Ship target detection method and system in SAR images based on feature balance network

Abstract

The present invention discloses a method and system for detecting ship targets in SAR images based on a feature balance network, comprising the following steps: extracting features from preprocessed SAR images; inputting multi-level features into an FPN network to obtain a multi-scale feature map; inputting the multi-scale feature map into an RPN network to obtain a horizontal candidate box for the ship target; mapping the coordinate information of the horizontal candidate box into the multi-scale feature map, calculating the candidate box region features through an ROI Align algorithm, and generating a target box feature vector by pooling the calculation results; predicting the target box feature vector through an RCNN detection head to obtain a ship target detection result. The present invention integrates the local features of deformable convolution and the global features based on attention, which can better provide image features for ship detection, and adopts a multi-stage target rotation box representation prediction, which can more smoothly predict the information of the target rotation box, better extract ship features, and improve detection accuracy.

Description

Technical Field

[0001] The present invention relates to the field of ship detection, and in particular to a SAR image ship target detection method and system based on a feature balance network. Background Art

[0002] Synthetic Aperture Radar (SAR) has the advantage of constructing radar images under all-weather and all-day conditions. Currently, SAR image ship detection technology can be widely used in the fields of maritime vessel management, geological detection, target detection and early warning, and has very important application value.

[0003] Traditional SAR ship detection technology usually uses sliding window and threshold screening methods for detection. The selection of thresholds is very dependent on the data set distribution modeling and has poor generalization. Detection technology based on deep learning models can autonomously learn the distribution of data and the characteristics of ship targets, which can effectively improve detection accuracy and generalization. However, the main difficulties of SAR ship detection are as follows:

[0004] 1. In near-shore scenarios, ships are docked densely and port shore buildings will interfere with ship detection;

[0005] 2. Maritime ships include a variety of large, medium and small civilian ships, and the size of the ships varies greatly;

[0006] 3. SAR images are usually bird's-eye view images, and the angles and postures of ships are varied. Using horizontal box representation learning will bring more feature interference. Summary of the invention

[0007] In view of the above-mentioned deficiencies in the prior art, the method and system for SAR image ship target detection based on feature balance network provided by the present invention solve the problems of poor accuracy and generalization of the existing SAR ship target detection algorithms.

[0008] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is:

[0009] A method for detecting ship targets in SAR images based on a feature balance network is provided, which comprises the following steps:

[0010] S1, obtaining a SAR image to be detected and preprocessing it to obtain a preprocessed SAR image;

[0011] S2, extracting features from the preprocessed SAR image through a feature balancing network to obtain multi-level features;

[0012] S3, input the multi-level features into the FPN network to obtain a multi-scale feature map;

[0013] S4, input the multi-scale feature map into the RPN network to obtain the coordinate information of the horizontal candidate box of the ship target;

[0014] S5, mapping the coordinate information of the horizontal candidate box of the ship target obtained in step S4 to the multi-scale feature map, calculating the regional features of the candidate box by using the ROI Align algorithm, and pooling the calculation results to generate a first target box feature vector;

[0015] S6. Use the first RCNN detection head to predict the first target frame feature vector to obtain first target category information, first target rotation frame coordinate information, and first target rotation frame angle information;

[0016] S7, mapping the first target rotation frame obtained in step S6 to the multi-scale feature map, calculating the candidate frame region features through the ROI Align algorithm, and pooling the calculation results to generate a second target frame feature vector;

[0017] S8, using the second RCNN detection head to predict the second target frame feature vector, to obtain second target category information, second target rotation frame coordinate information, and second target rotation frame angle information;

[0018] S9. Output the second target category information and the second target rotation frame as the current ship target detection result.

[0019] Furthermore, the specific method of step S1 includes the following sub-steps:

[0020] S1-1, standardizing the SAR image to be detected by using the pre-set mean and variance of the image data;

[0021] S1-2, normalizing the standardized SAR image to be detected;

[0022] S1-3, performing zero-filling on the normalized SAR image to be detected so that the width and height of the filled image are both integer multiples of 32, thereby obtaining a preprocessed SAR image.

