognition experiment of SAR, visible light, and active radar one-dimensional range profile is conducted. For a fixed sample set, the ratio of the training set and the test set is adjusted, and the training set and the test set are randomly selected in proportion each time, and the average method of multiple experiments is adopted to fully examine the performance of the classifier model. The experimental results are shown in Table 2. It can be seen that as the number of training set samples decreases from 90% to 30%, the overall classification accuracy, average classification accuracy, and Kappa coefficient have a gradual downward trend, but the overall recognition rate is maintained at a high level, which fully demonstrates that the features used for classification are effective, the classification performance of the classifier model is very high, and the consistency of the model has also reached a high level.

[0243] In order to explore the robustness of the algorithm to interference, Gaussian noise and coherent speckle noise were added to the visible light and SAR sample images involved in the classification. The experimental results show that the classification accuracy is slightly affected. This shows that the one-dimensional range profile signal feature of the active radar has a high degree of recognizability. In order to further explore the robustness of the active radar feature classification, Gaussian noise of different intensity levels was added to the one-dimensional range profile signal of the active radar in the experiment to examine the recognition rate of the algorithm under different signal-to-noise ratios. The experimental results are shown in Table 3.

[0244] Table 2 SAR / visible light / active radar one-dimensional range image

[0245] Proportion 90％ 80％ 70％ 60％ 50％ 40％ 30％ Training / Testing 1020 / 114 907 / 227 793 / 341 680 / 454 567 / 567 453 / 681 340 / 794 Aircraft Carrier 1 1 1 1 1 0.9831 0.9962 Destroyer 1 1 1 0.9748 1 0.9957 0.9610 Sea surface 1 1 0.9907 0.9930 1 1 0.9960 OA 1 1 0.9971 0.9890 1 0.9927 0.9836 AA 1 1 0.9956 0.9835 1 0.9890 0.9754 Kappa 1 1 0.9969 0.9893 1 0.9929 0.9844

[0246] Table 3 Experimental results of noise robustness of SAR / visible light / active radar one-dimensional range profile multi-mode fusion recognition

[0247] Active radar one-dimensional range Level 1 Level 2 Level 3 Level 4 Level 5 Level 6 Average signal-to-noise ratio (dB) 45 40 35 30 25 20 Aircraft Carrier 0.8667 0.8667 0.6667 0.7000 0.8667 0.9000 Destroyer 1 0.9630 1 0.8889 0.5185 0.4444 Sea surface 1 0.9556 1 0.9556 1 0.6889 OA 0.9690 0.9380 0.9225 0.8682 0.7674 0.6357 AA 0.9518 0.9037 0.8778 0.7940 0.6556 0.4574 Kappa 0.9556 0.9284 0.8889 0.8481 0.7951 0.6778

[0248] From the experimental results, it can be seen that when the noise intensity reaches level three and the average signal-to-noise ratio reaches 35dB, the overall classification accuracy can still reach 92.25%, and the average classification accuracy is 87.78%, which fully demonstrates that the active radar one-dimensional range profile signal feature extraction method is effective, the sample features are more robust to noise, and the classification efficiency of the classifier is generally good. However, with the further increase of the noise intensity, the classification index decreases significantly.

[0249] Secondly, we conducted multi-mode heterogeneous fusion recognition experiments of SAR, visible light, and passive radar. As the number of training set samples decreased from 90% to 30%, the overall classification accuracy, average classification accuracy, and Kappa coefficient showed a gradual downward trend, but the overall recognition rate was also maintained at a high level, indicating that the features used for classification were effective, the classification performance of the classifier model was high, and the consistency of the model also reached a high level. The specific experimental results are shown in Table 4.

