Aerial target detection method based on transformer and jump path aggregation FPN

Abstract

The application discloses an aerial target detection method based on a transformer and a skip-path aggregation FPN, and comprises the following steps: acquiring aerial image data to be detected; inputting the aerial image data to be detected into a pre-trained aerial small target detection network model to perform target detection, so as to obtain a target detection result of the aerial image data to be detected; wherein the aerial small target detection network model is a deep aggregation residual transformation network combined with a deformable convolution and a transformer module to perform feature extraction, a skip-path aggregation feature pyramid network to perform feature fusion, and a double-path decoupling detection head to perform target detection. In the picture captured by the aerial view of the unmanned aerial vehicle, higher detection precision can be achieved, especially in the scene where a large number of small targets exist.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of target detection, and particularly relates to a method for aerial target detection based on a transformer and a jump path aggregation FPN. BACKGROUND

[0002] Aerial target detection technology is a technology that combines unmanned aerial vehicles (UAVs), high-resolution cameras, computer vision, and deep learning algorithms, aiming to realize the automatic identification, positioning, and tracking of ground targets. In recent years, due to the rapid development of UAV technology and the improvement of computing power, aerial target detection technology has been widely applied in intelligent transportation, agricultural monitoring, security monitoring, and other fields. Target detection technology mainly includes traditional target detection methods and deep learning-based target detection methods. Traditional target detection methods mainly include sliding window-based detection methods, clustering-based detection methods, and image segmentation-based detection methods. These methods rely on hand-designed feature extractors such as SIFT, HOG, LBP, etc., and have the advantages of fast computing speed and good real-time performance. However, these methods are sensitive to changes in target size, shape, pose, and illumination, which can easily lead to a decline in detection performance. Deep learning-based target detection methods mainly include two categories: region-based convolutional neural networks (R-CNN series) and single-stage target detection networks (such as YOLO, SSD, etc.). R-CNN series methods mainly include R-CNN, Fast R-CNN, Faster R-CNN, etc. These methods first generate candidate target regions using a region proposal network (RPN), then extract features for these regions using a convolutional neural network (CNN), and perform target classification and position regression through a fully connected layer. R-CNN series methods have high detection accuracy, but the computing speed is slow and the real-time performance is poor. Single-stage target detection networks such as YOLO and SSD directly perform feature extraction and target detection on the entire image without generating candidate regions. These methods have fast computing speed and good real-time performance, but the detection accuracy is relatively low.

[0003] Compared with traditional ground images, aerial images have the following characteristics: high resolution, large field of view, large change in viewing angle, diverse target scale and posture, etc. These characteristics pose higher requirements for aerial target detection technology, making the performance of traditional target detection methods unsatisfactory on aerial images. In view of the characteristics of aerial images, researchers have proposed many more adaptable target detection methods. Among them, a common method is to combine deep learning with traditional feature extraction methods, taking advantage of the powerful representation ability of deep learning and the stability of traditional feature extraction methods to improve the accuracy of aerial target detection. Another method is to improve existing deep learning target detection networks, such as introducing multi-scale feature fusion, attention mechanism, etc., to adapt to the diversity of target scale and posture in aerial images. However, existing technologies are still not satisfactory in the field of aerial small target detection, and further improvement is needed. SUMMARY

[0004] The present application provides a kind of based on transformer and jump path aggregation FPN's aerial target detection method, on unmanned aerial vehicle aerial view image has very high small target detection precision.

[0005] The first aspect of the present application provides a kind of based on transformer and jump path aggregation FPN's aerial target detection method, comprising the following steps: obtaining aerial image data to be detected;The aerial image data to be detected is input into the pre-trained aerial small target detection network model to carry out target detection, and the target detection result of the aerial image data to be detected is obtained;Wherein, the aerial small target detection network model is the deep aggregation residual transformation network combined with deformable convolution and transformer module for feature extraction, uses jump path aggregation feature pyramid network for feature fusion, and uses double road decoupling detection head for target detection.