[0023] Furthermore, in step S2, the feature balancing network includes a first feature balancing module, a second feature balancing module, a third feature balancing module and a fourth feature balancing module connected in sequence; each feature balancing module includes a first convolutional layer, a second convolutional layer, a deformable convolutional layer, a spatial attention layer, a feature splicing layer and a channel attention layer connected in sequence; the specific method for the feature balancing module to extract features from the input image includes the following sub-steps:

[0024] S2-1, performing different convolution operations on the input image through the first convolution layer and the second convolution layer respectively, and obtaining a first convolution result and a second convolution result whose channel dimension is half of the input image respectively;

[0025] S2-2, performing a deformable convolution operation on the second convolution result through a deformable convolution layer;

[0026] S2-3, perform spatial attention operation on the deformable convolution operation result through the spatial attention layer;

[0027] S2-4, concatenate the spatial attention operation result and the first convolution result through the feature concatenation layer, and use the concatenation result as the input of the channel attention layer;

[0028] S2-5, adjust the weight of the concatenation result along the channel dimension through the channel attention layer to obtain a single-level feature;

[0029] The single-level features corresponding to the four feature balancing modules constitute the multi-level features corresponding to the feature balancing network.

[0030] Furthermore, in step S3, the FPN network includes four 1×1 convolutional layers, namely, a first 1×1 convolutional layer, a second 1×1 convolutional layer, a third 1×1 convolutional layer, and a fourth 1×1 convolutional layer; the input of the first 1×1 convolutional layer is the single-level feature output by the first feature balancing module; the input of the second 1×1 convolutional layer is the single-level feature output by the second feature balancing module; the input of the third 1×1 convolutional layer is the single-level feature output by the third feature balancing module; the input of the fourth 1×1 convolutional layer is the single-level feature output by the fourth feature balancing module; wherein:

[0031] The output of the fourth 1×1 convolutional layer is the first scale feature map;

[0032] The first scale feature map is upsampled by 2 times and concatenated with the output of the third 1×1 convolutional layer to form the second scale feature map.

[0033] The concatenation of the second-scale feature map after 2-fold upsampling and the output of the second 1×1 convolutional layer is the third-scale feature map;

[0034] The third-scale feature map is upsampled by 2 times and concatenated with the output of the first 1×1 convolutional layer to obtain the fourth-scale feature map.

[0035] The first scale feature map, the second scale feature map, the third scale feature map and the fourth scale feature map together constitute a multi-scale feature map.

[0036] Furthermore, in step S4, if the confidence of the horizontal candidate box of the ship target is lower than 0.05, the horizontal candidate box of the ship target is deleted.

[0037] Furthermore, the sizes of the first target frame feature vector and the second target frame feature vector are both 7×7.

[0038] Furthermore, in step S6 and step S8, if the confidence of the target rotation box is lower than 0.05, the target rotation box is deleted.

[0039] A SAR image ship target detection system based on a feature balance network is provided, which includes an image preprocessing module, a feature balance network, an FPN network, an RPN network, a first target frame feature vector calculation module, a first RCNN detection head, a second target frame feature vector calculation module and a second RCNN detection head; wherein:

[0040] An image preprocessing module is used to preprocess the SAR image to be detected to obtain a preprocessed SAR image;

[0041] Feature balancing network, used to extract features from preprocessed SAR images to obtain multi-level features;

[0042] FPN network, used to obtain multi-scale feature maps corresponding to multi-level features;

[0043] RPN network, used to process the multi-scale feature map and obtain the coordinate information of the horizontal candidate box of the ship target;

[0044] The first target frame feature vector calculation module is used to map the coordinate information of the horizontal candidate frame of the ship target obtained by the RPN network to the multi-scale feature map, calculate the candidate frame region features through the ROI Align algorithm, and pool the calculation results to generate the first target frame feature vector;

