An unmanned aerial vehicle aerial image target detection method, device and storage medium

Abstract

The application discloses an unmanned aerial vehicle aerial image target detection method and device and a storage medium, wherein the method comprises the following steps: a backbone network extracts features of an unmanned aerial vehicle aerial image to obtain preliminary features; a cross-stage local hollow space pyramid pooling module is used to obtain features of different receptive fields, and channel aggregation is performed to obtain deep features; the obtained deep features are input into a path aggregation network module for feature fusion, and a coordinate attention multi-scale stacking module is used to extract more rich spatial position information to obtain feature representations of different scales; and the feature representations of different scales are subjected to target classification and boundary box regression through a convolution module and a nonlinear activation function. According to the application, the coordinate attention multi-scale stacking module is used to inject coordinate information into channels while realizing multi-scale feature extraction, so that more detailed information is retained to help the detection of small targets in the image. The application can be widely applied to the field of target detection.

Description

Technical Field

[0001] This invention relates to the field of target detection, and in particular to a method, apparatus and storage medium for target detection in aerial images taken by unmanned aerial vehicles. Background Technology

[0002] With the rapid development of computer hardware and artificial intelligence algorithms, object detection, as an important component of computer vision, can accurately locate objects in images through object classification and bounding box regression, and has been widely used in daily life. In recent years, drone aerial photography technology has become increasingly mature, providing a new direction for object detection. By performing object detection on drone aerial images, we can quickly locate victims in fires or earthquakes, accelerating rescue efforts; we can also survey vehicles on roads to identify congested sections and manually clear them.

[0003] Despite the numerous applications of object detection in UAV aerial imagery, there are relatively few algorithms specifically designed for it. This is primarily due to a lack of datasets, the small size or uneven scale distribution of targets within the image, and issues such as occlusion and overlap. Therefore, UAV aerial image detection networks need to maximize their sensitivity to feature information, introduce attention mechanisms, and expand the receptive field for multi-scale prediction. One-stage object detection algorithms, similar to the YOLO series, perform well on common datasets like VOC and MSCOCO, but they lack attention mechanisms. Some improved algorithms use attention mechanisms but lack coordinate attention, easily losing global positional information and making small targets prone to being missed. Furthermore, pooling operations can also cause information loss. For example, while the ASPP network achieves a larger receptive field without changing the feature map resolution, dilated convolutions lead to information loss, reducing detection accuracy. Summary of the Invention

[0004] In order to at least partially solve one of the technical problems existing in the prior art, the present invention aims to provide a method, apparatus and storage medium for target detection in drone aerial images.

[0005] The technical solution adopted in this invention is:

[0006] A target detection method for drone aerial images includes the following steps:

[0007] Acquire drone aerial images, perform data augmentation and preprocessing on the drone aerial images, and set prior bounding boxes;

[0008] A backbone network composed of multi-scale stacked coordinate attention modules is used to extract features from UAV aerial images to obtain preliminary features.

[0009] Based on the preliminary features, the cross-stage local void space pyramid pooling module is used to obtain features of different receptive fields, and channel aggregation is performed to obtain deep features.

[0010] The obtained deep features are input into the path aggregation network module for feature fusion, and the coordinate attention multi-scale stacking module is used to extract richer spatial location information to obtain feature representations at different scales.

[0011] Feature representations at different scales are used for target classification and bounding box regression through convolutional modules and nonlinear activation functions.

[0012] Furthermore, data augmentation is performed on the drone aerial images, including:

[0013] Data augmentation processing of UAV aerial images is performed using at least one of the three methods: Mosaic method, Mixup method, or random flipping method.

[0014] In the Mosaic method, multiple images are randomly cropped and then stitched together into a single image; in the Mixup method, two sample and label data are added proportionally to generate new sample and label data.

[0015] Data preprocessing included normalizing ground truth bounding boxes and resizing the input image to 640×640×3. Additionally, nine different prior box sizes were used to achieve multi-scale feature map bounding box regression.

[0016] Furthermore, the backbone network includes a series of coordinate attention multi-scale stacked modules and downsampling transition modules to obtain feature representations with location information;

[0017] The downsampling transition module includes two 1×1 convolutional layers, one 3×3 convolutional layer, three batch normalization layers, and three SiLU activation function layers.

