A ship target detection and identification method combining SAR and optical images

Abstract

The application belongs to the technical field of target detection and identification based on multi-source information fusion, and provides a ship target detection and identification method combined with SAR and optical images to improve the target detection efficiency; first, the attention mechanism is introduced in the target detection model of the optical image, and better detection performance is obtained on the optical image; second, the mapping relationship between the image coordinates and the geographic coordinates is used to obtain the SAR image slice, and the detection result of the optical image is confirmed again, so that the interference of the false target on the optical image is reduced, and the robustness of the model is improved; finally, the decision-level fusion method of D-S evidence theory based on the evidence correlation coefficient is introduced, the target class probability distribution in the SAR image and the optical image is combined, the conflicting evidence is effectively alleviated, the accuracy and reliability of the target type identification are improved, and the interference ability of the model to the false target is also improved, and the accuracy of the false target detection is improved to 97.5%.

Description

Technical Field

[0001] This invention belongs to the technical field of target detection and recognition based on multi-source information fusion. More specifically, this invention provides a method for ship target detection and recognition based on SAR and optical images using improved DS evidence theory. Background Technology

[0002] Remote sensing image target interpretation is one of the important research directions in the fields of computer vision and image processing. It can be used to detect specific categories of object instances in digital images. Because remote sensing sensors acquire images over a wide range, target detection based on remote sensing images has a wide range of applications in military and civilian fields such as intelligence reconnaissance, target surveillance, military strikes, disaster relief, and industrial development.

[0003] Currently, mainstream target detection algorithms are based on single data sources such as optical remote sensing images or SAR remote sensing images. Optical images, due to their rich color information, high resolution, and strong semantic information, have become the most mainstream data source. With the rapid development of target detection networks such as YOLO, the performance of target detection methods based on optical remote sensing images has been greatly improved; however, some drawbacks and problems that are difficult to solve immediately still exist: First, the quality of the acquired optical remote sensing data is greatly affected by the external environment. In adverse weather conditions, insufficient lighting, or unsatisfactory shooting distances and angles, the acquired images may appear blurry, leading to the loss of semantic information of key object instances in the digital image; Second, because optical remote sensing images are a true reflection of ground colors, deception techniques targeting optical images, such as camouflage, concealment, and stealth, make target detection algorithms based solely on optical remote sensing images less robust and reliable.

[0004] Meanwhile, with the continuous improvement of SAR imaging technology, coupled with its advantages of all-weather, all-day operation, strong penetration, and rich texture information, research on target interpretation of SAR remote sensing images has been continuously deepened. SAR images are most commonly used for ship and building target detection, but compared to optical images, they have poorer readability, lower resolution, and are easily affected by speckle noise and geometric distortion, resulting in lower detection and recognition performance in remote sensing target interpretation. To better leverage the advantages of optical and SAR sensors in practical tasks, researchers have applied the concept of "multi-source information fusion" to the specific application of remote sensing target interpretation. With the rapid development and innovation of sensing technology, wireless communication technology, and aerospace technology in recent years, a large number of optical and SAR satellites have been successfully launched and are operating worldwide, providing researchers with a vast research space for increasingly abundant heterogeneous remote sensing data.

[0005] Currently, optical ship datasets such as HRSC2016, xView, FGSD, DIOR, ShipRSImageNet, and SAR datasets such as SSDD, FUSAR-Ship, AIR-SARShip, and OpenSARShip are all open-source, providing possibilities for ship target detection and recognition by combining SAR and optical images. However, the following problems still exist in practical engineering applications: First, due to the different imaging methods of SAR and optical images, it is difficult to achieve strict spatiotemporal synchronization, causing significant interference to the fusion of heterogeneous data. Second, due to the different heights and angles at which optical and SAR image sensors acquire data, as well as the diversity of polarization modes of SAR sensors, the same target appears in different forms in different sensors, increasing the difficulty of detection. Third, the SAR images in the current mainstream open-source datasets are usually acquired by satellites and lack detailed information, resulting in poor performance when used for classification and recognition tasks. Based on this, remote sensing target interpretation technology based on multi-source information fusion remains a research hotspot and focus. However, current research lacks a scheme to effectively combine SAR and optical sensors, resulting in low target detection efficiency. How to achieve efficient fusion of optical and SAR remote sensing data remains an unsolved problem. Summary of the Invention

