A method and device for detecting aircraft in SAR images

Abstract

The application discloses a SAR image airplane detection method and device, comprising: acquiring a SAR image to be detected; and detecting the SAR image according to a preset airplane detection model, so that the airplane detection model first extracts scattering features of the SAR image by using a preset network, obtains a plurality of first feature maps with different scales, then fuses features of the plurality of first feature maps according to a preset feature pyramid network, obtains a second feature map, and then sequentially performs candidate region determination and size unification on the second feature map and then processes the second feature map, and outputs a SAR image detection result; wherein the SAR image detection result comprises position information of a plurality of candidate targets on the SAR image; and the candidate target is an airplane. The method can effectively extract discrete feature information of an airplane target on the SAR image, and identify position information of the airplane target with discreteness, thereby improving the SAR image airplane detection effect.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of radar image target detection, and particularly relates to a SAR image airplane detection method and device. BACKGROUND

[0002] Synthetic Aperture Radar (SAR) is an active microwave imaging system on the ground, which has the advantages of all-weather, all-day, and strong penetration, and is widely used in disaster monitoring, environmental monitoring, military reconnaissance and other fields. Target detection is an important topic of SAR image intelligent interpretation, and airplanes are an important target of SAR image target detection. It is of great significance to accurately detect airplane targets using the unique advantages of SAR imaging.

[0003] Traditional SAR image airplane target detection algorithms are mainly based on constant false alarm rate algorithms. However, with the rapid development of SAR imaging systems, people can obtain high-resolution SAR images containing a large amount of target and background area detail information. However, traditional methods are difficult to handle the complex features of high-resolution SAR images, so deep learning with strong representation ability gradually becomes the mainstream method of SAR image airplane detection with the support of a large number of high-resolution SAR images. Convolutional Neural Network (CNN) is an important model in deep learning technology, which has excellent performance in processing image data and is widely used in target detection field, and has become the mainstream method of SAR image airplane detection.

[0004] Compared with traditional detection methods, SAR image airplane detection technology based on convolutional neural network fully utilizes the feature extraction ability of deep learning technology, and the target detection effect is improved. However, since the airplane target has a highly detailed structure, it appears in SAR images as a series of discrete points obtained by backscattering of each component, that is, the airplane in SAR image is composed of a series of discrete points, and does not have the continuous edge features like optical images. The convolution operator in the convolutional neural network structure is difficult to capture the information of these discrete points, resulting in high false alarm and high missing detection, etc. Therefore, directly applying the convolutional neural network target detection method in computer vision field still faces some challenges. SUMMARY

[0005] The present application provides a SAR image airplane detection method and device to solve the technical problem that the convolutional neural network target detection method cannot fully extract the discrete features presented by the SAR image airplane target, and has high false alarm and high missing detection.

[0006] To solve the above technical problems, the present application embodiment provides a SAR image airplane detection method, comprising:

[0007] acquire a SAR image to be detected;

[0008] detect the SAR image according to a preset aircraft detection model, first extract scattering features of the SAR image using a preset network to obtain a plurality of first feature maps with different scales, then fuse the plurality of first feature maps according to a preset feature pyramid network to obtain a second feature map, and then sequentially perform candidate region determination and size unification on the second feature map and then perform processing to output a SAR image detection result; wherein the SAR image detection result includes position information of a plurality of candidate targets on the SAR image; the candidate target is an aircraft.

[0009] In the process of detecting the SAR image using the aircraft detection model, the scattering features of the SAR image are extracted to obtain a plurality of first feature maps with different scales, the feature pyramid network is used to fuse the first feature maps with different scales to output the second feature map, and the second feature map is processed after candidate region determination and size unification, thereby outputting the SAR image detection result. This method can effectively extract the discrete feature information of the aircraft target on the SAR image and identify the position information of the aircraft target with discreteness, thereby improving the SAR image aircraft detection effect.

