A multi-source ship target identification method based on two-stage collaborative fusion

Abstract

The application discloses a kind of multi-source ship target identification methods based on two-stage cooperative fusion, it is related to signal processing technical field, first with the optical camera on the plane and radar sensor acquisition target sea area data, including original optical image and radar image data, then using Gaussian filter to the data collected is preprocessed;Next, using deep learning algorithm is carried out sea-land segmentation, constructs loss function and optimizes network parameter to obtain segmentation mask, then according to the mask result removes land background area, obtains the optical and radar image after processing;Afterwards, using the registration method that improved SURF feature and graph cut algorithm are combined, by screening feature point pair and solving minimum value of energy function, optimal transformation parameter is obtained, image registration is realized;For the image after registration, based on adaptive fusion rule to fuse image;Finally, corresponding feature is extracted by hand feature extraction and based on deep learning feature extraction and is fused to obtain final target recognition result.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of signal processing, in particular to a multi-source ship target recognition method based on two-stage cooperative fusion. BACKGROUND

[0002] Multi-source refers to the use of data from different types of sensors for information fusion. Common multi-source data includes optical images, radar images, etc. Optical images provide high-resolution image information, which can clearly present the texture, color and details of objects; radar images obtain target information through electromagnetic wave reflection, especially in adverse weather conditions such as fog, night, rain and snow, etc., which has strong robustness. By fusing these two kinds of data, the limitations of single data source can be made up, and the accuracy and robustness of target recognition can be improved.

[0003] Multi-source ship target recognition refers to the technology of integrating multiple information sources such as optical images and radar images to comprehensively analyze and recognize ship targets. Ship target recognition is of great significance in maritime surveillance, maritime safety, military reconnaissance and other fields. Optical images and radar images each have different advantages and disadvantages. Optical images perform well in visible light, but are greatly affected by environmental conditions; while radar images can provide stable target detection capability in complex environments, but have relatively low resolution. By fusing optical images and radar images, the advantages of both data sources can be fully utilized, and more comprehensive and accurate ship target recognition results can be provided under different scenarios and conditions. For example, in daylight or good visibility conditions, optical images can clearly present the shape characteristics of ships, while in night or poor visibility conditions, radar images can obtain the outline information of ships through reflected waves, thereby improving the recognition accuracy and adaptability.

[0004] Multi-source ship target recognition technology has great application potential, but still faces many technical challenges in practical application. First, optical images and radar images come from different sensors, with different imaging mechanisms and data characteristics. Optical images usually have high spatial resolution and rich texture information, but they are easily affected by weather, lighting, fog, and other environmental factors, leading to loss or blur of target information. Radar images, on the other hand, detect targets through electromagnetic wave reflection, have strong anti-interference ability, and can penetrate clouds and night environment, but their resolution is relatively low and they can only provide contour information of the target. Therefore, how to overcome the differences and complementarity of the two data sources while ensuring recognition accuracy is the primary challenge of multi-source ship target recognition. Second, image data fusion is one of the technical difficulties. Optical images and radar images not only differ in spatial resolution, but also differ in imaging principles and data formats. For example, optical images are based on visible light reflection, while radar images rely on electromagnetic wave reflection characteristics, making the image fusion process complex. Common image fusion methods include pixel-level fusion, feature-level fusion, and decision-level fusion, but different methods have different fusion effects and adapt to different scenarios. Pixel-level fusion performs well in detail recovery, but is easily affected by noise; feature-level fusion can extract more recognizable features, but may lose some detailed information during fusion; and decision-level fusion is a comprehensive process after target recognition, but this method requires high accuracy of the recognition model. In addition, data synchronization and alignment are also important challenges in multi-source image fusion. Different types of sensors may have inconsistencies in image data due to differences in time, space, and perspective, which can affect the quality of the fusion result. Therefore, how to accurately align images from different sources and effectively fuse data is the key to improving the performance of multi-source ship target recognition. Finally, feature extraction and selection also pose challenges to multi-source ship target recognition technology. In the fused image, there is a lot of redundant information and noise, and traditional manual feature extraction methods may not meet the requirements of high precision and high real-time performance for accurate ship target recognition. Although deep learning methods have shown strong capabilities in feature extraction, how to efficiently extract useful features and optimize them is still a problem to be solved.