[0006] The purpose of this invention is to provide a method for ship target detection and recognition that combines SAR and optical images, thereby achieving efficient fusion of optical and SAR remote sensing data and improving the target detection efficiency of ship target detection and recognition. First, a CBAM attention mechanism is introduced into the YOLO V7 target detection network model for optical images, achieving better detection performance on optical images. Second, SAR image slices are obtained using the mapping relationship between image coordinates and geographic coordinates, and the detection results of the optical images are reconfirmed based on the SAR image slices, reducing the interference of false targets on the optical images and improving the robustness of the model. Finally, a decision-level fusion method based on the DS evidence theory with evidence correlation coefficients is introduced, combining the probability distribution of target categories in SAR and optical images, effectively mitigating conflicting evidence, improving the accuracy and reliability of target type recognition, and enhancing the model's ability to resist interference from false targets, increasing the accuracy of false target detection to 97.5%.

[0007] The present invention adopts the following technical solution to achieve the above-mentioned technical objectives:

[0008] A method for ship target detection and recognition combining SAR and optical imagery, characterized by the following steps:

[0009] Acquire spatiotemporally synchronized optical and SAR images of the target object;

[0010] The optical image to be detected is input into a pre-trained target detection model to obtain the coordinate parameters of the target detection box and the probability distribution of the target category based on the optical image.

[0011] By mapping and inverse mapping the image coordinates and geographic coordinates, the coordinate parameters of the target detection box in the optical image are mapped to the SAR image to be detected, thus obtaining the target detection box of the SAR image, and SAR image slices are obtained based on the target detection box.

[0012] A pre-trained SAR ship slice detection model is used to classify and judge SAR image slices to obtain SAR target slices;

[0013] Inputting SAR target slices into a pre-trained SAR image ship classification model yields a target category probability distribution based on the SAR image.

[0014] The DS evidence fusion method based on evidence correlation coefficient is used to fuse the target category probability distribution based on optical image and the target category probability distribution based on SAR image to obtain the final target category probability distribution.

[0015] Furthermore, the target detection model adopts the YOLO V7 target detection model with an attention mechanism. The model input is an optical image, specifically including a backbone part, a neck part, and a head part. The original optical image is input to the backbone part. The backbone part outputs three feature maps (low-level, mid-level, and high-level) to the neck part. The neck part performs feature fusion and outputs the results to the head part. The head part outputs the coordinate parameters of the target detection box and the probability distribution of the target category. In particular, a CBAM attention module is added before each output branch of the backbone part, and a CBAM attention module is added between the RepVGG module and the convolution module of each scale branch of the head part.

[0016] Furthermore, the SAR ship slice detection model employs an SVM-based binary classifier to extract HOG and SIFT features from SAR image slices, and performs feature stitching, using the results as model input.

[0017] Furthermore, the SAR image ship classification model adopts the ViT model, and the model input is a SAR target slice.

[0018] Furthermore, the specific process of coordinate parameter mapping is as follows:

[0019] Obtain the six parameters trans_opt of the optical image to be detected, and convert the target detection box coordinates of the optical image from image coordinates to geographic projection coordinates:

[0020] px=trans_opt[0]+col_opt×trans_opt[1]+row_opt×trans_opt[2]

[0021] py=trans_opt[3]+col_opt×trans_opt[4]+row_opt×trans_opt[5]

[0022] Where px and py are the horizontal and vertical coordinates of the geographic projection coordinates, col_opt is the column coordinate of the pixel in the optical image, row_opt is the row coordinate of the pixel in the optical image, and trans_opt[0], trans_opt[1], trans_opt[2], trans_opt[3], trans_opt[4], and trans_opt[5] are the six parameters of the optical image raster data.

[0023] Obtain the six parameters trans_sar of the SAR image to be detected, and solve the equations to obtain row_sar and col_sar:

[0024] px=trans_sar[0]+col_sar×trans_sar[1]+row_sar×trans_sar[2]

[0025] py=trans_sar[3]+col_sar×trans_sar[4]+row_sar×trans_sar[5]

[0026] Wherein, col_sar is the column coordinate of a pixel in the SAR image, row_sar is the row coordinate of a pixel in the SAR image; trans_sar[0], trans_sar[1], trans_sar[2], trans_sar[3], trans_sar[4], and trans_sar[5] correspond to the six parameters of the SAR image raster data.