[0006] Optionally, in an embodiment of the present application, the aerial small target detection network model includes a backbone network, a neck network, a region extraction network, a region of interest pooling layer and a detection head;

[0007] The backbone network includes a starting processing layer and four stage processing layers, and the starting processing layer is composed of 3 3x3 convolution and a maximum pooling layer;Each stage processing layer is composed of a plurality of aggregation residual transformation bottleneck levels combined with deformable convolution and transformer, and four feature maps of aerial image data are output through the four stage processing layers;

[0008] The neck network includes the jump path aggregation feature pyramid network, which is used to obtain the four feature maps from the backbone network, and input into the jump path aggregation feature pyramid network after reducing the channel through the convolution module, and the output of the jump path aggregation feature pyramid network is sent to the region extraction network and the region of interest pooling layer;

[0009] The region extraction network is used to generate a candidate region;

[0010] The region of interest pooling layer is used to perform region of interest feature map extraction on the feature map output by the skip-path aggregation FPN using the candidate region generated by the region extraction network, and the region of interest pooling is used to make all feature map outputs have the same size and send the feature map to the detection head;

[0011] The detection head is a cascaded RCNN multi-stage structure, each stage uses a double-decoupling detection head, for the feature map output by the region of interest pooling layer, one path uses four convolution layers for processing and then performs boundary box regression to obtain the target detection box of the aerial image data, and the other path uses two fully connected layers for processing and then performs classification to obtain the target classification result of the aerial image data.

[0012] Optionally, in an embodiment of the present application, the pooling mode in the region of interest pooling layer is modulated deformable region of interest pooling.

[0013] The second aspect embodiment of the present application provides an aerial target detection device based on a transformer and a skip-path aggregation FPN, comprising: an acquisition module configured to acquire aerial image data to be detected;

[0014] A detection module is configured to input the aerial image data to be detected into a pre-trained aerial small target detection network model to perform target detection, and obtain a target detection result of the aerial image data to be detected; wherein the aerial small target detection network model is a deep aggregation residual transformation network combined with deformable convolution and a transformer module for feature extraction, a skip-path aggregation feature pyramid network for feature fusion, and a double-decoupling detection head for target detection.

[0015] Optionally, in an embodiment of the present application, the aerial small target detection network model comprises a backbone network, a neck network, a region extraction network, a region of interest pooling layer, and a detection head;

[0016] The backbone network comprises a starting processing layer and four stage processing layers, the starting processing layer is composed of 3 3x3 convolution and a maximum pooling layer; each stage processing layer is composed of a plurality of aggregation residual transformation bottleneck layers combined with deformable convolution and a transformer, and four feature maps of the aerial image data are output through the four stage processing layers;

[0017] The neck network comprises the skip-path aggregated feature pyramid network, which is used to acquire the four-layer feature maps from the backbone network, and input the four-layer feature maps into the skip-path aggregated feature pyramid network after reducing channels through a convolution module, and output of the skip-path aggregated feature pyramid network is input into the region extraction network and the region of interest pooling layer;

[0018] The region extraction network is used to generate a candidate region.

[0019] The region of interest pooling layer is used to perform region of interest feature map extraction on the feature maps output by the skip-path aggregated feature pyramid network by using the candidate region generated by the region extraction network, and all feature map output sizes are made the same by using region of interest pooling, and the feature maps are input into the detection head.

[0020] The detection head is a cascaded RCNN multi-stage structure, each stage uses a double-path decoupled detection head, for the feature maps output by the region of interest pooling layer, one path uses four convolution layers to process and perform boundary box regression to obtain a target detection box of the aerial image data, and the other path uses two fully connected layers to process and perform classification to obtain a target classification result of the aerial image data.

[0021] Optionally, in an embodiment of the present application, the pooling mode in the region of interest pooling layer is modulated deformable region of interest pooling.

[0022] The third aspect embodiment of the present application provides an electronic device, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, and the processor executes the program to perform the aerial target detection method based on the transformer and the skip-path aggregated FPN as described in the above embodiments.

[0023] The fourth aspect embodiment of the present application provides a computer readable storage medium, which stores a computer program, and the program is executed by a processor to perform the aerial target detection method based on the transformer and the skip-path aggregated FPN as described in the above embodiments.

[0024] The aerial target detection method based on the transformer and the skip-path aggregated FPN of the embodiments of the present application combines the aggregated residual transformation bottleneck layer of the deformable convolution and the transformer, can adaptively adjust the receptive field, capture the long-distance dependence of the image, and emphasize important channel information and suppress noise. The skip-path aggregated pyramid network proposed in the present application can better fuse multi-scale feature information, especially small target information with large resolution. The detection head is decoupled in the present application, and two branches of convolution and full connection are used for boundary box regression and classification, so that the classification pays more attention to the central content area, and the regression pays more attention to the edge information, and the detection effect is improved.