[0005] In summary, although multi-source ship target recognition technology has great potential in improving target recognition accuracy and expanding application scope, due to technical problems such as sensor differences, image fusion, data synchronization, and feature selection, further research and innovation are needed to improve its application effect and stability in complex environments. SUMMARY

[0006] The purpose of this invention is to provide a multi-source ship target recognition method based on two-level collaborative fusion, which solves the problems of poor accuracy and low efficiency of target recognition results caused by the influence of sensor differences, image fusion, data synchronization and feature selection in the existing technology.

[0007] To achieve the above objectives, this invention provides a multi-source ship target identification method based on two-level collaborative fusion, comprising the following steps:

[0008] Step 1: Using optical cameras and radar sensors mounted on the aircraft, simultaneously collect data on the target sea area to obtain raw optical image data. and radar image data ,in The pixel coordinates in the image plane are represented by a Gaussian filter, which is then applied to obtain the filtered optical image. and filtered radar images ;

[0009] Step 2: Use deep learning algorithms to perform land-sea segmentation on the preprocessed optical and radar images, respectively, to obtain the optical and radar images after zeroing out the land areas of the original images, which are respectively... , ;

[0010] Step 3: A registration method based on a combination of improved SURF (Speeded Up Robust Features) and graph cut algorithms is used to... and Perform image registration to obtain the registered optical image. and radar images ;

[0011] Step 4: For the registered optical image and radar images An adaptive fusion rule based on regional energy, gradient, and structure tensor information is used to obtain the fused image. ;

[0012] Step 5: Extract corresponding features from the fused image obtained in Step 4 using manual feature extraction and deep learning-based feature extraction. Then, fuse the extracted features to obtain the final target recognition result.

[0013] The preferred formula for Gaussian filtering is as follows:

[0014] ;

[0015] in It is a Gaussian kernel, defined as:

[0016] ;

[0017] In the formula, represents the filtered image, represents the original image, represents the standard deviation of the Gaussian kernel, controlling the width and height of the kernel, represents the horizontal coordinate increment of the input data point, represents the vertical coordinate increment of the input data point.

[0018] Preferably, a deep learning algorithm is used in step 2 to perform sea-land segmentation on the preprocessed optical image and radar image respectively, and the process of obtaining the results of the optical image and radar image with the land area of the original image set to zero is as follows:

[0019] S21, construct a loss function Optimize the sea-land segmentation network parameters to obtain the sea-land segmentation mask; wherein the loss function The expression is as follows:

[0020] ;

[0021] In the formula, is the number of samples, is the number of categories (sea and land), is the true label of the sample belongs to the category , is the probability value of the model predicting that the sample belongs to the category , is the weight coefficient of the category , which is adaptively adjusted according to the proportion of each category in the sample, for example, if the sea pixel ratio is , the land pixel ratio is , then the sea category weight , and the land category weight ;

[0022] S22, according to the results of the segmentation mask, only keep the pixel part corresponding to the coastal and sea surface area in the image, wherein sea and land are represented by specific pixel values, sea area is 1, land area is 0, remove the land background area, and obtain the results of the optical image and radar image with the land area of the original image set to zero, the expression is as follows:

[0023] ;

[0024] ;

[0025] In the formula, and respectively, are the results of the optical image and the radar image after the land region of the original image is set to zero, is a sea-land segmentation mask representing the optical image, is a sea-land segmentation mask representing the radar image. is a sea-land segmentation mask representing the optical image, is a sea-land segmentation mask representing the optical image, is a sea-land segmentation mask representing the radar image, is a sea-land segmentation mask representing the radar image.