[0027] Furthermore, the DS evidence fusion method based on the evidence correlation coefficient is as follows:

[0028] Let the probability distributions of target categories based on optical images and SAR images be evidence m1 and m2, respectively. Calculate the evidence correlation coefficient r based on the probability distributions of m1 and m2. BPA :

[0029]

[0030] Where c(m1,m2) represents the degree of correlation between evidence m1 and m2:

[0031]

[0032] Where i,j=1,2,3,A i and B j The focal elements of evidence m1 and m2 are respectively, m1(A i () indicates that evidence m1 belongs to category A. i The probability, m2(B) j ) indicates that evidence m2 belongs to category B. j The probability of |·|; |·| represents the potential of the focal element (i.e., the number of elements in the focal element);

[0033] Using the evidence correlation coefficient as the support, we obtain the support matrix:

[0034]

[0035] Among them, S 22 =S 11 =1,S 21 =S 12 =r BPA (m1,m2);

[0036] Calculate other evidence for any piece of evidence m p Total support level:

[0037]

[0038] Where p,q=1,2,…n, n is the number of pieces of evidence;

[0039] Calculate evidence m p Credibility of Crd(m) p ):

[0040]

[0041] in,

[0042] The basic probability assignment of new evidence is obtained using a weighted average method:

[0043]

[0044] Where m represents new evidence, m(A) i ) indicates that new evidence m belongs to category A. i Basic probability assignment;

[0045] By using the Dempster combination rule to perform two fusions on the new evidence m, the probability distribution of the fused target category is obtained, thus completing the ship target detection and identification.

[0046] Based on the above technical solution, the beneficial effects of the present invention are as follows:

[0047] This invention provides a method for ship target detection and recognition based on SAR and optical images using improved DS evidence theory, which has the following advantages:

[0048] 1) The method of the present invention introduces the CBAM attention mechanism into the YOLO V7 target detection network model for optical images, which achieves better detection performance on optical images compared with the baseline model YOLO V7;

[0049] 2) The method of the present invention uses the mapping relationship between image coordinates and geographic coordinates to obtain a slice of the target area in the SAR image. By integrating the coordinate information of the optical image and the target category information in the SAR image, the detection results of the optical image are confirmed for the second time, which reduces the interference of false targets on the optical image and improves the robustness of the model.

[0050] 3) The method of this invention introduces the DS evidence theory based on the evidence correlation coefficient. By combining the target category probability distribution in SAR images and the target category probability distribution in optical images, it effectively alleviates conflicting evidence, improves the accuracy and reliability of target type identification, and also enhances the model's ability to resist interference from false targets.

[0051] 4) The method of the present invention applies a decision-level fusion method, which solves the problem of difficulty in obtaining strictly spatiotemporally synchronized optical images and SAR images, and reduces the model's dependence on the strict spatiotemporal synchronization requirement of training data. Attached Figure Description

[0052] Figure 1 This is a flowchart illustrating the ship target detection and recognition method combining SAR and optical images according to the present invention.

[0053] Figure 2 These are the original optical images and SAR images for spatiotemporal synchronization in this embodiment of the invention.

[0054] Figure 3 This is a schematic diagram of the structure of the YOLO V7 object detection model with an attention mechanism introduced in an embodiment of the present invention, wherein (a) is a schematic diagram of the backbone part and (b) is a schematic diagram of the head part.

[0055] Figure 4 This is a visualization of the optical image ship target detection and recognition results of the YOLO V7 target detection model with an attention mechanism introduced in this embodiment of the invention.

[0056] Figure 5 This is a visualization result of ship target detection and recognition using combined SAR and optical images in an embodiment of the present invention. Detailed Implementation

[0057] To make the objectives, technical solutions, and effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. The following embodiments are used to explain the invention and are not intended to limit the scope of the invention.