[0045] The first RCNN detection head is used to predict the first target frame feature vector to obtain first target category information, first target rotation frame coordinate information, and first target rotation frame angle information;

[0046] A second target frame feature vector calculation module is used to map the first target rotation frame obtained by the first RCNN detection head into a multi-scale feature map, calculate the candidate frame region features through the ROI Align algorithm, and pool the calculation results to generate a second target frame feature vector;

[0047] The second RCNN detection head is used to predict the second target frame feature vector to obtain the second target category information, the second target rotation frame coordinate information and the second target rotation frame angle information;

[0048] The detection result output module is used to output the second target category information and the second target rotation frame as the current ship target detection result.

[0049] The beneficial effects of the present invention are as follows: the present invention integrates the local features of deformable convolution and the global features based on attention, which can better provide more robust and balanced image features for ship detection. The present invention adopts a multi-stage target rotation box representation prediction, which can more smoothly predict the coordinate information and angle information of the target rotation box, and better extract the ship features in the top view, thereby improving the detection accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 is a schematic diagram of the process of this method;

[0051] Figure 2 It is a structural diagram of the feature balance network and the FPN network;

[0052] Figure 3 It is a structural diagram of the feature balance module;

[0053] Figure 4 This is the data processing flow diagram of this system;

[0054] Figure 5 This is a schematic diagram comparing the test results of the present application with other structural solutions after the characteristic balancing network of the present application is replaced with other structures in the embodiments. DETAILED DESCRIPTION

[0055] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0056] like Figure 1 As shown, the SAR image ship target detection method based on feature balance network includes the following steps:

[0057] S1, obtaining a SAR image to be detected and preprocessing it to obtain a preprocessed SAR image;

[0058] S2, extracting features from the preprocessed SAR image through a feature balancing network to obtain multi-level features;

[0059] S3, input the multi-level features into the FPN network to obtain a multi-scale feature map;

[0060] S4, input the multi-scale feature map into the RPN network to obtain the coordinate information of the horizontal candidate box of the ship target;

[0061] S5, mapping the coordinate information of the horizontal candidate box of the ship target obtained in step S4 to the multi-scale feature map, calculating the regional features of the candidate box by using the ROI Align algorithm, and pooling the calculation results to generate a first target box feature vector;

[0062] S6. Use the first RCNN detection head to predict the first target frame feature vector to obtain first target category information, first target rotation frame coordinate information, and first target rotation frame angle information;

[0063] S7, mapping the first target rotation frame obtained in step S6 to the multi-scale feature map, calculating the candidate frame region features through the ROI Align algorithm, and pooling the calculation results to generate a second target frame feature vector;

[0064] S8, using the second RCNN detection head to predict the second target frame feature vector, to obtain second target category information, second target rotation frame coordinate information, and second target rotation frame angle information;

[0065] S9. Output the second target category information and the second target rotation frame as the current ship target detection result.

[0066] The specific method of step S1 includes the following sub-steps:

[0067] S1-1, standardizing the SAR image to be detected by using the pre-set mean and variance of the image data;

[0068] S1-2, normalizing the standardized SAR image to be detected;

[0069] S1-3, performing zero-filling on the normalized SAR image to be detected so that the width and height of the filled image are both integer multiples of 32, thereby obtaining a preprocessed SAR image.