[0026] Preferably, a registration method based on improved SURF features and graph cut algorithm is used in step 3 to and perform image registration, and obtain the registered optical image and the radar image The process is as follows:

[0027] S31, the feature points are extracted by using the improved SURF algorithm, and in the feature descriptor generation stage, the texture features based on local binary pattern are introduced to enhance the description ability of the feature points and the robustness to light changes; for the feature points in the optical image and the feature points in the radar image, the reliable feature point pairs are screened out by using the feature matching algorithm based on the bidirectional nearest neighbor distance ratio (BB-NNDR);

[0028] S32, the graph cut algorithm is used to solve the minimum value of the energy function, and the optimal transformation parameter is obtained, so that the radar image can be accurately transformed to the same coordinate system as the optical image. The energy function is composed of data items and smoothing items The formula is as follows:

[0029] ;

[0030] ;

[0031] ;

[0032] ;

[0033] ;

[0034] wherein, is the set of all pixel points in the image, is the set of matched feature point pairs, is a set of adjacent pixel pairs, is a transformation matrix for mapping points in the radar image to the optical image coordinate system, is a balance coefficient, is a set of adjacent pixel pairs and the distance between, is a set of adjacent pixel pairs, is a parameter for controlling the degree of smoothing, is a similarity measure function based on conditional probability, is a similarity measure, denotes a smoothing measure between adjacent pixel pairs;

[0035] S33, find the correspondence between the images by S31 and S32, realize accurate registration, and obtain the registered optical image and radar image .

[0036] Preferably, for the registered optical image and radar image , an adaptive fusion rule based on regional energy, gradient and structure tensor information is used to obtain the fused image The process is as follows:

[0037] S41, calculate the regional energy , gradient amplitude and structure tensor eigenvalue in the local window centered on , , , , the expressions are as follows:

[0038] ;

[0039] ;

[0040] ;

[0041] Through eigenvalue decomposition of the structure tensor matrix , get and ;

[0042] S42, calculate the weight coefficient of the optical image and the weight coefficient of the radar image, the expressions are as follows:

[0043] ;

[0044] ;

[0045] wherein, represents the local energy of the image in a local window centered at , and are the horizontal and vertical pixel coordinate values in the local window , represents the gradient magnitude of the image at the coordinate , and are the partial derivatives of the image in the and directions, and the subscripts opt and rad represent optical and radar, respectively; represents the eigenvalue of the structure tensor ;

[0046] S43, obtaining the fused image , expressed as follows:

[0047] .

[0048] Preferably, the manual feature extraction extracts the contour of the ship by using an edge detection algorithm based on a Canny operator, accurately determines the boundary of the ship target by performing morphological processing (such as inflation and corrosion operation) and contour tracking algorithm on the edge image, and then calculates multiple geometric shape features of the ship, including the aspect ratio, the perimeter, and the area change rate.

[0049] Preferably, the deep learning feature extraction adopts an architecture combining a multi-branch convolutional neural network (M-CNN) and an attention mechanism; the multi-branch convolutional neural network processes different scales and feature levels of the fused image, each branch containing a plurality of convolutional layers, pooling layers and residual connection modules, and extracts rich image features; the attention mechanism learns the importance weight of different regions in the image, so that the model focuses on the ship target region and suppresses the interference of background noise; in the training stage, a large number of labeled ship target fusion image samples (including the above-mentioned manual features as auxiliary input) are used for training; the features extracted by the manual feature extraction and the image features extracted by each branch network are spliced and fused before the fully connected layer, and finally serve as the comprehensive input feature vector of the model.