[0018] The coordinate attention multi-scale stacked module includes three 1×1 convolutional layers, four 3×3 convolutional layers, seven batch normalization layers, eight SiLU activation layers, and one coordinate attention layer. The coordinate attention layer includes a global average pooling layer in the X direction, a global average pooling layer in the Y direction, three 1×1 convolutional layers, a batch normalization layer, and three Sigmoid activation layers.

[0019] Furthermore, the calculation process expression for the coordinate attention layer is as follows:

[0020]

[0021]

[0022] f=σ(F1(Z h Z w ]))

[0023] g h =σ(F h (f h ))

[0024] g w =σ(F w (f w ))

[0025]

[0026] Where, x c It is an input feature with c channels, y c These are the output features, and W and H are the width and height of the input features. and It is the output of global average pooling of the input features along the horizontal and vertical directions, where i and j are used to determine the position of the feature points; σ represents the Sigmoid activation function, F1, F2... h F w Let f represent a 1×1 convolution transformation function, where f is an intermediate feature. h and f w Let f be divided into two feature representations along the horizontal and vertical dimensions, and g h and g w The output of the two-dimensional feature representation is processed by a 1×1 convolution transformation function and an activation function.

[0027] Furthermore, the method of obtaining features of different receptive fields using the cross-stage local void spatial pyramid pooling module includes:

[0028] The cross-stage local void space pyramid pooling module divides the initial input features into two parts according to channels. One part passes through a 1×1 convolutional layer to retain more detailed information, while the other part passes through a void space pyramid pooling layer to obtain different receptive fields.

[0029] The cross-stage local void space pyramid pooling module includes three 1×1 convolutional layers, one 3×3 convolutional layer, four batch normalization layers, four SiLU activation function layers, and a void space pyramid pooling layer.

[0030] The void space pyramid pooling layer includes three 1×1 convolutional layers, three 3×3 convolutional layers, one pooling layer, one batch normalization layer, one ReLU activation layer, and one upsampling layer.

[0031] The expression for calculating the size of the dilated convolution feature map is as follows:

[0032] n = k + (k-1)*(d-1)

[0033]

[0034] Where d is the introduced hyperparameter, (d-1) is the number of spaces in the input feature map; k is the original convolution kernel size, n is the convolution kernel size after adding spaces, i is the size of the input feature map, p is the padding value, s is the stride, and o is the size of the final output feature map.

[0035] Furthermore, the path aggregation network module includes four 1×1 convolutional layers, four coordinate attention multi-scale stacking modules, two downsampling transition modules, and two upsampling layers;

[0036] The entire path aggregation network is implemented by upsampling deep features and continuously fusing them with shallow features. The fused shallow feature map is then continuously fused with deep features through a coordinate attention multi-scale stacking module and a downsampling transition module, thereby obtaining three feature representations with different scales and rich semantic information.

[0037] Furthermore, the calculation process expression of the path aggregation network module is as follows:

[0038] P′4=F ca ([F1(P4),F up (F1(P5))])

[0039] P″3=F ca ([F1(P3),F up (F1(P′4))]

[0040] P″4=F ca ([F dp (P″3),P′4])

[0041] P″5=F ca ([F dp (P″4),F1(P5)])

[0042] Wherein, P3 and P4 are feature maps of different scales extracted by the backbone network, P5 is a deep feature map after passing through the cross-stage local void spatial pyramid pooling module, P′4 is an intermediate feature map obtained during the upsampling fusion process, and F ca For coordinate attention multi-scale stacked module transformation function, F up and F dp F1 is the transformation function for the upsampling and downsampling transition modules, P″3, P″4 and P″5 are the transformation functions for the 1×1 convolution, and P″3, P″4 and P″5 are the output feature representations of the three different scales after final fusion.

[0043] Furthermore, the step of performing target classification and bounding box regression by using feature representations of different scales through convolutional modules and nonlinear activation functions includes:

[0044] Bounding boxes are generated on the feature map, target classification and bounding box regression are performed using a loss function, and the highest-scoring predicted box is selected by non-maximum suppression, thereby achieving accurate target detection.