[0250] Table 4 SAR / visible light / passive radar

[0251] Proportion 90％ 80％ 70％ 60％ 50％ 40％ 30％ Training / Testing 1020 / 114 907 / 227 793 / 341 680 / 454 567 / 567 453 / 681 340 / 794 Aircraft Carrier 0.95 0.9615 0.9339 0.9551 0.8977 0.9721 0.9837 Destroyer 0.9737 0.9615 0.9646 0.9551 0.9886 0.9012 0.8967 Sea surface 1 1 1 1 1 1 1 OA 0.9737 0.9736 0.9648 0.9692 0.9577 0.9545 0.9559 AA 0.9605 0.9603 0.9472 0.9537 0.9364 0.9317 0.9338 Kappa 0.9746 0.9744 0.9662 0.9701 0.9621 0.9578 0.9601

[0252] Similarly, in order to further explore the robustness of passive radar feature classification, Gaussian noise of different intensity levels was added to the passive radar signal in the experiment to examine the recognition rate of the algorithm under different signal-to-noise ratios. The experimental results are shown in Table 5. From the experimental results, it can be seen that when the noise intensity reaches level 4 and the average signal-to-noise ratio reaches 17dB, the overall classification accuracy can still reach 94.57%, and the average classification accuracy is 91.77%, which fully demonstrates that the passive radar signal feature extraction method is effective, the sample features are more robust to noise, and the classification performance of the classifier is generally good. However, with the further increase of the noise intensity, the classification index decreases significantly.

[0253] Table 5 SAR / visible light / passive radar multi-mode fusion recognition noise robustness experimental results

[0254] Passive radar noise intensity Level 1 Level 2 Level 3 Level 4 Level 5 Level 6 Average signal-to-noise ratio (dB) 20 19 18 17 16 15 Aircraft Carrier 1 1 1 1. 1 1 Destroyer 1 1 1 0.8704 0.5185 0.1481 Sea surface 1 1 1 1 1 1 OA 1 1 1 0.9457 0.7984 0.6434 AA 1 1 1 0.9177 0.7064 0.5015 Kappa 1 1 1 0.9568 0.8395 0.7160

[0255] It can be seen from the above experiments that the algorithm of the present invention can well complete the recognition of near-shore ship targets, and has a certain robustness to the interference of complex backgrounds, and the recognition accuracy can reach up to 100%. The high classification accuracy is due to the fact that the algorithm integrates multi-modal feature information, which can filter out useless and redundant features in various complex and uncertain situations, and increase the differences between different types of targets, thereby improving the target recognition accuracy. In addition, as the number of training samples gradually decreases, the recognition results only fluctuate slightly, which verifies that the algorithm still has the advantage of high accuracy in a small sample environment, and is suitable for high-dynamic unmanned platforms that lack measured data.

[0256] A multi-mode fusion sea surface target recognition system for a high-dynamic unmanned platform, which includes a platform sensor subsystem, a multi-mode feature processing and recognition subsystem, an information transmission control subsystem, and a remote display and control subsystem;

[0257] The platform sensor subsystem includes an optical imaging sensor module, a SAR sensor module, an active radar sensor module and a passive radar sensor module. The system is a front-end receiving device for obtaining optical images, SAR images, active radar signals and passive radar signals of sea surface targets. The signals obtained by the optical imaging sensor module, the SAR sensor module, the active radar sensor module and the passive radar sensor module provide test samples for the target recognition system; at the same time, the azimuth information of the target relative to itself is obtained, and after coordinate transformation, the absolute position information of the target is obtained, including the latitude and longitude coordinates and distance information of the target.

[0258] The multi-mode feature processing and recognition subsystem includes a preprocessing module, a feature extraction module and a sea surface target feature fusion classification module; the preprocessing module includes an image filtering processing module corresponding to the optical imaging sensor module, an image filtering processing module corresponding to the SAR sensor module, a signal sampling processing module corresponding to the active radar sensor module and a signal sampling processing module corresponding to the passive radar sensor module; the image and signal data receiving and preprocessing operation functions of the preprocessing module are realized by field programmable gate array hardware, and the processing results are stored in the onboard memory.

[0259] The image filtering processing module corresponding to the optical imaging sensor module realizes the noise filtering, distortion correction and target segmentation preprocessing functions of the optical image; the image filtering processing module corresponding to the SAR sensor module realizes the image coherent speckle noise suppression and CFAR target detection functions; the signal sampling processing module corresponding to the active radar sensor module completes the discrete sampling of the one-dimensional range image signal of the active radar, and collects the one-dimensional discrete signal of the same length according to the input requirements of the classifier model, which can be 1000 sampling points, which is convenient for feature extraction operation processing; the signal sampling processing module corresponding to the passive radar sensor module completes the discrete sampling of the one-dimensional range image signal of the passive radar, and collects the one-dimensional discrete signal of the same length according to the input requirements of the classifier model, such as 1000 sampling points, which is convenient for feature extraction operation processing.