[0070] like Figure 2 and Figure 3 As shown, the feature balancing network in step S2 includes a first feature balancing module, a second feature balancing module, a third feature balancing module and a fourth feature balancing module connected in sequence; each feature balancing module includes a first convolutional layer, a second convolutional layer, a deformable convolutional layer, a spatial attention layer, a feature splicing layer and a channel attention layer connected in sequence; the specific method for the feature balancing module to extract features from the input image includes the following sub-steps:

[0071] S2-1, performing different convolution operations on the input image through the first convolution layer and the second convolution layer respectively, and obtaining a first convolution result and a second convolution result whose channel dimension is half of the input image respectively;

[0072] S2-2, performing a deformable convolution operation on the second convolution result through a deformable convolution layer;

[0073] S2-3, perform spatial attention operation on the deformable convolution operation result through the spatial attention layer;

[0074] S2-4, concatenate the spatial attention operation result and the first convolution result through the feature concatenation layer, and use the concatenation result as the input of the channel attention layer;

[0075] S2-5, adjust the weight of the concatenation result along the channel dimension through the channel attention layer to obtain a single-level feature;

[0076] The single-level features corresponding to the four feature balancing modules constitute the multi-level features corresponding to the feature balancing network.

[0077] In step S3, the FPN network includes four 1×1 convolutional layers, namely the first 1×1 convolutional layer, the second 1×1 convolutional layer, the third 1×1 convolutional layer and the fourth 1×1 convolutional layer; the input of the first 1×1 convolutional layer is the single-level feature output by the first feature balancing module; the input of the second 1×1 convolutional layer is the single-level feature output by the second feature balancing module; the input of the third 1×1 convolutional layer is the single-level feature output by the third feature balancing module; the input of the fourth 1×1 convolutional layer is the single-level feature output by the fourth feature balancing module; wherein:

[0078] The output of the fourth 1×1 convolutional layer is the first scale feature map;

[0079] The first scale feature map is upsampled by 2 times and concatenated with the output of the third 1×1 convolutional layer to form the second scale feature map.

[0080] The concatenation of the second-scale feature map after 2-fold upsampling and the output of the second 1×1 convolutional layer is the third-scale feature map;

[0081] The third-scale feature map is upsampled by 2 times and concatenated with the output of the first 1×1 convolutional layer to obtain the fourth-scale feature map.

[0082] The first scale feature map, the second scale feature map, the third scale feature map and the fourth scale feature map together constitute a multi-scale feature map.

[0083] In step S4, if the confidence of the horizontal candidate box of the ship target is lower than 0.05, the horizontal candidate box of the ship target is deleted.

[0084] The sizes of the first target frame feature vector and the second target frame feature vector are both 7×7.

[0085] In step S6 and step S8, if the confidence of the target rotation box is lower than 0.05, the target rotation box is deleted.

[0086] like Figure 4 As shown, the SAR image ship target detection system based on feature balance network includes an image preprocessing module, a feature balance network, an FPN network, an RPN network, a first target frame feature vector calculation module, a first RCNN detection head, a second target frame feature vector calculation module and a second RCNN detection head; wherein:

[0087] An image preprocessing module is used to preprocess the SAR image to be detected to obtain a preprocessed SAR image;

[0088] Feature balancing network, used to extract features from preprocessed SAR images to obtain multi-level features;

[0089] FPN network, used to obtain multi-scale feature maps corresponding to multi-level features;

[0090] RPN network, used to process the multi-scale feature map and obtain the coordinate information of the horizontal candidate box of the ship target;

[0091] The first target frame feature vector calculation module is used to map the coordinate information of the horizontal candidate frame of the ship target obtained by the RPN network to the multi-scale feature map, calculate the candidate frame region features through the ROI Align algorithm, and pool the calculation results to generate the first target frame feature vector;

[0092] The first RCNN detection head is used to predict the first target frame feature vector to obtain first target category information, first target rotation frame coordinate information, and first target rotation frame angle information;

[0093] A second target frame feature vector calculation module is used to map the first target rotation frame obtained by the first RCNN detection head into a multi-scale feature map, calculate the candidate frame region features through the ROI Align algorithm, and pool the calculation results to generate a second target frame feature vector;

[0094] The second RCNN detection head is used to predict the second target frame feature vector to obtain the second target category information, the second target rotation frame coordinate information and the second target rotation frame angle information;

[0095] The detection result output module is used to output the second target category information and the second target rotation frame as the current ship target detection result.