[0058] This embodiment provides a method for ship target detection that combines SAR and optical imagery, the process of which is as follows: Figure 1 As shown, the module involves an optical target detection module, a secondary verification module, and a DS decision module, and specifically includes the following steps:

[0059] Step 1: Obtain optical and SAR datasets for three types of civilian vessels—cargo ships, tankers, and container ships—based on open-source datasets;

[0060] Step 1-1: Optical Image Data Acquisition

[0061] This embodiment uses four datasets: FGSCR-42, ShipRSImageNet, Ships in Satellite Imagery, and SpaceNet-6. It acquires 2978 optical images for three types of civilian ships: cargo ships, tankers, and container ships, as an optical ship target detection dataset. Among them, there are 1503 images of cargo ships, 595 images of tankers, and 878 images of container ships.

[0062] Steps 1-2: SAR Image Data Acquisition

[0063] This embodiment obtained 190 cargo ship slices, 467 oil tanker slices, and 247 container ship slices, totaling 904 ship slices, based on the ICEYE and SpaceNet-6 datasets. Based on slice size clustering, and with a ship slice to background slice ratio of 1:5, background slices with aspect ratios of 1:1, 2:1, and 1:2 were obtained, totaling 4610 background slices. The ship slices (positive samples) and background slices (negative samples) together constituted the SAR slice detection dataset; and the ship slices were used as the SAR ship classification dataset.

[0064] like Figure 2 The images shown are examples of spatiotemporally synchronized original optical and SAR images, where the left image is the optical image and the right image is the SAR image.

[0065] Step 2: Train the YOLO V7 target detection model with attention mechanism based on the optical ship target detection dataset. Input the optical image to be detected into the trained YOLO V7 target detection model with attention mechanism. Input the coordinate parameters of the target detection box in the detection result into Step 3 for coordinate mapping. Then input it into Step 4 for secondary confirmation to determine whether it is a real target. Input the probability distribution of ship category in the detection result into Step 6 for DS fusion ship type recognition.

[0066] The YOLO V7 object detection model that incorporates an attention mechanism is as follows: Figure 3 As shown, the network consists of a backbone, a neck, and a head. The raw optical image is input to the backbone. The backbone outputs three feature maps (low-level, mid-level, and high-level) to the neck. The neck performs feature fusion and outputs the results to the head. The head outputs the coordinate parameters and class probability distribution of the target detection box. The network structure diagram of the backbone is shown below. Figure 3 As shown in (a), the network includes an ELAN module, an MP module, and a CBAM attention module. The ELAN module is responsible for feature extraction, the MP module is responsible for downsampling, and the CBAM module is responsible for allocating attention in the spatial and channel directions of the feature map to improve feature extraction in subsequent networks. This invention adds a CBAM attention module before each output branch of the YOLO V7 backbone, enabling the network to allocate different weights to the feature map space and channels during feature fusion in the neck section. The network structure diagram of the head section is shown below. Figure 3 As shown in (b), it includes: RepVGG module, CBAM module and convolution module. The RepVGG module is responsible for improving network performance and inference speed, CBAM is responsible for allocating attention to feature maps, and Conv module is responsible for unifying the number of channels in the output feature map. This invention introduces CBAM module between RepVGG module and convolution module in each scale branch.

[0067] In this embodiment, the training parameters are set as follows: batch size is set to 32, image size is 640×640, and a total of 300 epochs are trained to complete the offline training of the YOLO V7 ship target detection model with attention mechanism.

[0068] Step 3: Extract the coordinate parameters (image coordinates) of the target detection box of the optical image to be detected, and map the coordinates of the target detection box of the optical image to the SAR image to be detected through the mapping and inverse mapping relationship between image coordinates and geographic coordinates to obtain the target detection box of the SAR image, and obtain SAR image slices based on the target detection box.

[0069] Step 3-1: Use Python's GDAL remote sensing data processing library to obtain the six parameters trans_opt of the optical image to be detected, and convert the target detection box coordinates in the optical image from image coordinates to geographic projection coordinates:

[0070] px=trans_opt[0]+col_opt×trans_opt[1]+row_opt×trans_opt[2]

[0071] py=trans_opt[3]+col_opt×trans_opt[4]+row_opt×trans_opt[5]