[0025] Additional aspects and advantages of the present application will be apparent from the following description, taken in conjunction with the accompanying drawings, illustrating by way of example the principles of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0026] The above and / or additional aspects and advantages of the present application will become apparent and be more fully understood from the following description, taken in conjunction with the accompanying drawings, in which:

[0027] Figure 1 A flowchart of a method for aerial target detection based on a transformer and a skip-path aggregated FPN according to an embodiment of the present application;

[0028] Figure 2 A schematic diagram of a network model structure for aerial small target detection according to an embodiment of the present application;

[0029] Figure 3 A schematic diagram of a transformer structure according to an embodiment of the present application;

[0030] Figure 4 A schematic diagram of an aggregated residual transformation bottleneck layer structure combining deformable convolution and transformer according to an embodiment of the present application;

[0031] Figure 5 A schematic diagram of a skip-path aggregated feature pyramid structure according to an embodiment of the present application;

[0032] Figure 6 An example diagram of a device for aerial target detection based on a transformer and a skip-path aggregated FPN according to an embodiment of the present application;

[0033] Figure 7 A schematic diagram of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0034] Embodiments of the present application are described in detail below with reference to the attached drawings, which show by way of example, embodiments in which like reference numerals refer to like elements or elements having similar functions. The embodiments described below are examples intended to provide an explanation of the present application and are not intended to restrict the present application.

[0035] Figure 1 A flowchart of a method for aerial target detection based on a transformer and a skip-path aggregated FPN according to an embodiment of the present application.

[0036] As shown in Figure 1 the method for aerial target detection based on a transformer and a skip-path aggregated FPN includes the following steps:

[0037] In step S101, aerial image data to be detected is acquired.

[0038] In the embodiment of the present application, the targets in aerial image data can include pedestrians, cyclists, bicycles, electric vehicles, cars, trucks, buses and the like of various categories.

[0039] In step S102, the aerial image data to be detected is input into the aerial small target detection network model pre-trained to perform target detection, and the target detection result of the aerial image data to be detected is obtained; wherein the aerial small target detection network model is a deep aggregation residual transformation network combined with deformable convolution and transformer module for feature extraction, a skip path aggregation feature pyramid network for feature fusion, and a dual-path decoupling detection head for target detection.

[0040] The aerial image data to be detected is input into the aerial small target detection network model pre-trained for target detection, and the target position in the aerial image and the category to which the target belongs can be obtained.

[0041] The embodiment of the present application uses a UAV to collect aerial image data and performs labeling to obtain a data set, divides the data set into a training set, a validation set and a test set, trains, validates and tests the aerial small target detection network model using the data set, and obtains the trained aerial small target detection network model.

[0042] The embodiment of the present application uses a deep aggregation residual transformation network combined with deformable convolution and transformer module as a backbone network, proposes a skip path aggregation feature pyramid network to strengthen small target feature fusion, adopts a cascaded RCNN detection framework idea, and uses a dual-path decoupling detection head to construct an aerial small target detection network model, which is called DTDHC RCNN.

[0043] The DTDHC RCNN algorithm is trained on the training set, multiple data enhancement strategies are adopted during training, and the best weight file is obtained through parameter optimization of the validation set.

[0044] The best weight file trained is loaded into the DTDHC RCNN algorithm to obtain a target detection network, and then the detection result is tested on the test set, and the subsequently collected aerial image is detected.

[0045] In the embodiment of the present application, as shown in Figure 2 The aerial small target detection network model includes a backbone network, a neck network, a region proposal network (RPN), a region of interest pooling layer and a detection head.

[0046] The backbone network comprises a starting processing layer and four stage processing layers, the starting processing layer is composed of 3 3x3 convolutions and a max pooling layer; each stage processing layer is composed of a plurality of aggregated residual transformation bottleneck layers combined with deformable convolution and transformer, and the P2, P3, P4 and P5 four-layer feature maps of the aerial image data are output through the four stage processing layers. Figure 3 As shown in Figure 4 As shown in

[0047] The neck network comprises a skip-path aggregated feature pyramid network, which is used to obtain the four-layer feature maps from the backbone network, and after the channel is reduced through a convolution module, the four-layer feature maps are input into the skip-path aggregated feature pyramid network, and the output of the skip-path aggregated feature pyramid network is sent to a region extraction network and a region of interest pooling layer. Figure 5 As shown in

[0048] The region extraction network is used to generate candidate regions.