[0050] Therefore, the application has the following beneficial effects by adopting the above-mentioned multi-source ship target recognition method based on two-level collaborative fusion:

[0051] (1) By means of scene segmentation technology, the effective part containing only coastal and sea surface area is accurately retained, and irrelevant background information such as land is discarded, which greatly reduces the data processing amount, makes the computing resources be used efficiently, and significantly improves the overall processing efficiency, thereby laying a good foundation for subsequent target recognition work;

[0052] (2) For the ship target, the features such as geometric shape and length-width ratio which are easy to obtain manually and simple to calculate are innovatively extracted first, and these features are used as powerful auxiliary input of the deep learning model; so that the deep learning model does not need to spend a lot of time and resources to repeatedly extract these simple features, but can concentrate more energy on mining deep and more discriminative feature information, effectively reducing the training and recognition time of the model, and greatly improving the recognition efficiency;

[0053] (3) A two-level collaborative fusion strategy of pixel level and feature level is adopted, the original information of optical image and radar image is fully integrated in pixel level fusion, and the feature combination is further optimized in feature level fusion, which complement each other, so that the feature description of the ship target is more comprehensive and accurate, thereby comprehensively improving the accuracy of target recognition, and providing a more reliable technical means for ship target recognition in complex marine environment.

[0054] The technical solutions of the present application will be further described in detail below with the help of drawings and embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 The present application is a kind of two-level collaborative fusion based on multi-source ship target recognition method. DETAILED DESCRIPTION

[0056] The following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed application, but only represents selected embodiments of the application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the present application.

[0057] Please refer to Figure 1 , a kind of two-level collaborative fusion based on multi-source ship target recognition method, comprising the following steps:

[0058] Step 1, use optical camera and radar sensor carried on the airplane to synchronously collect data of target sea area, and obtain original optical image data and radar image data , wherein represents the pixel coordinates in the image plane, then the Gaussian filter preprocessing is carried out, and the filtered optical image is obtained and filtered radar image ; the formula of Gaussian filtering is as follows:

[0059] ;

[0060] wherein is a Gaussian kernel, defined as:

[0061] ;

[0062] in the formula, denotes a filtered image, denotes an original image, denotes a standard deviation of a Gaussian kernel, controlling the width and height of the kernel, denotes a horizontal coordinate increment of an input data point, denotes a vertical coordinate increment of an input data point.

[0063] Step 2, use a deep learning algorithm to respectively perform sea-land segmentation on the preprocessed optical image and radar image, to obtain the results of the optical image and the radar image with the land area of the original image being zeroed, respectively as , ; the specific process is as follows:

[0064] S21, construct a loss function Optimize the sea-land segmentation network parameters to obtain a sea-land segmentation mask; the loss function adopts a combination of a dice coefficient loss function and a cross-entropy loss function to optimize the network parameters by deep learning, to solve the problem of imbalance between sea and land sample categories; wherein, the expression of the loss function is as follows:

[0065] ;

[0066] in the formula, is the number of samples, is the number of categories (two categories of sea and land), is the true label of the sample belonging to the category , is the probability value of the model predicting that the sample belongs to the category , is the weight coefficient of the category , which is adaptively adjusted according to the proportion of each category in the sample, for example, if the proportion of sea pixels is and the proportion of land pixels is , then the weight of the sea category is and the weight of the land category is ;

[0067] S22, according to the result of the segmentation mask, only the pixel part corresponding to the coastal and sea surface area in the image is retained, wherein the sea and land are respectively represented by a specific pixel value, the sea area is 1, and the land area is 0, the land background area is removed, and the results of the optical image and the radar image with the land area of the original image being zeroed are obtained, and the expression is as follows:

[0068]

[0069]

[0070] In the formula, and are the results of the optical image and the radar image with the land area of the original image being zeroed respectively, represents the sea-land segmentation mask of the optical image, represents the sea-land segmentation mask of the radar image. represents that the pixel value of the point with coordinates (x, y) in the sea-land segmentation mask of the optical image is 1, represents that the pixel value of the point with coordinates (x, y) in the sea-land segmentation mask of the optical image is 0, represents that the pixel value of the point with coordinates (x, y) in the sea-land segmentation mask of the radar image is 1, represents that the pixel value of the point with coordinates (x, y) in the sea-land segmentation mask of the radar image is 0.