[0010] As a preferred scheme, the SAR image is first extracted using a preset network to obtain a plurality of first feature maps with different scales, specifically:

[0011] The SAR image is extracted and down-sampled using a convolution layer and a max-pooling layer to obtain a processed SAR image;

[0012] The processed SAR image is extracted using a scattering information extraction unit after a plurality of residual units to obtain a first feature map after feature extraction;

[0013] The first feature map after feature extraction is down-sampled using a convolution kernel to obtain a final processed first feature map.

[0014] As a preferred scheme, the processed SAR image is extracted using a scattering information extraction unit after a plurality of residual units to obtain a first feature map after feature extraction, specifically:

[0015] The first feature map after residual unit processing is obtained after a plurality of residual units are extracted;

[0016] The first feature map after residual unit processing is obtained by performing feature extraction on the first feature map after residual unit processing through a deformable difference convolution module;

[0017] The first feature map after processing is obtained by processing the first feature map after processing through an adaptive noise reduction module.

[0018] As a preferred solution, the first feature map after processing is obtained by performing feature extraction on the first feature map after processing through the residual unit through a deformable difference convolution module, specifically:

[0019] The first feature map after processing is obtained by performing channel dimension reduction on the first feature map after processing through the residual unit using convolution;

[0020] The first feature map after gradient information extraction is obtained by extracting gradient information of the first feature map after processing through central difference convolution;

[0021] The first feature map after target discrete feature extraction is obtained by extracting target discrete features in the first feature map after gradient information extraction using deformable convolution;

[0022] The first feature map after processing is obtained by restoring the channel number of the first feature map after target discrete feature extraction using convolution and performing cross-layer connection with the first feature map after processing through the residual unit to realize feature fusion.

[0023] In the embodiment of the application, the channel number of the feature map is reduced through convolution, and the gradient information of the feature map is extracted through central difference convolution, which enhances the scattering center features of the target, and then deformable convolution is used to extract the target discrete features in the feature map, which can better locate the discrete points of the target, and is beneficial to extracting the discrete features of the SAR image aircraft target.

[0024] As a preferred solution, the first feature map after processing is obtained by processing the first feature map after processing through an adaptive noise reduction module, specifically:

[0025] The first feature map after processing is obtained by highlighting important regions of the first feature map after feature extraction using a spatial attention network;

[0026] The first feature map after feature extraction is obtained by thresholding the first feature map after processing using a soft threshold denoising network.

[0027] In the embodiment of the application, the adaptive noise reduction module highlights the important regions of the feature map by using a spatial attention network, and then thresholding the feature map through a soft threshold denoising network, which highlights the key information and suppresses the background noise, thereby improving the effect of SAR image aircraft detection.

[0028] As a preferred solution, the second feature map is sequentially processed after candidate region determination and size unification, and the SAR image detection result is output, specifically:

[0029] The region generation network is used to set a plurality of anchor boxes corresponding to each pixel point of the plurality of second feature maps;

[0030] The plurality of anchor boxes are subjected to non-maximum suppression to output corresponding candidate regions;

[0031] The position information of the plurality of candidate regions is used to extract a plurality of feature maps of the corresponding regions in the second feature map, and the size of the second feature map is unified to obtain a processed second feature map;

[0032] The processed second feature map is subjected to full connection network processing to output a SAR image detection result.

[0033] As a preferred solution, the SAR image detection result is output, specifically, a position box is used to identify the target position in the SAR image.

[0034] As a preferred solution, the aircraft detection model is obtained by training a SAR image aircraft detection network using a training sample set, and the training manner is specifically as follows:

[0035] A SAR image airport slice training sample set is obtained, and the training sample set includes a plurality of SAR image airport slices and a label file corresponding to each slice;

[0036] The training sample set is input into the SAR image aircraft detection network to train the SAR image aircraft detection network, and an aircraft detection model trained is obtained.