[0096] In the specific implementation process, the traditional feature extraction network generally directly calculates the next layer of features by multi-layer convolution on the feature map. The feature balancing module of the present invention adopts a separated structure to calculate the features of multiple parts, and uses spatial attention and channel attention to autonomously weight and adjust the features of multiple parts, so as to better fuse the features and obtain the output features. The spatial attention operation formula is: Ms(F)=Sigmoid(Conv([AvgPool(F);MaxPool(F)])), where AvgPool is an average pooling operation, MaxPool is a spatial pooling operation, and Conv is a convolution operation. Spatial attention can learn to autonomously adjust the weight of each position in the feature. The channel attention operation formula is: Mc(F)=Sigmoid(MLP(AvgPool(F))+MLP(MaxPool(F))), where MLP is a multi-layer perceptron operation. The weight of each channel feature layer in the feature can be autonomously adjusted through the channel attention operation.

[0097] When different feature balancing modules extract features, they will calculate and extract feature layers with decreasing scales. Ordinary convolution operations are performed on any point on the feature map. It can be expressed as: ,in The features of ordinary convolution operation are mainly based on the features of 8 points around each point. The deformable convolution operation adopted by the present invention introduces a learnable offset for each point, which can be expressed as: , the offset is generated by another convolution operation. Because the offset calculation will automatically learn and adjust during the training process, compared with traditional convolution, deformable convolution can learn an offset that is more suitable for the ship contour in SAR image feature extraction, thereby extracting more robust ship features.

[0098] The feature sizes of the fourth-scale feature map to the first-scale feature map calculated by the FPN network decrease successively, and the number of channels increases successively. Usually, the low-level feature size is larger, the step length of adjacent pixels in the original image is smaller, and the generated target box is smaller, which is suitable for detecting small-scale targets; similarly, high-level features mean longer step lengths, which are suitable for large-scale targets; therefore, detection on multiple scale features can adapt to the multi-scale changes of ships. Usually, low-level features contain more positional information, and high-level features contain more semantic information. Therefore, the present invention uses the FPN network to fuse high-level ship information to low-level ship features, which is conducive to the detection of small-scale ship targets. Each layer of features in the FPN network is adjusted to the same number of channels using a 1×1 convolution operation, and the high-level features are expanded by 2 times upsampling, and then added to the low-level features to obtain the low-level features. From the highest layer (first) to the lower layer (fourth) features are fused in sequence to obtain multi-scale features.

[0099] In one embodiment of the present invention, two types of SAR image data, namely, offshore scene and nearshore port scene, are selected, and the complete solution provided by the present invention is compared with the solution after replacing the feature balance network in the present invention with other structures. Figure 5 (a) is the solution replaced with the ResNet module, (b) is the solution replaced with the deformable convolution module, (c) is the solution replaced with the attention module, and (d) is this solution. Figure 5 The green rectangle in the middle is the correctly detected target frame, the yellow circle indicates the missed target frame, and the red rectangle indicates the target frame that is incorrectly detected by the method. Figure 5 It can be seen that the solution of replacing the feature balance network with the ResNet module not only has many missed detections, but also has false detections in the land area; the solution after replacing the feature balance network with the deformable convolution has a good ship feature extraction capability, and can detect ship targets with smaller targets, near-shore interference and other more difficult targets, but it also has false detections in the land area; the solution after replacing the feature balance network with the attention module can better distinguish between land areas and ship areas, and can reduce false detections, but there will still be missed detections and wrong detections. This method cleverly combines the capabilities of the deformable convolution and attention modules, can detect ship targets more accurately, and greatly reduces false detections. For some smaller target ships, it has better detection accuracy.