[0260] The feature extraction module includes an image feature extraction module corresponding to the optical imaging sensor module and the SAR sensor module, which mainly completes the multi-resolution robust feature extraction and selection functions of the image target. The feature extraction is implemented by a fast sampling filtering method, and the feature selection adopts a principal component analysis (PCA) algorithm, and finally the image sample is converted into a feature vector; the feature extraction module also includes a signal feature extraction module corresponding to the active radar sensor module and the passive radar sensor module, which mainly completes the calculation of 10 feature quantities of enhanced average absolute value, average amplitude change, wavelength, maximum fractal length, root mean square, zero crossing, slope sign change, time moment, Teager-Kaiser energy operator, and skewness, and then combines them into a 10-dimensional feature vector; the feature extraction module is implemented by field programmable gate array hardware, and the feature vector is stored in an onboard flash memory.

[0261] The sea surface target feature fusion classification module mainly completes model storage, multi-feature kernel matrix operation, fusion of multi-mode heterogeneous features, recognition test, and recognition result storage functions. The sea surface target feature fusion classification module is composed of Huawei Ascend high-performance AI embedded acceleration computing unit, embedded multimedia storage card, high-speed interface device and target recognition algorithm, including the storage of trained sea surface target recognition algorithm model, feature call reading of input samples, calculation fusion of multi-mode feature kernel matrix, test classification of input target samples, and finally completes the recognition of sea surface targets, that is, the determination of target categories.

[0262] The information transmission control subsystem includes an intelligent control module and an information transmission module connected to the sea surface target feature fusion classification module of the multi-mode feature processing and identification subsystem. The information transmission control subsystem mainly completes the system internal and external transmission of target identification results and the coordinated control function; the information transmission module mainly completes the wired and wireless transmission functions of target identification data, target coordinate data and other information; wired transmission is mainly responsible for data transmission within the system; wireless transmission is mainly responsible for data transmission between the system and the ground control station; according to the information transmission distance, wireless digital data link sending and receiving equipment is used, or data link and communication relay equipment are used to achieve the corresponding functions; the intelligent control module mainly controls the normal and orderly operation of each sensor module, sea surface target feature fusion classification module, information transmission module and other modules through the transmission of control signals in the entire sea surface target identification system. The function of this module is directly completed by the onboard control device of the Ascend AI accelerator.

[0263] The remote display and control subsystem is a ground display control module. The remote display and control subsystem is set on the ground control station host, which is convenient for the training, integration and debugging of the target recognition algorithm, and also convenient for the visualization of the target recognition function; the remote display and control subsystem realizes the visualization output and positioning identification of the target in the digital map of the remote ground station through the target coordinates and target recognition information obtained wirelessly, and completes the real-time tracking and character overlay functions of the target; the subsystem is mainly composed of a general-purpose computer host, a display device, and a Geographic Information System (GIS) hardware and software device.

[0264] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A multi-mode fusion sea surface target recognition method for a highly dynamic unmanned platform, characterized in that: The following steps are included: Step 1: Preprocessing of sea surface target data. The image data is acquired through the high-dynamic unmanned platform sensor module, including image filtering preprocessing of optical images and SAR images from sea surface targets, and signal sampling preprocessing of active radar signals and passive radar signals from sea surface targets; Step 2: Extraction of sea surface target features. The four types of data, namely, optical image, SAR image, active radar signal and passive radar signal, are subjected to feature extraction respectively. For optical image and SAR image data, the fast sampling filtering method is used after preprocessing, and the PCA algorithm is used for feature extraction to convert image samples into feature vectors. For active radar signal and passive radar signal data, the feature extraction after preprocessing mainly completes the calculation of 10 feature quantities, namely, enhanced average absolute value, average amplitude change, wavelength, maximum fractal length, root mean square, zero crossing, slope sign change, time moment, Teager-Kaiser energy operator and skewness, and then combines them into a 10-dimensional feature vector. Step 3: Sea surface target identification, The modal features of the four data features of sea surface targets, namely optical images, SAR images, active radar signals and passive radar signals, are extracted and identified through a sea surface target recognition method based on multi-mode heterogeneous features fusion and multi-core machine learning, i.e., the determination of target categories. It mainly completes the functions of model storage, multi-feature kernel matrix operation, fusion of multi-mode heterogeneous features, recognition test and storage of recognition results.