[0037] To solve the same technical problem, the embodiment of the application also provides a SAR image aircraft detection device, which comprises an image acquisition module and an image detection module.

[0038] The image acquisition module is configured to acquire a SAR image to be detected.

[0039] The image detection module is configured to detect the SAR image according to a preset aircraft detection model, extract scattering features of the SAR image using a preset network to obtain a plurality of first feature maps with different scales, perform feature fusion on the plurality of first feature maps according to a preset feature pyramid network to obtain a second feature map, and then sequentially perform candidate region determination, size unification, and full connection layer processing on the second feature map to output a SAR image detection result; wherein the SAR image detection result comprises position information of a plurality of candidate targets on the SAR image; and the candidate target is an aircraft.

[0040] As a preferred solution, the image detection module comprises a scattering information extraction network sub-module, a feature pyramid network sub-module, a region generation network sub-module, a region of interest pooling sub-module, and a full connection network sub-module.

[0041] The scattering information extraction network sub-module is configured to extract scattering features of the SAR image to obtain a plurality of first feature maps with different scales.

[0042] The feature pyramid network sub-module is configured to fuse the first feature maps and output a second feature map.

[0043] The region generation network sub-module is configured to process the second feature map to generate a candidate region.

[0044] The region of interest pooling sub-module is configured to unify the sizes of the second feature map and the candidate region.

[0045] The fully connected network sub-module is configured to perform position regression and class judgment on the feature map with unified size, and finally output a detection result. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 FIG. 1 is a flowchart of an embodiment of the SAR image airplane detection method provided by the present application;

[0047] Figure 2 FIG. 2 is a flowchart of a scattering information extraction unit of an embodiment of the SAR image airplane detection method provided by the present application;

[0048] Figure 3 FIG. 3 is a center difference convolution diagram of an embodiment of the SAR image airplane detection method provided by the present application;

[0049] Figure 4 FIG. 4 is a deformable convolution diagram of an embodiment of the SAR image airplane detection method provided by the present application;

[0050] Figure 5 FIG. 5 is a soft threshold denoising network diagram of an embodiment of the SAR image airplane detection method provided by the present application;

[0051] Figure 6 FIG. 6 is a detection result diagram of an embodiment of the SAR image airplane detection method provided by the present application;

[0052] Figure 7 FIG. 7 is a structural diagram of another embodiment of the SAR image airplane detection method provided by the present application. DETAILED DESCRIPTION

[0053] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the present application.

[0054] Embodiment one

[0055] Please refer to Figure 1 , Figure 1 A flowchart of a SAR image aircraft detection method provided by the present application, comprising steps 101 to 102, and the specific steps are as follows:

[0056] S101: Obtain a SAR image to be detected.

[0057] In this embodiment, a SAR airport image is obtained.

[0058] S102: Detect the SAR image according to a preset aircraft detection model, so that the aircraft detection model first uses a preset network to extract scattering features of the SAR image, obtains a plurality of first feature maps with different scales, and then performs feature fusion on the plurality of first feature maps according to a preset feature pyramid to obtain a second feature map, and then sequentially performs candidate region determination, size unification, and full connection layer processing on the second feature map, and outputs the SAR image detection result; wherein the SAR image detection result comprises position information of a plurality of candidate targets on the SAR image; the candidate target is an aircraft.

[0059] In this embodiment, after the SAR image is input into the aircraft detection model, the scattering information extraction backbone network is used to extract features of the SAR image to obtain a plurality of feature maps with different scales, and then the feature fusion is performed on the obtained feature maps through the feature pyramid network to output a multi-scale feature map. After the feature pyramid network receives the multi-scale feature map, the target candidate region is determined on the multi-scale feature map, and then the feature map and the target candidate region are unified in size. The size-unified feature map is processed through the full connection layer to identify and accurately position the candidate target, and finally the SAR image aircraft detection result is output.