[0071] This step greatly reduces the data amount and computational complexity of subsequent processing, and excludes the interference of the land background, thereby laying a foundation for the subsequent accurate target recognition.

[0072] Step 3, considering the imaging principle difference between the optical image and the radar image and the platform motion and other factors, high-precision registration needs to be performed on the two before fusion; an image registration method based on the combination of an improved SURF (Speeded Up Robust Features) feature and a graph cut algorithm is adopted to perform image registration on and , to obtain the registered optical image and the registered radar image ; the specific process is as follows:

[0073] S31, the improved SURF algorithm is used to extract feature points, in the feature descriptor generation stage, the texture feature based on the local binary pattern is introduced, the description ability of the feature points and the robustness to light changes are enhanced, and the feature points in the optical image and the feature points in the radar image are screened out by a feature matching algorithm based on the bidirectional nearest neighbor distance ratio (BB-NNDR) to obtain reliable feature point pairs;

[0074] ​​S32, the minimum value of the energy function is solved by using the graph cut algorithm, and the optimal transformation parameter is obtained, so that the radar image can be accurately transformed into the same coordinate system as the optical image. The energy function consists of data items and smooth items , and the formula is as follows:

[0075] ;

[0076] ;

[0077] ;

[0078] ;

[0079] ;

[0080] wherein, is the set of all pixel points in the image, is the set of matched feature point pairs, is the set of adjacent pixel point pairs, is the transformation matrix for mapping the points in the radar image to the optical image coordinate system, is the balance coefficient, is the distance between adjacent pixel point pairs and , is the set of adjacent pixel point pairs, is a parameter for controlling the smoothness degree, is a similarity measure function based on conditional probability, is a similarity measure, indicates a smoothness measure between adjacent pixel point pairs;

[0081] S33, the corresponding relationship between the images is found through S31 and S32, accurate registration is realized, and the registered optical image and radar image are obtained.

[0082] Step 4, for the registered optical image and radar image , an adaptive fusion rule based on regional energy, gradient and structure tensor information is used to obtain the fused image ; the specific process is as follows:

[0083] S41, the regional energy , gradient amplitude and structure tensor information in the local window centered on the image are calculated and structural tensor eigenvalue , , , the expression is as follows:

[0084] ;

[0085] ;

[0086] ;

[0087] By performing eigenvalue decomposition on the structural tensor matrix , the and are obtained;

[0088] S42, calculate the weight coefficient of the optical image and the weight coefficient of the radar image , the expression is as follows:

[0089] ;

[0090] ;

[0091] In the formula, represents the local energy of the image in the local window centered at , and are the horizontal and vertical coordinate values of the pixel points in the local window , represents the gradient amplitude of the image at the coordinate , and are the partial derivatives of the image in the and directions, and the subscripts opt and rad represent optical and radar, respectively; represents the eigenvalue of the structural tensor ;

[0092] S43, obtain the fused image , the expression is as follows:

[0093] .

[0094] This adaptive fusion rule can dynamically allocate weights according to the local features of the image, so that better fusion results can be obtained in both target areas (such as ship parts) and background areas (such as sea surface), improving the contrast and clarity of the target and background, and enhancing the recognizability of the image.