2. A multi-mode fusion sea surface target recognition method for a highly dynamic unmanned platform as claimed in claim 1, characterized in that: The optical image of the sea surface target in step 1 is subjected to image filtering preprocessing, which is to use a general sea surface target detection method based on image threshold segmentation and morphology to achieve preliminary extraction and processing of the sea surface target, which specifically includes the following steps: (1) Using a certain threshold, perform threshold segmentation on the image to obtain a binary image; (2) Perform morphological processing on the binary image, including dilation operation, opening operation, and erosion operation, to obtain a binary image of the target; (3) Calculate the center of gravity of each target in the binary image and mark it; (4) Expand outward from the center of gravity, mark the target area, and obtain the target slice image.

3. A multi-mode fusion sea surface target recognition method for a highly dynamic unmanned platform as claimed in claim 2, characterized in that: The determined threshold refers to converting the original image from a color image to a grayscale image, first performing histogram equalization on the image, and then using a global threshold method to maximize the difference in variance between the foreground and background parts of the image; The dilation operation is implemented by using a structural element on the binary image. The element is a binary matrix, usually a small rectangle, circle or ellipse. In the dilation operation, for each pixel, if there is a white pixel overlapping with the structural element, the pixel is also set to white. The effect of the operation is to make the target object larger, the connectivity stronger, fill holes, and increase the perimeter of the object; the opening operation is a combination of corrosion and then dilation, which can be used to remove small holes, slender objects and noise points, make the edge of the object smooth, and will not destroy the connectivity of the object; the erosion operation is the process of reducing the shape of the target object in the image, which is implemented by using a structural element on the binary image. In the erosion operation, for each pixel, if the structural element is completely contained in the black pixel, the pixel is set to black. The effect of the operation is to make the target object smaller, separate connected objects, remove noise points outside the object and reduce the perimeter; The centroid of each target in the binary image is determined by calculating the total moment of the target pixel set. The specific calculation method is as follows: 1) Extract the target pixel set in the binary image, assuming that the target pixel set is S; 2) Calculate the total moment of the target pixel set S, denoted as M00. The total moment can be calculated by the secondary moment in the horizontal and vertical directions: Where I(x,y) is the pixel value in the binary image. If (x,y)∈S, then I(x,y)=1, otherwise I(x,y)=0; 3) Calculate the numerator of the total moment, which represents the secondary moment of the target in the horizontal and vertical directions, denoted as M10 and M01 respectively. 4) Calculate the coordinates of the center of gravity (xc, yc), x c =M 10 M 00 , the coordinates of the centroid (xc, yc) represent the center position of the target in the binary image. In actual programming, these calculations can be simplified using functions provided by libraries such as NumPy.

4. The multi-mode fusion sea surface target recognition method for a highly dynamic unmanned platform as claimed in claim 1, characterized in that: The SAR image of the sea surface target in step 1 is subjected to image filtering preprocessing, which is to realize the preliminary extraction and processing of the sea surface target by using an image segmentation algorithm based on a constant false alarm rate (CFAR), and specifically comprises the following steps: (1) Preprocessing the image to suppress noise, including performing a logarithmic transformation on the image to convert the multiplicative speckle noise into additive noise; using a general template filtering method to perform a convolution operation on the image and the template. Common template filtering methods include Gaussian filtering and mean filtering; performing an exponential operation on the image obtained after template filtering to transform it back to the image in the original image space; (2) The target area is obtained by CFAR detection of the preprocessed segmentation, that is, the region of interest (ROI). The key to CFAR detection technology is to determine the adaptive threshold. Assuming that p(x) is the probability density function of the radar background clutter distribution model, let Obviously, F(x) is an increasing function on [0,+∞). By solving the equation, Get the threshold I, where p fa For the false alarm rate, the traditional CFAR method is mainly a method for processing one-dimensional radar signals. When used for SAR images, a sliding window method is usually used to traverse the image: the central pixel of the sliding window is used as the test point (Cell Under Test, CUT), and the pixels around the window are used to estimate the clutter distribution, completing target detection at a fixed constant false alarm rate.