[0060] Optionally, the scattering features of the SAR image are extracted using a preset network to obtain a plurality of first feature maps with different scales, and the specific steps are as follows:

[0061] The convolution layer and the max-pooling layer are used to extract features and downsample the SAR image to obtain a processed SAR image;

[0062] After the plurality of residual units, the scattering information extraction unit is used to extract features of the processed SAR image to obtain the first feature map after feature extraction;

[0063] The convolution kernel is used to downsample the first feature map after feature extraction to obtain the final processed first feature map.

[0064] In this embodiment, the input SAR airport image is extracted by the scattering information extraction backbone network, and first, the input SAR image is extracted and down-sampled by the convolution layer and the maximum pooling layer. Then, in the first stage of the scattering information extraction backbone network, a scattering information extraction unit is used after two residual units to output a feature map. In the second stage, a scattering information extraction unit is used after three residual units to output a feature map. In the third stage, a scattering information extraction unit is used after five residual units to output a feature map. In the fourth stage, a scattering information extraction unit is used after two residual units to output a feature map. Finally, a convolution kernel is used to process the above feature map by the convolution layer to obtain a feature map. The size of the feature map obtained in each stage gradually decreases.

[0065] For example, a 512x512 image is input, and first, the 7x7 convolution layer and the maximum pooling layer are used for feature extraction and down-sampling. Then, in the first stage of the scattering information extraction backbone network, a scattering information extraction unit is used after two residual units to output a 64x64 feature map. In the second stage, a scattering information extraction unit is used after three residual units to output a 32x32 feature map. In the third stage, a scattering information extraction unit is used after five residual units to output a 16x16 feature map. In the fourth stage, a scattering information extraction unit is used after two residual units to output an 8x8 feature map. Finally, a 3x3 convolution kernel is used in the last convolution layer to output a 4x4 feature map.

[0066] Optionally, the scattering information extraction unit is used after the multiple residual units to extract features from the processed SAR image to obtain the first feature map after feature extraction, which is specifically:

[0067] The first feature map after the residual unit processing obtained after the multiple residual units are extracted;

[0068] The first feature map after the residual unit processing is obtained by the deformable difference convolution module for feature extraction of the first feature map after the residual unit processing;

[0069] The first feature map after feature extraction is obtained by the adaptive noise reduction module for processing the first feature map after processing.

[0070] In this embodiment, the SAR image is input into the scattering information extraction unit after the residual unit. The scattering information extraction unit first extracts the discrete scattering center point features of the SAR image aircraft target in the feature map by the deformable difference convolution module, and then highlights the key information and suppresses the background noise of the feature map by the adaptive noise reduction module to obtain the feature map after scattering feature extraction.

[0071] Optionally, the first feature map processed by the residual unit is processed by the deformable difference convolution module to obtain a processed first feature map, specifically:

[0072] The first feature map processed by the residual unit is processed by using convolution to reduce the channel number to obtain the processed first feature map;

[0073] The gradient information of the processed first feature map is extracted by using central difference convolution to obtain the first feature map with extracted gradient information;

[0074] The target discrete feature in the first feature map with extracted gradient information is extracted by using deformable convolution to obtain the first feature map with extracted target discrete feature;

[0075] The first feature map with extracted target discrete feature is processed by using convolution to restore the channel number, and is cross-layer connected with the first feature map processed by the residual unit to realize feature fusion, thereby obtaining the processed first feature map.

[0076] In the embodiment, as shown in Figure 2 , the deformable difference convolution module first reduces the channel number of the input feature map by convolution; the central difference convolution first performs central difference calculation on the convolution kernel and the corresponding region of the feature map to extract the gradient information of the feature map, as shown in Figure 3 ; as shown in Figure 4 , the deformable convolution is used to first predict the offset of the pixel points of the convolution kernel, locate the scattering center point of the target in the feature map, and finally use convolution to restore the channel number of the feature map and cross-layer connect with the input feature map to realize feature fusion, thereby outputting the feature map.