[0072] Wherein, px and py are the horizontal and vertical coordinates of the geographic projection coordinates, col_opt is the column coordinate of the pixel in the optical image, row_opt is the row coordinate of the pixel in the optical image, trans_opt is the six parameters of the optical image raster data, trans_opt[0] is the projection coordinate of the horizontal coordinate of the upper left corner of the optical image, trans_opt[1] is the row rotation parameter of the optical image, trans_opt[2] is the pixel width of the optical image, that is, the resolution of the image in the horizontal space, trans_opt[3] is the projection coordinate of the vertical coordinate of the upper left corner of the optical image, trans_opt[4] is the column rotation parameter of the optical image, and trans_opt[5] is the pixel width of the optical image, that is, the resolution of the image in the vertical space;

[0073] Step 3-2: Use the gdal library to obtain the six parameters trans_sar of the SAR image. The projected coordinates calculated from the optical image are used to solve the following linear equation in two variables to obtain row_sar and col_sar:

[0074] px=trans_sar[0]+col_sar×trans_sar[1]+row_sar×trans_sar[2]

[0075] py=trans_sar[3]+col_sar×trans_sar[4]+row_sar×trans_sar[5]

[0076] Wherein, col_sar is the column coordinate of a pixel in the SAR image, row_sar is the row coordinate of a pixel in the SAR image; trans_sar is the six parameters of the SAR image raster data, trans_sar[0] is the projection coordinate of the horizontal coordinate of the upper left corner of the SAR image, trans_sar[1] is the row rotation parameter of the SAR image, trans_sar[2] is the pixel width of the SAR image, that is, the resolution of the image in the horizontal space, trans_sar[3] is the projection coordinate of the vertical coordinate of the upper left corner of the SAR image, trans_sar[4] is the column rotation parameter of the SAR image, and trans_sar[5] is the pixel width of the SAR image, that is, the resolution of the image in the vertical space.

[0077] Step 4: Based on the SAR slice detection dataset, extract HOG features and SIFT features to train an SVM-based binary classifier as the SAR ship slice detection model; use the trained detection model to classify the SAR image slices obtained in Step 3. If it is a ship target, input it into Step 5 for reclassification; otherwise, discard the false target.

[0078] The specific process in this embodiment is as follows:

[0079] Step 4-1: First, adjust the SAR image slices to a size of 224×224;

[0080] Step 4-2: Use the HOGDescriptor descriptor in Python's OpenCV library, set winSize to 64×128, blockSize to 16×16, and nbins to 9, to extract HOG features from SAR image slices.

[0081] Step 4-3: Use the cv2.xfeatures2d.SIFT_create method in Python's OpenCV library to extract SIFT features from SAR image slices;

[0082] Step 4-4: Use the vstack method in Python's NumPy library to concatenate two features, and use the sklearn library to generate an SVM classifier, setting its kernel to the RBF function;

[0083] Steps 4-5: Train the SVM classifier based on the SAR slice detection dataset, and use the trained SVM classifier to classify the SAR image slices obtained in step 3. If it is a ship target, input it into step 5 for reclassification; otherwise, discard the false target.

[0084] Step 5: Train the ViT model based on the SAR ship classification dataset as the SAR image ship classification model; input the ship target slices obtained in Step 4 into the trained ViT model to obtain its target category probability distribution, which is then used as the input for Step 6;

[0085] The specific process in this embodiment is as follows:

[0086] Step 5-1: Since SAR images are all single-channel images, the PIL library in Python is used to read the images, and the convert method is used to convert them into RGB JPG format;

[0087] Step 5-2: Adjust the image size to 224×224 and normalize the image.

[0088] Step 5-3: Train the ViT model based on the SAR ship classification dataset, and use the trained ViT model to classify the ship target slices obtained in Step 4 to obtain the target category probability distribution, which is then input into the DS decision framework in Step 6.

[0089] Step 6: Use the DS evidence fusion framework based on evidence correlation coefficient to fuse and identify the optical target category probability distribution obtained in Step 2 and the SAR target category probability distribution obtained in Step 5;

[0090] Step 6-1: Let the probability distribution of optical target categories obtained in Step 2 and the probability distribution of SAR target categories obtained in Step 5 be evidence m1 and m2, respectively. Calculate the evidence correlation coefficient r based on the probability distributions of m1 and m2. BPA The calculation formula is as follows:

[0091]

[0092] Where c(m1,m2) represents the degree of correlation between evidence m1 and m2, and the calculation formula is as follows:

[0093]