5. The multi-mode fusion sea surface target recognition method for a highly dynamic unmanned platform as claimed in claim 1, characterized in that: The feature extraction of the optical image and SAR image data in the step 2 adopts a fast and robust feature extraction method for sea surface target images, including a method for extracting local multi-resolution features, a method for designing and implementing an image filter kernel, and a method for removing redundant features of image targets; The method for extracting local multi-resolution features is specifically implemented in the following steps: (1) Use a non-uniform sampling filter kernel to perform non-uniform filtering sampling on the target in each band. The sampling is uniform within the fovea, and the maximum sampling rate is 1, which is the original resolution of the image. In the peripheral area, the sampling rate gradually decreases with the distance from the center point. In this way, the input information is greatly compressed, which greatly reduces the difficulty of storage and processing. (2) After extracting feature maps in different scale spaces, the central-peripheral difference strategy is imitated and the interpolation difference method is used to perform peripheral difference transformation on the multi-scale space to filter out background noise and amplify the contour information of the detected target; (3) The contrast scale space feature map is stretched row by row to obtain a one-dimensional feature vector, and the PCA algorithm is used to reduce its dimensionality to obtain the final multi-scale feature; The design and implementation method of the image filter kernel draws on the "Center-Surround Antagonism, CSA" mechanism of the biological visual system. By simulating the change of sampling rate with distance in the visual attention mechanism, the present invention constructs a non-uniform sampling filtering method, in which uniform sampling with a sampling rate of 1 is performed within the fovea, that is, the image is not compressed; outside the fovea, the sampling rate decreases as the distance from the sampling point to the center point increases, and the farther away from the fovea, the smaller the sampling rate; In summary, the filter kernel is set to: The image size is N×N, where N is usually an integer power of 2, and δ represents the radius of the fovea. When it is within δ, it is uniformly sampled. Based on the discrete wavelet transform (DWT) method of discretizing the scale according to the power series on the scale-shift phase plane, the points on the phase plane are non-uniform. This aspect also uses the DWT idea to set the scale parameter. For a fixed scale parameter m, the corresponding sampling rate is γα -m ,make: h(x)=ca -m (3) Since m is an integer power of 2, m=-g(log2h(x)) (4) Among them, g(y) is the rounding down of y. For the filter kernel, the scale parameter m is the same within the same concentric circle. In the central area, the scale parameter m=0; outside the central area, the scale parameter increases with the increase of distance; The method for removing redundant features of image targets refers to extracting features of different scales, removing redundant information between different scales, so as to more effectively detect stable key points and strengthen the spatial texture features of the image, that is, performing differential processing on feature images obtained at different scales, and outputting the layered results as features; wherein the fine-scale feature map represents the output features of the light-giving center cells, and the coarse-scale feature map represents the output features of the light-removing center cells. By performing differential processing on the feature maps between different scales, the texture features of the image can be strengthened, and the image noise can be filtered, so as to fully utilize the image scale information; In order to improve the classification speed and reduce the feature dimension, the correlation between eigenvectors must be eliminated; use principal component analysis (PCA) to whiten the data, calculate the eigenvalues ​​and eigenvectors of the correlation coefficient matrix of the sample, sort the eigenvalues ​​in descending order, and adjust the eigenvectors accordingly; calculate the contribution rate of each eigenvalue, extract the principal component eigenvector according to the given extraction rate, and project the sample on the extracted eigenvector to extract the features.