[0077] Optionally, the processed first feature map is processed by the adaptive noise reduction module to obtain the first feature map with extracted features, specifically:

[0078] The important region of the first feature map with extracted features is highlighted by using the spatial attention network to obtain the processed first feature map;

[0079] The processed first feature map is thresholded by using the soft threshold denoising network to obtain the first feature map with extracted features.

[0080] In the embodiment, the spatial attention network is first used to highlight the important region of the feature map from the spatial direction; as shown in Figure 5 , the soft threshold denoising network is used to threshold the feature map from the channel direction by using a group of adaptive thresholds learned by the model, thereby obtaining the feature map processed by the scattering feature extraction, background noise suppression and target feature highlighting.

[0081] According to the obtained feature maps, a plurality of different feature maps are fused by a feature pyramid network to output a multi-scale feature map.

[0082] Optionally, the second feature map is sequentially subjected to candidate region determination and size unification, and then processed to output a SAR image detection result, specifically:

[0083] An anchor box corresponding to each pixel point of each of the plurality of second feature maps is set using a region generation network.

[0084] The plurality of anchor boxes are subjected to non-maximum suppression to output a corresponding candidate region.

[0085] Using position information of the plurality of candidate regions, a plurality of feature maps of corresponding regions in the second feature map are extracted, and the plurality of feature maps are subjected to size unification to obtain a processed second feature map.

[0086] The processed second feature map is subjected to full connection layer processing to output a SAR image detection result.

[0087] In this embodiment, a region generation network is used to receive the multi-scale feature map, generate a plurality of anchor boxes at each pixel point of each feature map, perform non-maximum suppression on the plurality of anchor boxes generated by the region candidate unit to output a plurality of regions of interest as target candidate regions, then input the feature map and the position information of the region of interest into a region of interest pooling unit, perform size unification on the feature map of the corresponding region of the region of interest in the second feature map, then perform full connection layer processing on the feature map after size unification to identify the candidate target and mark the position information, and finally output a SAR image aircraft detection result.

[0088] Optionally, the SAR image detection result is output, specifically:

[0089] The position of the target in the SAR image is marked by a position box.

[0090] In this embodiment, the SAR airport image is output to perform aircraft target detection, and the position of the target is marked by a position box in the image, as shown in Figure 6 , all aircraft targets in the SAR airport image are correctly marked using the method.

[0091] Exemplarily, the SAR image aircraft detection model proposed in the present application is compared with other models in terms of detection effect, and it can be seen that the recall rate, accuracy rate and mAP of the detection model adopted in the present application are obviously improved, and it can be seen that the detection effect output by this embodiment is better.

[0092] Table 1: Comparison of detection effects

[0093] Recall (%) Accuracy (%) mAP (%) Detection model using RetinaNet 91.9 82.2 89.0 Detection model using Faster R-CNN 91.8 91.7 90.2 Detection model using network with scattering information extraction 96.6 92.9 95.4

[0094] Optionally, the aircraft detection model is obtained by training the aircraft detection model using a training sample set, and the training manner is specifically as follows:

[0095] An SAR image airport slice training sample set is obtained, and the training sample set includes a plurality of SAR image airport slices and a label file corresponding to each slice;

[0096] The training sample set is input into the SAR image aircraft detection network for training, and a trained aircraft detection model is obtained.

[0097] By introducing a scattering extraction module into the scattering information extraction backbone network to extract scattering features of the SAR image, and by designing a deformable convolution module and an adaptive noise reduction module in the scattering extraction module to extract scattering information and suppress background noise, the position information of an aircraft target with weak scattering information and discreteness can be identified on the SAR image after feature fusion of the feature map by a feature pyramid, region selection, size unification, and full connection processing, thereby effectively improving the SAR image aircraft detection effect and alleviating the problems of high false alarm and high missed detection in the current detection method.