[0100] In this embodiment, experiments are also conducted on the public dataset HRSID. Various replaced methods are trained on the HRSID training set. The test results on the HRSID test set are shown in Table 1 below. The test index is calculated using the coco dataset test index calculation method. The test index used in conventional target detection tasks is the average detection accuracy (mAP) with an intersection-over-union ratio of 0.5 as the threshold. The AP used in this embodiment represents the average AP with an intersection-over-union ratio of 0.5 between the predicted box and the true annotation box in steps of 0.05 to 0.95, which is a more stringent test index. APs represents the test accuracy on small targets, APm represents the test accuracy on medium-sized targets, and AP1 represents the test accuracy on large-scale targets.

[0101] Table 1: Comparison of experimental results

[0102]

[0103] It can be seen from Table 1 that this method has a significant improvement effect on the detection accuracy of SAR image ship applications; it also has a significant improvement effect on ship target detection at multiple scales.

[0104] In another embodiment, the present invention is compared with four common detection methods, namely Faster RCNN, FCOS, ROITransformer and Oriented RCNN, and is also tested on the HRSID dataset. The experimental results are shown in Table 2 below.

[0105] Table 2: Comparison of experimental results

[0106]

[0107] It can be seen from Table 2 that compared with the target detection methods in the conventional optical field, this method has obvious improvement effects on the difficulties of SAR image ship applications and has better detection accuracy than other target detection methods.

[0108] In summary, the present invention combines the local features of deformable convolution with the global features based on attention, which can better provide more robust and balanced image features for ship detection. The present invention adopts a multi-stage target rotation box representation prediction, which can more smoothly predict the coordinate information and angle information of the target rotation box, and better extract the ship features in the top view, thereby improving the detection accuracy.

Claims

1. A ship target detection method in SAR images based on feature balance network. It is characterized in that The following steps are involved: S1, obtaining a SAR image to be detected and preprocessing it to obtain a preprocessed SAR image; S2, extracting features from the preprocessed SAR image through a feature balancing network to obtain multi-level features; S3, input the multi-level features into the FPN network to obtain a multi-scale feature map; S4, input the multi-scale feature map into the RPN network to obtain the coordinate information of the horizontal candidate box of the ship target; S5, mapping the coordinate information of the horizontal candidate box of the ship target obtained in step S4 to the multi-scale feature map, calculating the regional features of the candidate box by using the ROI Align algorithm, and pooling the calculation results to generate a first target box feature vector; S6. Use the first RCNN detection head to predict the first target frame feature vector to obtain first target category information, first target rotation frame coordinate information, and first target rotation frame angle information; S7, mapping the first target rotation frame obtained in step S6 to the multi-scale feature map, calculating the candidate frame region features through the ROI Align algorithm, and pooling the calculation results to generate a second target frame feature vector; S8, using the second RCNN detection head to predict the second target frame feature vector, to obtain second target category information, second target rotation frame coordinate information, and second target rotation frame angle information; S9, outputting the second target category information and the second target rotation frame as the current ship target detection result; In step S2, the characteristic balancing network includes a first characteristic balancing module, a second characteristic balancing module, a third characteristic balancing module and a fourth characteristic balancing module connected in sequence; Each feature balancing module includes a first convolution layer, a second convolution layer, a deformable convolution layer, a spatial attention layer, a feature concatenation layer, and a channel attention layer connected in sequence; the specific method for extracting features from an input image by the feature balancing module includes the following sub-steps: S2-1, performing different convolution operations on the input image through the first convolution layer and the second convolution layer respectively, and obtaining a first convolution result and a second convolution result whose channel dimension is half of the input image respectively; S2-2, performing a deformable convolution operation on the second convolution result through a deformable convolution layer; S2-3, perform spatial attention operation on the deformable convolution operation result through the spatial attention layer; S2-4, concatenate the spatial attention operation result and the first convolution result through the feature concatenation layer, and use the concatenation result as the input of the channel attention layer; S2-5, adjust the weight of the concatenation result along the channel dimension through the channel attention layer to obtain a single-level feature; The four feature balancing modules correspond to the single-level features The multi-level features that make up the feature balancing network.