[0094] Where i,j=1,2,3,A i and B j The focal elements of evidence m1 and m2 are respectively, m1(A i () indicates that evidence m1 belongs to category A. i The probability, m2(B) j ) indicates that evidence m2 belongs to category B. j The probability of |·|; |·| represents the potential of the focal element (i.e., the number of elements in the focal element);

[0095] Step 6-2: Calculate the evidence correlation coefficients between each pair of pieces of evidence in 6-1 as the support, and obtain the support matrix:

[0096]

[0097] Among them, S 22 =S 11 =1,S 21 =S 12 =r BPA (m1,m2);

[0098] Step 6-3: Define other evidence in relation to a given piece of evidence m p The total level of support is given by the following formula:

[0099]

[0100] Where p,q=1,2,…n, n is the number of pieces of evidence;

[0101] Step 6-4: Calculate the evidence m using the support matrix. p Credibility of Crd(m) p The calculation formula is as follows:

[0102]

[0103] in,

[0104] Step 6-5: Use a weighted average method to obtain the basic probability assignment of new evidence:

[0105]

[0106] Where m represents new evidence, m(A) i ) indicates that new evidence m belongs to category A. i Basic probability assignment;

[0107] Step 6-6: Use Dempster's combination rule to perform two fusions on the new evidence m to obtain the fused probability distribution.

[0108] Step 7: Visualize the detection and recognition results in the optical image, label true targets with their true category and corresponding probability, and label false targets as "False target".

[0109] The benefits of this invention will be further illustrated below with simulation tests:

[0110] In this embodiment, the hardware platform is Intel Xeon Silver 4214R + NVIDIA RTX3090 + 16GDDR4 RAM, and the software environment is CentOS 7.4 + CUDA 11.1 + PyTorch 1.10.0 + Python 3.7. PyCharm is used for algorithm development. The datasets used are self-made datasets for optical ship target detection, SAR slice detection, and SAR ship classification. The test set is the spatiotemporally synchronized valid data in the SpaceNet-6 dataset. The evaluation metrics are three commonly used metrics in the field of target detection: mAP, mAR, and number of floating-point calculations.

[0111] To demonstrate the effectiveness of this invention, a comparison was made with the YOLO V7 baseline model, which is based solely on optical images. The results are shown in Table 1.

[0112] Table 1

[0113] Model mAP (IOU = 0.50) mAR Number of floating point calculations YOLO V7 0.95 0.925 16.0 GFlops The present invention 0.975 0.948 15.2 GFlops

[0114] As shown in Table 1, compared with the baseline YOLO V7 network model, the present invention has a 2.5% higher mAP and a 2.3% higher mAR when IOU=0.50. At the same time, the number of floating-point calculations is reduced by 0.8 GFlops. While improving the reliability and accuracy of detection, it also takes into account time and reduces the complexity of the model.

[0115] Furthermore, this invention fully combines the advantages of optical images and SAR images, introducing a secondary verification mechanism and a DS decision framework, improving the detection accuracy of false targets to 97.5%, effectively solving the problem of false target deception in optical images, and improving the robustness of the model. Specifically, as... Figure 4 The results shown are those relying solely on optical image detection and recognition. Figure 5 The figure shows the ship target detection and identification results after introducing the secondary confirmation mechanism and DS evidence theory. As can be seen from the figure, the latter reduces the interference of false targets and improves the robustness of the model.

[0116] In summary, this invention addresses the problem of optical images being easily deceived by false targets. Starting from improving the robustness of existing remote sensing target detection algorithms, it proposes a ship target detection method that combines SAR and optical images. By utilizing the complementary characteristics between sensors and introducing the CBAM attention mechanism, the detection efficiency of the original detection model is improved, the detection error is reduced, and the robustness of the model is enhanced. This is an effective and practical remote sensing ship target detection method.

[0117] The above description is merely a specific embodiment of the present invention. Any feature disclosed in this specification may be replaced by other equivalent or similar features unless otherwise specified. All disclosed features, or steps in all methods or processes, may be combined in any way except for mutually exclusive features and / or steps.