[0049] The region of interest pooling layer is used to perform region of interest feature map extraction on the feature maps output by the skip-path aggregated feature pyramid network using the candidate regions generated by the region extraction network, and the region of interest pooling is used to make the output sizes of all feature maps the same, and the feature maps are sent to a detection head; the modulation deformable region of interest pooling is used in the pooling mode.

[0050] The detection head is a cascaded RCNN multi-stage structure, and each stage uses a dual-path decoupled detection head. For the feature maps output by the region of interest pooling layer, one path is processed through four convolution layers to perform boundary box regression, and the other path is processed through two fully connected layers to perform classification, so as to obtain the target detection frame of the aerial image data and the target classification result of the aerial image data.

[0051] When training the aerial small target detection network model, an adaptive momentum estimation optimizer with weight decay is used as the optimizer, the initial learning rate is set to 1e-2, the minimum learning rate is 0.0002, the weight decay coefficient is 0.05, the network is trained for 24 rounds, the number of samples for each training is set to 16, and the learning rate scheduling strategy is cosine annealing.

[0052] The aerial target detection method based on the transformer and the jump path aggregation FPN can adaptively adjust a receptive field, capture long-distance dependencies of an image, and emphasize important channel information and suppress noise by combining deformable convolution and aggregated residual transformation bottleneck layers of the transformer.

[0053] Next, the aerial target detection device based on the transformer and the jump path aggregation FPN is described with reference to the accompanying drawings.

[0054] Figure 6 An example diagram of the aerial target detection device based on the transformer and the jump path aggregation FPN according to the embodiment of the present application.

[0055] As shown in the figure, Figure 6 The aerial target detection device 10 based on the transformer and the jump path aggregation FPN includes an acquisition module 100 and a detection module 200.

[0056] The acquisition module 100 is configured to acquire aerial image data to be detected. The detection module 200 is configured to input the aerial image data to be detected into a pre-trained aerial small target detection network model to perform target detection, and obtain a target detection result of the aerial image data to be detected.

[0057] In the embodiment of the present application, the aerial small target detection network model includes a backbone network, a neck network, a region extraction network, a region of interest pooling layer, and a detection head.

[0058] The backbone network includes one starting processing layer and four stage processing layers. The starting processing layer is composed of three 3x3 convolution layers and one max pooling layer. Each stage processing layer is composed of a plurality of aggregated residual transformation bottleneck layers combined with deformable convolution and a transformer. The four stage processing layers output four feature maps of the aerial image data.

[0059] The neck network comprises a skip-path aggregation feature pyramid network, which is used to obtain four layers of feature maps from the backbone network, and after the channel is reduced through a convolution module, the feature maps are input into the skip-path aggregation feature pyramid network, and the output of the skip-path aggregation feature pyramid network is input into a region extraction network and a region of interest pooling layer.

[0060] The region extraction network is used to generate a candidate region.

[0061] The region of interest pooling layer is used to perform region of interest feature map extraction on the feature maps output by the skip-path aggregation feature pyramid network by using the candidate region generated by the region extraction network, and the region of interest pooling is used to make the output sizes of all the feature maps the same, and the feature maps are input into a detection head.

[0062] The detection head is a cascaded RCNN multi-stage structure, each stage uses a double-path decoupling detection head, for the feature maps output by the region of interest pooling layer, one path is processed by four convolution layers to perform boundary box regression, to obtain the target detection box of the aerial image data, and the other path is processed by two fully connected layers to perform classification, to obtain the target classification result of the aerial image data.

[0063] In the embodiment of the present application, the pooling mode in the region of interest pooling layer is modulated deformable region of interest pooling.

[0064] It should be noted that the foregoing explanation and description of the embodiment of the aerial target detection method based on the transformer and the skip-path aggregation FPN also applies to the embodiment of the aerial target detection device based on the transformer and the skip-path aggregation FPN, which will not be described here.