[0098] Embodiment Two

[0099] Correspondingly, referring to Figure 7 , Figure 7 is a structural schematic diagram of the SAR image aircraft detection device provided by the present application. As Figure 7 shown, the SAR image aircraft detection device includes an image acquisition module 701 and an image detection module 702, wherein the specific units of each module are as follows:

[0100] The image acquisition module is configured to acquire an SAR image to be detected.

[0101] The image detection module is configured to detect the SAR image according to a preset aircraft detection model, so that the aircraft detection model first uses a preset network to extract scattering features of the SAR image, obtains a plurality of first feature maps with different scales, then performs feature fusion on the plurality of first feature maps according to a preset feature pyramid network to obtain a second feature map, and then sequentially performs candidate region determination, size unification, and processing on the second feature map to output an SAR image detection result; wherein the SAR image detection result includes position information of a plurality of candidate targets on the SAR image; and the candidate target is an aircraft.

[0102] In the embodiment, the image detection module comprises a scattering information extraction network submodule 7021, a feature pyramid network submodule 7022, a region generation network submodule 7023, a region of interest pooling submodule 7024, and a fully connected network submodule 7025.

[0103] The scattering information extraction network submodule 7021 is configured to perform scattering feature extraction on the SAR image to obtain a plurality of first feature maps with different scales.

[0104] The scattering information extraction network submodule 7021 comprises a convolution layer unit 70211, a max-pooling layer unit 70212, a residual unit 70213, and a scattering information extraction unit 70214.

[0105] The scattering information extraction unit 70214 comprises a deformable difference convolution subunit 702141 and an adaptive noise reduction subunit 702142. The deformable difference convolution subunit 702141 is configured to extract and fuse discrete features of a SAR image target. The adaptive noise reduction subunit 702142 is configured to highlight key information and suppress background noise of the SAR image.

[0106] The feature pyramid network submodule 7022 is configured to perform feature fusion on the first feature maps and output a second feature map.

[0107] The region generation network submodule 7023 is configured to process the second feature map to generate a candidate region.

[0108] The region of interest pooling submodule 7024 is configured to unify the sizes of the second feature map and the candidate region.

[0109] The fully connected network submodule 7025 is configured to perform position regression and class judgment on the feature map with unified size, and finally output a detection result.

[0110] The SAR image airplane detection method can refer to the content described above for more detailed processes, but is not limited thereto.

[0111] Compared with the prior art, for the discreteness of the SAR image aircraft target, a SAR image aircraft detection model based on deep learning and target discreteness is constructed, the discrete features of the SAR image aircraft target are fully extracted, a plurality of feature maps with different scales are obtained, then a feature pyramid is used to fuse the feature maps with different scales to output a second feature map, the second feature map is subjected to region selection, size unification, full connection processing, and finally the position information of the aircraft target is marked on the SAR image. In the aircraft detection model process, by using the scattering information extraction unit, the discrete points of the target can be better located, which is beneficial to extract the discrete features of the SAR image aircraft target, and highlights the key information and suppresses the background noise, so that the SAR image aircraft detection effect can be effectively improved.