[0095] Step 5, corresponding features are extracted from the fused image obtained in step 4 through manual feature extraction and deep learning-based feature extraction, and then the extracted features are fused to obtain the final target recognition result; for the image fused at the pixel level, first, manual feature extraction is performed to obtain some intuitive and easy-to-calculate ship target features, which can provide important prior information for the subsequent deep learning model and reduce the learning burden and calculation amount of the model. Among them, the manual feature extraction uses an edge detection algorithm based on the Canny operator to extract the contour of the ship, accurately determines the boundary of the ship target through morphological processing (such as inflation, corrosion operation) and contour tracking algorithm on the edge image, and then calculates multiple geometric shape features of the ship, including the aspect ratio, perimeter and area change rate. The deep learning feature extraction adopts the architecture combining multi-branch convolutional neural network (M-CNN) and attention mechanism; the multi-branch convolutional neural network processes different scales and feature levels of the fused image, each branch contains several convolutional layers, pooling layers and residual connection modules, and extracts rich image features; the attention mechanism learns the importance weight of different regions in the image, so that the model focuses on the ship target area and suppresses the interference of background noise; in the training stage, a large number of labeled ship target fusion image samples (including the above-mentioned manual features as auxiliary input) are used for training; the features extracted by the manual feature extraction and the image features extracted by each branch network are spliced and fused before the fully connected layer, and finally used as the comprehensive input feature vector of the model.

[0096] Therefore, the application adopts the above-mentioned multi-source ship target recognition method based on two-level collaborative fusion, and combines multiple steps and technologies to achieve the final target recognition result; first, the optical camera and radar sensor on the aircraft are used to collect target sea area data, including original optical image and radar image data, and then Gaussian filtering is used for preprocessing the collected data. Next, a deep learning algorithm is used for sea-land segmentation, a loss function is constructed to optimize network parameters to obtain a segmentation mask, and then the land background area is removed according to the mask result to obtain the processed optical and radar images. Then, an improved SURF feature and graph cut algorithm combined registration method is used to extract feature points and enhance their description ability, and the optimal transformation parameters are obtained by screening feature point pairs and solving the minimum value of the energy function, so as to realize image registration. For the registered image, an adaptive fusion rule based on regional energy, gradient and structure tensor information is used to fuse the image. On the basis of the fused image, on the one hand, manual feature extraction is performed through algorithms such as edge detection based on the Canny operator, and on the other hand, a deep learning method combining multi-branch convolutional neural network and attention mechanism is used to extract features, and finally the manual features and deep learning features are fused to obtain the final target recognition result.

[0097] It should be pointed out finally that the above embodiments are only used to illustrate the technical solutions of the present application but not to limit it, and although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can still be modified or replaced equivalently, and these modifications or equivalent replacements should not make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present application.