2. According to the SAR image ship target detection method based on feature balance network as described in claim 1, It is characterized in that The specific method of step S1 includes the following sub-steps: S1-1, standardizing the SAR image to be detected by using the pre-set mean and variance of the image data; S1-2, normalizing the standardized SAR image to be detected; S1-3, performing zero-filling on the normalized SAR image to be detected so that the width and height of the filled image are both integer multiples of 32, thereby obtaining a preprocessed SAR image.

3. According to the SAR image ship target detection method based on feature balance network as described in claim 1, It is characterized in that In step S3, the FPN network includes four 1×1 convolutional layers, namely, the first 1×1 convolutional layer, the second 1×1 convolutional layer, the third 1×1 convolutional layer and the fourth 1×1 convolutional layer; the input of the first 1×1 convolutional layer is the single-level feature output by the first feature balancing module; The input of the second 1×1 convolutional layer is the single-level features output by the second feature balancing module; The input of the third 1×1 convolutional layer is the single-level feature output by the third feature balancing module; The input of the fourth 1×1 convolutional layer is the single-level features output by the fourth feature balancing module; where: The output of the fourth 1×1 convolutional layer is the first scale feature map; The first scale feature map is upsampled by 2 times and concatenated with the output of the third 1×1 convolutional layer to form the second scale feature map. The concatenation of the second-scale feature map after 2-fold upsampling and the output of the second 1×1 convolutional layer is the third-scale feature map; The third-scale feature map is upsampled by 2 times and concatenated with the output of the first 1×1 convolutional layer to obtain the fourth-scale feature map. The first scale feature map, the second scale feature map, the third scale feature map and the fourth scale feature map together constitute a multi-scale feature map.

4. According to the method for detecting ship targets in SAR images based on feature balance network in claim 1, It is characterized in that In step S4, if the confidence of the horizontal candidate box of the ship target is lower than 0.05, the horizontal candidate box of the ship target is deleted.

5. According to the method for detecting ship targets in SAR images based on feature balance network as described in claim 1, It is characterized in that The sizes of the first target frame feature vector and the second target frame feature vector are both 7×7.

6. The method for detecting ship targets in SAR images based on feature balance network according to claim 1, It is characterized in that In step S6 and step S8, if the confidence of the target rotation box is lower than 0.05, the target rotation box is deleted.

7. A system based on the SAR image ship target detection method based on feature balance network according to any one of claims 1 to 6, It is characterized in that It includes an image preprocessing module, a feature balancing network, an FPN network, an RPN network, a first target frame feature vector calculation module, a first RCNN detection head, a second target frame feature vector calculation module and a second RCNN detection head; wherein: An image preprocessing module is used to preprocess the SAR image to be detected to obtain a preprocessed SAR image; Feature balancing network, used to extract features from preprocessed SAR images to obtain multi-level features; FPN network, used to obtain multi-scale feature maps corresponding to multi-level features; RPN network, used to process the multi-scale feature map and obtain the coordinate information of the horizontal candidate box of the ship target; The first target frame feature vector calculation module is used to map the coordinate information of the horizontal candidate frame of the ship target obtained by the RPN network into the multi-scale feature map, calculate the candidate frame region features through the ROI Align algorithm, and pool the calculation results to generate the first target frame feature vector; The first RCNN detection head is used to predict the first target frame feature vector to obtain first target category information, first target rotation frame coordinate information, and first target rotation frame angle information; A second target frame feature vector calculation module is used to map the first target rotation frame obtained by the first RCNN detection head into a multi-scale feature map, calculate the candidate frame region features through the ROI Align algorithm, and pool the calculation results to generate a second target frame feature vector; The second RCNN detection head is used to predict the second target frame feature vector to obtain the second target category information, the second target rotation frame coordinate information and the second target rotation frame angle information; The detection result output module is used to output the second target category information and the second target rotation frame as the current ship target detection result.

Images