[0112] The above specific embodiments further illustrate the purpose, technical scheme and beneficial effects of the present application, and it should be understood that the above only describes specific embodiments of the present application and is not used to limit the protection scope of the present application. It is particularly pointed out that any modification, equivalent replacement, improvement, etc. made by those skilled in the art within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for detecting an aircraft in a SAR image, characterized in that The method comprises the following steps: acquiring a SAR image to be detected; detecting the SAR image according to a preset aircraft detection model, so that the aircraft detection model first uses a convolution layer and a maximum pooling layer to perform feature extraction and down-sampling on the SAR image, and obtain a processed SAR image; obtaining a first feature map processed by a residual unit after feature extraction by a plurality of residual units; using convolution to reduce the channel number of the first feature map processed by the residual unit, and obtaining a processed first feature map; extracting gradient information of the processed first feature map by central difference convolution, and obtaining a first feature map after extracting gradient information; extracting target discrete features in the first feature map after extracting gradient information by deformable convolution, and obtaining a first feature map after extracting target discrete features; restoring the channel number of the first feature map after extracting target discrete features by convolution, and performing cross-layer connection with the first feature map processed by the residual unit to realize feature fusion, thereby obtaining a processed first feature map; processing the processed first feature map by an adaptive noise reduction module to obtain a first feature map after extracting features; down-sampling the first feature map after extracting features by a convolution kernel to obtain a final processed first feature map; further performing feature fusion on the plurality of first feature maps according to a preset feature pyramid network to obtain a second feature map, and then sequentially performing candidate region determination, size unification and processing on the second feature map to output a SAR image detection result; wherein the SAR image detection result comprises position information of a plurality of candidate targets on the SAR image; the candidate target is an aircraft.

2. The SAR image aircraft detection method of claim 1, wherein, The further processing of the processed first feature map by the adaptive noise reduction module to obtain the first feature map after extracting features comprises: using a spatial attention network to highlight important regions of the first feature map after extracting features to obtain a processed first feature map; using a soft threshold denoising network to perform thresholding processing on the processed first feature map to obtain a first feature map after extracting features.

3. The SAR image aircraft detection method of claim 1, wherein, The sequential processing of the second feature map after candidate region determination and size unification to output the SAR image detection result comprises: using a region generation network to set a plurality of anchor boxes corresponding to each pixel point of the plurality of second feature maps; performing non-maximum suppression on the plurality of anchor boxes to output corresponding candidate regions; using position information of the plurality of candidate regions to extract a plurality of feature maps of corresponding regions in the second feature map, and performing size unification to obtain a processed second feature map; performing full connection network processing on the processed second feature map to output a SAR image detection result.

4. The SAR image aircraft detection method of claim 1, wherein, The output of the SAR image detection result comprises marking the target position with a position box in the SAR image.

5. The SAR image aircraft detection method of claim 1, wherein, The aircraft detection model is obtained by training a SAR image aircraft detection network using a training sample set, and the training method comprises the following steps: An SAR image airport slice training sample set is acquired, and the training sample set includes a plurality of SAR image airport slices and a label file corresponding to each slice; The training sample set is input into an SAR image airplane detection network for training, and a trained airplane detection model is obtained.

6. A SAR image aircraft detection apparatus characterized by comprising: The SAR image airplane detection method according to any one of claims 1-5, comprising an image acquisition module and an image detection module; The image acquisition module is configured to acquire an SAR image to be detected; The image detection module is configured to detect the SAR image according to a preset airplane detection model, so that the airplane detection model first extracts scattering features of the SAR image using a preset network to obtain a plurality of first feature maps with different scales, then fuses features of the plurality of first feature maps according to a preset feature pyramid network to obtain a second feature map, and then sequentially determines a candidate region, unifies sizes, and then processes the second feature map to output an SAR image detection result; wherein the SAR image detection result includes position information of a plurality of candidate targets on the SAR image; and the candidate target is an airplane.

7. The SAR image aircraft detection apparatus of claim 6, wherein The image detection module includes a scattering information extraction network submodule, a feature pyramid network submodule, a region generation network submodule, a region of interest pooling submodule, and a fully connected network submodule; The scattering information extraction network submodule is configured to extract scattering features of the SAR image to obtain a plurality of first feature maps with different scales; The feature pyramid network submodule is configured to fuse features of the first feature maps and output a second feature map; The region generation network submodule is configured to process the second feature map to generate a candidate region; The region of interest pooling submodule is configured to unify sizes of the second feature map and the candidate region; The fully connected network submodule is configured to perform position regression and class judgment on the feature map after size unification, and finally output a detection result.

Images