Claims

1. A multi-source ship target identification method based on two-stage collaborative fusion, characterized in that, Includes the following steps: Step 1, using optical camera and radar sensor mounted on the aircraft, synchronous data acquisition of the target sea area, to obtain the original optical image data and radar image data , wherein represent the pixel coordinates in the image plane, and then Gaussian filtering is performed on the pre-processing to obtain the filtered optical image and filtered radar image ; Step 2, respectively, use deep learning algorithm to sea-land segmentation on the pre-processed optical image and radar image, obtain the results of optical image and radar image after land area is zeroed on the original image, respectively , ; The process of obtaining the optical and radar images after zeroing out the land areas of the original image is as follows: S21, constructing a loss function The sea-land segmentation network parameters are optimized to obtain a sea-land segmentation mask; wherein, the loss function The expression of the loss function is as follows: ； In the formula, is the number of samples, is the number of categories, is the sample belongs to the category The true label is, is the model prediction sample belongs to the category The probability value is, is the weight coefficient of the category According to the proportion of each category in the sample, it is adaptively adjusted; S22. Based on the segmentation mask results, only the pixel portions corresponding to the coastal and sea surface regions in the image are retained, where the sea and land are represented by specific pixel values, with the sea area being 1 and the land area being 0. The land background region is removed to obtain the optical image and radar image after setting the land area of ​​the original image to zero, as shown in the following expression: ； ； wherein, and respectively are the results of the optical and radar images with the original image land areas zeroed out, denotes the sea-land segmentation mask for the optical image, denotes the sea-land segmentation mask for the radar image; denotes that the pixel value of the point with coordinates (x, y) in the sea-land segmentation mask for the optical image is 1, denotes that the pixel value of the point with coordinates (x, y) in the sea-land segmentation mask for the optical image is 0, denotes that the pixel value of the point with coordinates (x, y) in the sea-land segmentation mask for the radar image is 1, denotes that the pixel value of the point with coordinates (x, y) in the sea-land segmentation mask for the radar image is 0. Step 3: A registration method based on a combination of improved SURF features and graph cut algorithm is used to... and Perform image registration to obtain the registered optical image. and radar images ; Obtain the registered optical image and radar images The process is as follows: S31. Feature points are extracted using an improved SURF algorithm. During the feature descriptor generation stage, texture features based on local binary patterns are introduced; for feature points in optical images... and feature points in radar images Reliable feature point pairs are selected by a feature matching algorithm based on the bidirectional nearest neighbor distance ratio; S32. Use the graph cut algorithm to find the minimum value of the energy function, and obtain the optimal transformation parameters and energy function. From data items and smoothing terms The composition, formula is as follows: ； ； ； ； ； in, It is the set of all pixels in the image. It is the set of matched feature point pairs. It is a set of adjacent pixel pairs. It is a transformation matrix used to map points in a radar image to the optical image coordinate system. It is the balance coefficient. Adjacent pixel pairs and The distance between them It is a set of adjacent pixel pairs. It is a parameter that controls the smoothness. It is a similarity measure function based on conditional probability. It is a similarity measure. It represents a measure of smoothness between adjacent pixel pairs; S33. By finding the correspondence between the images through S31 and S32, precise registration is achieved, and the registered optical image is obtained. and radar images ; Step 4: For the registered optical image and radar images An adaptive fusion rule based on regional energy, gradient, and structure tensor information is used to obtain the fused image. The fused image is obtained by employing an adaptive fusion rule based on regional energy, gradient, and structure tensor information. The process is as follows: S41, Calculate the image by local window centered Internal regional energy Gradient magnitude and structure tensor eigenvalues , , The expression is as follows: ； ； ； By analyzing the structure tensor matrix Eigenvalue decomposition yields and ; S42. Calculate the weighting coefficients of the optical image. Weighting coefficients of radar images The expression is as follows: ； ； In the formula, Indicates for image , its in local window centered The energy within the region, and For local windows The horizontal and vertical coordinates of the pixels within. Representing an image In coordinates gradient magnitude at that point and It is an image exist and The partial derivative in the direction, with the subscripts opt and rad indicating optical and radar respectively; Represents the structure tensor eigenvalues; S43. Obtain the fused image. The expression is as follows: ； Step 5: Extract corresponding features from the fused image obtained in Step 4 using manual feature extraction and deep learning-based feature extraction. Then, fuse the extracted features to obtain the final target recognition result. Manual feature extraction uses an edge detection algorithm based on the Canny operator to extract the ship's outline. By performing morphological processing and contour tracking algorithms on the edge image, the boundary of the ship target is accurately determined. Then, multiple geometric features of the ship are calculated, including aspect ratio, perimeter, and area change rate.

2. The multi-source ship target identification method based on two-level collaborative fusion according to claim 1, characterized in that, The formula for Gaussian filtering is as follows: ； in It is a Gaussian kernel, defined as: ； In the formula, This represents the filtered image. Represents the original image. The standard deviation of the Gaussian kernel is used to control the width and height of the kernel. This represents the increment of the x-coordinate of the input data point. This represents the increment of the ordinate of the input data point.

3. The multi-source ship target identification method based on two-level collaborative fusion according to claim 2, characterized in that: Deep learning feature extraction employs an architecture that combines multi-branch convolutional neural networks with attention mechanisms. The multi-branch convolutional neural network processes different scales and feature levels of the fused image separately. Each branch contains several convolutional layers, pooling layers, and residual connection modules to extract rich image features. The attention mechanism learns the importance weights of different regions in the image, enabling the model to focus on the target area of ​​the ship and suppress the interference of background noise.

Images