[0065] The aerial target detection device based on the transformer and the skip-path aggregation FPN according to the embodiment of the present application combines the aggregation residual transformation bottleneck layer of the deformable convolution and the transformer, can adaptively adjust the receptive field, capture the long-distance dependence of the image, and emphasize important channel information and suppress noise. The skip-path aggregation pyramid network proposed in the present application can better fuse multi-scale feature information, especially small target information with large resolution. The detection head is decoupled in the present application, and two branches of convolution and full connection are used for boundary box regression and classification respectively, so that the classification pays more attention to the central content area, and the regression pays more attention to the edge information, which improves the detection effect.

[0066] Figure 7 The structure schematic diagram of the electronic device provided in the embodiment of the present application. The electronic device can comprise:

[0067] The memory 701, the processor 702, and the computer program stored in the memory 701 and executable on the processor 702.

[0068] The processor 702 implements the aerial target detection method based on the transformer and the jump path aggregation FPN provided in the above embodiments when executing a program.

[0069] Further, the electronic device further comprises:

[0070] The communication interface 703 is configured to communicate between the memory 701 and the processor 702.

[0071] The memory 701 is configured to store a computer program executable on the processor 702.

[0072] The memory 701 can include a high-speed RAM memory, and can also include a non-volatile memory, for example, at least one disk memory.

[0073] If the memory 701, the processor 702 and the communication interface 703 are independently implemented, the communication interface 703, the memory 701 and the processor 702 can be connected to each other through a bus and complete the communication between each other. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For the convenience of representation, Figure 7 Only one thick line is used in the figure, but it does not mean that there is only one bus or only one type of bus.

[0074] Optionally, in a specific implementation, if the memory 701, the processor 702 and the communication interface 703 are integrated on a chip, the memory 701, the processor 702 and the communication interface 703 can complete the communication between each other through an internal interface.

[0075] The processor 702 can be a central processing unit (CPU), or an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present application.

[0076] The embodiment also provides a computer readable storage medium, which stores a computer program, and the program is executed by a processor to implement the aerial target detection method based on the transformer and the jump path aggregation FPN as above.

[0077] In the description of the present specification, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Also, the specific features, structures, materials or characteristics described can be combined in any one or N embodiments or examples in a suitable manner. In addition, the person skilled in the art can combine and combine the different embodiments or examples described in the present specification and the features of the different embodiments or examples without contradiction.

[0078] In addition, the terms "first", "second" are only for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include at least one of the features. In the description of the present application, the meaning of "N" is at least two, for example, two, three, etc., unless otherwise specifically limited.

[0079] Any process or method descriptions in flow charts or otherwise described herein can be understood as representing code modules, segments, or portions of code that include one or more executable instructions for implementing the specified logical functions or steps, and the various embodiments of the present application include additional implementations in which the functions described in the steps are implemented in the same or a different order, including substantially simultaneously, as well as in reverse order, as will be understood by those skilled in the art. Accordingly, the present application is not limited to the specific implementations described herein, which are exemplary only, unless otherwise specifically stated.

[0080] It should be understood that parts of the present application can be implemented in hardware, software, firmware or a combination thereof. In the above-described embodiments, N steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. As in another embodiment, if implemented in hardware, any of the following technologies known in the art or their combinations can be used: discrete logic circuit with logic gate circuit for implementing logical functions on data signals, application specific integrated circuit with suitable combination of logic gate circuit, programmable gate array (PGA), field programmable gate array (FPGA) and the like.

[0081] Those skilled in the art can understand that all or part of the steps carried out by the above-mentioned embodiment method can be completed by a program instructing the relevant hardware, and the program can be stored in a computer readable storage medium. The program includes one or a combination of steps of the method embodiments when executed.