6. The multi-mode fusion sea surface target recognition method for a highly dynamic unmanned platform as claimed in claim 1, characterized in that: The feature extraction of active radar signal and passive radar signal data in step 2 adopts the general feature extraction method of active and passive radar one-dimensional range profile. For active radar, the characteristics of target range profile length, number of scattering centers, range profile symmetry, dispersion, and range profile entropy can reflect the characteristics of the target to a certain extent; while for passive radar, the characteristic parameters that can be used for target identification include pulse width, pulse arrival time, pulse repetition period or repetition frequency, pulse arrival angle, pulse repetition interval (PRI) variation, carrier frequency variation, etc. Affected by environmental factors, it is not possible to extract all this information in each frame of active and passive radar signals, or this characteristic information may not be able to accurately reflect the type of target, but for both sensors, the target can be described in the form of a one-dimensional range profile signal. In order to simplify the feature extraction process of active and passive radar signals, the radar one-dimensional range profile signal is treated as a conventional digital signal. Its most basic features include wavelength and mean absolute value (MAV). Based on this, 10 simple and effective signal feature parameters are extracted. Finally, these 10 feature parameters constitute the feature vector of the target, which has a small amount of calculation and is more general. Let x i represents the i-th component value of the signal x, L represents the total length of the signal, and its mathematical definition is as follows: 1) Enhanced Mean absolute value (EMAV) 2) Average Amplitude Change (AAC) 3) Waveform Length (WL) 4) Maximum Fractal Length (MFL) 5) Root Mean Square (RMS) 6) Zero Crossing (ZC) 7) Slope Sign Change (SSC) 8) Temporal Moment (TM) 9) Teager–Kaiser Energy Operator (Log Teager Kaiser Energy Operator, LTKEO) 10) Skew 7. The multi-mode fusion sea surface target recognition method for a highly dynamic unmanned platform as claimed in claim 1, characterized in that: The method for sea surface target recognition by multi-modal heterogeneous feature fusion multi-core machine learning in step 3 includes designing a typical sea surface target multi-model sample library, designing a general sea surface target recognition framework by fusion of multi-modal features and multi-core learning methods, a target recognition algorithm based on multi-modal heterogeneous feature fusion and multi-core machine learning, and evaluation of experimental results of multi-modal heterogeneous feature fusion multi-core machine learning; The design of the typical sea surface target multi-sample library includes obtaining the target's visible light image, SAR image, active radar one-dimensional range image, and passive radar one-dimensional range image to construct the multi-sample library. The visible light image sample is a target slice intercepted from the public Google earth high-definition image, with a size of 512×512; the SAR target image corresponding to the visible light image is simulated and generated by the image style transfer method based on the Generative Adversarial Networks (GANs). The source domain target image is the collected visible light image, the training image is a public sea surface SAR image data set, and the target domain image is the generated SAR image target slice, which has the same size as the visible light image, also 512×512; the active radar one-dimensional range image data is simulated and generated by a linear frequency modulation signal with a bandwidth of 30MHz. The effective operating frequency range of the passive radar is simulated and generated by four frequency bands of 0.85 / 0.94GHz, 0.95 / 1.2GHz, 2.9 / 3.1GHz, and 5.4 / 5.8GHz; The sea surface target recognition framework that integrates the multi-modal features and multi-core learning methods is designed in a universal way. It extracts multi-scale features of the target based on a variety of missile-borne sensors, organically integrates the multi-modal features with the multi-core in the multi-core learning framework, obtains the fusion kernel matrix by mapping the samples to the regenerated Hilbert space, and then optimizes the multi-core combination coefficients through iterative training, and finally obtains the trained multi-scale multi-core classifier. By inputting the test samples, the target is finally recognized. In order to capture the scale variability of complex targets, the target information features obtained by different sensors are extracted and visible light image kernels, SAR image kernels, active radar data kernels and passive radar data kernels at different scales are constructed, and weighted combinations are performed on them. The specific form is as follows: Visible light image: SAR imagery: Active Radar: Passive Radar: in represents the k-th scale feature extracted from the i-th target, δ k Represents the bandwidth of the Gaussian kernel function. Then, the above four base kernels are substituted into the multi-scale multi-kernel learning framework, and then the multi-scale kernel classifier is trained based on this to realize the fusion classification of multi-source heterogeneous features; The target recognition algorithm based on multi-mode heterogeneous feature fusion and multi-core machine learning is based on the principle of multi-mode heterogeneous feature fusion multi-scale kernel classifier target recognition framework, and specifically includes the following steps: (1) Preprocessing: Slice, segment, translate, rotate, scale, and normalize visible light images and SAR images, and divide the samples into training sets and test sets; (2) Feature extraction: Based on the extraction of multi-resolution robust features of visible light images and SAR images, and the extraction of feature parameters of active and passive radar data, feature vectors representing various types of targets are obtained, and a multi-source feature library of various types of targets is established; (3) Feature fusion: The base kernel matrix is ​​calculated on each feature library according to the kernel functions of different scales, and the synthetic kernel matrix is ​​calculated by linear weighting; (4) Learning combination weights: Use the multi-core learning algorithm to iteratively train the SVM classifier and update the combination weights until the iteration stop condition is met; (5) Target identification: import test set features and use the trained classifier for identification; The target recognition algorithm based on multi-mode heterogeneous feature fusion and multi-core machine learning uses the SimpleMKL algorithm of the simple multi-core learning method as the multi-core classifier. By adding constraints to the sparse kernel combination weights, a weighted 2-norm regularization formula is designed to solve the MKL problem. The classifier has a simple principle and is efficient to implement. In the SimpleMKL algorithm, the synthetic kernel model that satisfies Mercer's theorem is a weighted convex sum of several simple basic kernel functions. The dot product operation in the complex feature space is replaced by the synthetic kernel function; In the formula, K m is a positive definite kernel function, and a definition of <·,·> m The reproducing kernel Hilbert space H of the dot product operation m It can be simply a classical kernel with different parameters, such as linear kernel, polynomial kernel, Gaussian kernel, etc. M is the number of basis kernels, d m is the weight coefficient of the corresponding basis kernel function. Then, by adding constraints to the weights of the sparse kernel combination, a weighted l2-norm regularized convex and differentiable primitive objective formula is designed to solve the MKL problem, which can effectively obtain an approximate primitive solution, as shown in Equation (16). in, For J(d), the derivative of the Lagrangian function is set to zero according to the original variable, and its dual problem is obtained as follows: Note that, let K(x i ,x j )replace The above formula is actually the dual form of the standard support vector machine, which can be calculated by any standard support vector machine solver; therefore, under this objective function framework, the MKL problem is transformed into learning the kernel function weight coefficient d through a single optimization problem in a supervised environment m and the kernel coefficient α i ; SimpleMKL uses a two-step method to iterate the standard SVM model. In each iteration, the Lagrangian operator α of the SVM is first fixed and the kernel weight d is updated. Then the updated kernel weight d is fixed and the Lagrangian operator α of the SVM is updated at the same time until the stopping criterion is met. In this way, the hot-start-based two-step training method can reduce the complexity of the standard support vector machine solver. The stopping criterion can be based on the duality gap, the KKT condition, the rate of change of d between two iterations, or simply based on the maximum number of iterations. The multi-core construction method used by SimpleMKL is also a simple linear combination of classic kernels with different parameters. In order to give full play to the effectiveness of multi-resolution features of image targets, the present invention integrates multi-core learning and scale analysis methods, and designs an automatic target recognition method based on the organic fusion of local multi-resolution feature extraction and multi-scale kernels. This method is based on the idea of ​​integrating multi-level features, uses an SVM model based on a multi-scale kernel method, and uses kernel functions of different scales to map image features of different levels separately, and then synthesizes them, which can greatly improve the discriminability of features and improve classification accuracy. The calculation of the above formula does not require the knowledge of the input space D = {(x i ,y i )}(i=1,2,L,N) and feature mapping function Φ(·), the feature set S is implicitly mapped to the multi-scale multi-kernel space through the kernel technique, which reduces the amount of calculation; before feature fusion, the basic kernel function K in the multi-scale feature fusion process is used m (·,·), usually a basis kernel function with multi-scale representation capability is selected, such as the Gaussian radial basis kernel, as shown below: where σ1<···<σ M When σ is small, the classification effect is good for samples with drastic changes, while when σ is large, the classification effect is good for samples with gentle changes, and better generalization performance can be obtained. In specific implementation, the value of σ can refer to the law of scale change in wavelet transform, and σ can be defined by the following formula: σ i =2 i σ,i=0,1,2L (21); The evaluation of the experimental results of multi-mode heterogeneous feature fusion multi-core machine learning uses three indicators as evaluation criteria: overall classification accuracy (OA), average classification accuracy (AA) and Kappa coefficient. The overall classification accuracy refers to the ratio of the number of correctly classified samples to the total number of samples. The average classification accuracy refers to the average classification accuracy of each type of sample. Let C be the number of categories, N be the number of samples, and N i is the number of samples in the i-th category, M is the confusion matrix obtained from the experiment, and the calculation methods of OA, AA and Kappa coefficients are as follows:

Images