Claims

1. A method for ship target detection and recognition combining SAR and optical images, characterized in that, Includes the following steps: Acquire spatiotemporally synchronized optical and SAR images of the target object; The optical image to be detected is input into a pre-trained target detection model to obtain the coordinate parameters of the target detection box and the probability distribution of the target category based on the optical image. By mapping and inverse mapping the image coordinates and geographic coordinates, the coordinate parameters of the target detection box in the optical image are mapped to the SAR image to be detected, thus obtaining the target detection box of the SAR image, and SAR image slices are obtained based on the target detection box. A pre-trained SAR ship slice detection model is used to classify and judge SAR image slices to obtain SAR target slices; Inputting SAR target slices into a pre-trained SAR image ship classification model yields a target category probability distribution based on the SAR image. The DS evidence fusion method based on evidence correlation coefficient is used to fuse the target category probability distribution based on optical image and the target category probability distribution based on SAR image to obtain the final target category probability distribution. The target detection model employs the YOLO V7 target detection model with an attention mechanism. The model input is an optical image, specifically comprising a backbone, a neck, and a head. The original optical image is input to the backbone, which outputs three feature maps (low-level, mid-level, and high-level) to the neck. The neck performs feature fusion and outputs the results to the head, which outputs the coordinate parameters of the target detection bounding box and the probability distribution of the target category. A CBAM attention module is added before each output branch of the backbone, and a CBAM attention module is added between the RepVGG module and the convolutional module in each scale branch of the head. The SAR ship slice detection model uses an SVM-based binary classifier to extract HOG and SIFT features from SAR image slices, and then performs feature stitching, using the result as the model input. The SAR image ship classification model adopts the ViT model, and the model input is SAR target slices.

2. The ship target detection and recognition method based on combined SAR and optical images as described in claim 1, characterized in that, The specific process of coordinate parameter mapping is as follows: Obtain the six parameters trans_opt of the optical image to be detected, and convert the target detection box coordinates of the optical image from image coordinates to geographic projection coordinates: px=trans_opt[0]+col_opt×trans_opt[1]+row_opt×trans_opt[2] py=trans_opt[3]+col_opt×trans_opt[4]+row_opt×trans_opt[5] Where px and py are the horizontal and vertical coordinates of the geographic projection coordinates, col_opt is the column coordinate of the pixel in the optical image, row_opt is the row coordinate of the pixel in the optical image, and trans_opt[0], trans_opt[1], trans_opt[2], ... trans_opt[3], trans_opt[4], and trans_opt[5] correspond to the six parameters of optical image raster data; Obtain the six parameters trans_sar of the SAR image to be detected, and solve the equations to obtain row_sar and col_sar: px=trans_sar[0]+col_sar×trans_sar[1]+row_sar×trans_sar[2] py=trans_sar[3]+col_sar×trans_sar[4]+row_sar×trans_sar[5] Wherein, col_sar is the column coordinate of a pixel in the SAR image, row_sar is the row coordinate of a pixel in the SAR image; trans_sar[0], trans_sar[1], trans_sar[2], trans_sar[3], trans_sar[4], and trans_sar[5] correspond to the six parameters of the SAR image raster data.

3. The ship target detection and recognition method based on combined SAR and optical images as described in claim 1, characterized in that, The DS evidence fusion method based on the evidence correlation coefficient is as follows: Let the probability distributions of target categories based on optical images and SAR images be evidence m1 and m2, respectively. Calculate the evidence correlation coefficient r based on the probability distributions of m1 and m2. BPA : Where c(m1,m2) represents the degree of correlation between evidence m1 and m2: A i ,B j ∈[cargo,tanker,container] Where i,j=1,2,3,A i and B j The focal elements of evidence m1 and m2 are respectively, m1(A i () indicates that evidence m1 belongs to category A. i The probability, m2(B) j ) indicates that evidence m2 belongs to category B. j The probability of; Using the evidence correlation coefficient as the support, we obtain the support matrix: Among them, S 22 =S 11 =1,S 21 =S 12 =r BPA (m1,m2); Calculate other evidence for any piece of evidence m p Total support level: Where p,q=1,2,…n, n is the number of pieces of evidence; Calculate evidence m p Credibility of Crd(m) p ): in, The basic probability assignment of new evidence is obtained using a weighted average method: Where m represents new evidence, m(A) i ) indicates that new evidence m belongs to category A. i Basic probability assignment; By using the Dempster combination rule to perform two fusions on the new evidence m, the probability distribution of the fused target category is obtained, thus completing the ship target detection and identification.

Images