Claims

1. A method for aerial target detection based on transformer and jump path aggregation FPN, characterized in that, The method comprises the following steps: obtaining aerial image data to be detected; inputting the aerial image data to be detected into a pre-trained aerial small target detection network model for target detection to obtain a target detection result of the aerial image data to be detected; wherein the aerial small target detection network model is a deep aggregation residual transformation network combined with a deformable convolution and a transformer module for feature extraction, a skip-path aggregation feature pyramid network for feature fusion, and a dual-path decoupling detection head for target detection; the aerial small target detection network model comprises a backbone network, a neck network, a region extraction network, a region of interest pooling layer, and a detection head; the backbone network comprises one starting processing layer and four stage processing layers, the starting processing layer is composed of three 3x3 convolutions and one max pooling layer; each stage processing layer is composed of a plurality of aggregation residual transformation bottleneck layers combined with deformable convolution and transformer, and four feature maps of aerial image data are output through the four stage processing layers; the neck network comprises the skip-path aggregation feature pyramid network, which is used to obtain the four feature maps from the backbone network, and input the four feature maps into the skip-path aggregation feature pyramid network after reducing the channels through a convolution module; the skip-path comprises a horizontal skip connection and a down-sampling skip connection, and the output of the skip-path aggregation feature pyramid network is input into the region extraction network and the region of interest pooling layer; the region extraction network is used to generate candidate regions; the region of interest pooling layer is used to perform region of interest feature map extraction on the feature maps output by the skip-path aggregation feature pyramid network using the candidate regions generated by the region extraction network, and use region of interest pooling to make the output sizes of all feature maps the same, and input the feature maps into the detection head; the detection head is a cascaded RCNN multi-stage structure, each stage uses a dual-path decoupling detection head, for the feature maps output by the region of interest pooling layer, one path uses four convolution layers to process and perform boundary box regression to obtain the target detection frame of the aerial image data, and the other path uses two fully connected layers to process and perform classification to obtain the target classification result of the aerial image data.

2. The method of claim 1, wherein, The pooling mode in the region of interest pooling layer is modulated deformable region of interest pooling.

3. An aerial target detection device based on transformer and jump path aggregation FPN, characterized in that, The method comprises the following steps: an acquisition module, configured to acquire aerial image data to be detected; a detection module, configured to input the aerial image data to be detected into a pre-trained aerial small target detection network model for target detection to obtain a target detection result of the aerial image data to be detected; wherein the aerial small target detection network model is a deep aggregation residual transformation network combined with a deformable convolution and a transformer module for feature extraction, a skip-path aggregation feature pyramid network for feature fusion, and a dual-path decoupling detection head for target detection; the aerial small target detection network model comprises a backbone network, a neck network, a region extraction network, a region of interest pooling layer, and a detection head; the backbone network comprises one starting processing layer and four stage processing layers, the starting processing layer is composed of three 3x3 convolutions and one max pooling layer; each stage processing layer is composed of a plurality of aggregation residual transformation bottleneck layers combined with deformable convolution and transformer, and four feature maps of aerial image data are output through the four stage processing layers; The backbone network comprises a starting processing layer and four stage processing layers, the starting processing layer is composed of three 3x3 convolutions and a max pooling layer; each stage processing layer is composed of a plurality of aggregated residual transformation bottleneck levels combined with deformable convolution and transformer, and four layers of feature maps of aerial image data are output through the four stage processing layers; The neck network comprises the skip path aggregated feature pyramid network, which is used to acquire the four layers of feature maps from the backbone network, and input into the skip path aggregated feature pyramid network after reducing channels through a convolution module, the skip path comprises a horizontal skip connection and a down-sampling skip connection, and output of the skip path aggregated feature pyramid network is sent to the region extraction network and the region of interest pooling layer; The region extraction network is used to generate a candidate region; The region of interest pooling layer is used to perform region of interest feature map extraction on the feature map output by the skip path aggregated feature pyramid network by using the candidate region generated by the region extraction network, and all feature map output sizes are made same by using region of interest pooling, and the feature map is sent to the detection head; The detection head is a cascaded RCNN multi-stage structure, each stage uses a double-way decoupled detection head, for the feature map output by the region of interest pooling layer, one way uses four convolution layers to process and then performs boundary box regression to obtain a target detection box of aerial image data, and the other way uses two fully connected layers to process and then performs classification to obtain a target classification result of aerial image data.

4. The apparatus of claim 3, wherein, The pooling mode in the region of interest pooling layer is modulated deformable region of interest pooling.

5. An electronic device, comprising: Comprise: A memory, a processor and a computer program stored on the memory and executable on the processor, the processor executes the program to realize the aerial target detection method based on the transformer and the skip path aggregated FPN according to any one of claims 1-2.

6. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to realize the aerial target detection method based on the transformer and the skip path aggregated FPN according to any one of claims 1-2.

Images