[0045] The bounding box regression calculation process is expressed as follows:

[0046] b x =σ(t) x )×2-0.5+c x

[0047] b y =σ(t) y )×2-0.5+c y

[0048] b w =p w (σ(t w )×2) 2

[0049] b h =p h (σ(t h )×2) 2

[0050] Among them, c x and c y t represents the coordinates of the top-left corner of the grid, used to determine the grid position; x t y t w and t h p is the offset relative to the grid. w and p h To separate the width and height of the bounding box relative to the grid, b x b y b w and b h Let be the coordinates of the predicted point and the width and height of the predicted bounding box, and σ be the Sigmoid activation function.

[0051] Another technical solution adopted in this invention is:

[0052] A target detection device for drone aerial images, comprising:

[0053] At least one processor;

[0054] At least one memory for storing at least one program;

[0055] When the at least one program is executed by the at least one processor, the at least one processor implements the method described above.

[0056] Another technical solution adopted in this invention is:

[0057] A computer-readable storage medium storing a processor-executable program, which, when executed by a processor, is used to perform the method described above.

[0058] The beneficial effects of this invention are as follows: By employing a coordinate attention multi-scale stacking module, this invention injects coordinate information into the channels while extracting features at multiple scales, alleviating the problem of positional information loss caused by continuous convolution and preserving more detailed information to aid in the detection of small targets in images. Furthermore, to obtain a larger and more accurate receptive field without increasing computational load, this invention combines a cross-stage local network structure and a holed spatial pyramid pooling network structure. This reduces computational load while allowing adjustment of the number of holes in the inserted feature map to obtain feature representations for different receptive fields, making target detection results in UAV aerial images simpler and more accurate. Attached Figure Description

[0059] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following description is provided with accompanying drawings of the relevant technical solutions in the embodiments of the present invention or the prior art. It should be understood that the accompanying drawings described below are only for the purpose of clearly illustrating some embodiments of the technical solutions of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0060] Figure 1 This is a coordinate attention multi-scale stacked network diagram in an embodiment of the present invention;

[0061] Figure 2 This is a diagram of the coordinate attention module in an embodiment of the present invention;

[0062] Figure 3 This is a diagram of the coordinate attention multi-scale stacking module in an embodiment of the present invention;

[0063] Figure 4 It is a typical hollow space pyramid pooling module diagram;

[0064] Figure 5 This is the cross-stage local void space pyramid pooling module in this embodiment of the invention;

[0065] Figure 6 This is a flowchart illustrating the steps of a target detection method for UAV aerial images based on a coordinate attention multi-scale stacked network, as described in an embodiment of the present invention. Detailed Implementation

[0066] The embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention. The step numbers in the following embodiments are set only for ease of explanation, and there is no limitation on the order between the steps. The execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0067] In the description of this invention, it should be understood that the orientation descriptions, such as up, down, front, back, left, right, etc., are based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention.

[0068] In the description of this invention, "several" means one or more, "more than" means two or more, "greater than," "less than," and "exceeding" are understood to exclude the stated number, while "above," "below," and "within" are understood to include the stated number. The use of "first" and "second" in the description is merely for distinguishing technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features, or implicitly indicating the order of the indicated technical features.

[0069] In the description of this invention, unless otherwise explicitly defined, terms such as "set up," "install," and "connect" should be interpreted broadly, and those skilled in the art can reasonably determine the specific meaning of the above terms in this invention in conjunction with the specific content of the technical solution.

[0070] like Figure 6 As shown, this embodiment provides a target detection method for UAV aerial images based on a coordinate attention multi-scale stacked network. This method uses a coordinate attention mechanism to construct a stacked module and a feature fusion module, which are used for feature extraction in the backbone network and feature fusion between low-level and deep features with different semantic information, respectively. Furthermore, by combining CSP and ASPP, a new CSP_ASPP module is designed, which can increase the receptive field while reducing computational load. Moreover, the demultiplexing operation of the CSP structure can retain more detailed information, which is beneficial to improving the target detection accuracy of UAV aerial images. The method specifically includes the following steps:

[0071] Step S1: The acquired drone aerial images are augmented and preprocessed, and the size of the prior bounding box is preset.

[0072] In this embodiment, the input drone aerial images undergo Mosaic data augmentation, Mixup data augmentation, and random flipping. The Mosaic method randomly crops four images and then stitches them together into a single image. Mixup adds two sample and label data proportionally to generate new sample and label data. Data preprocessing includes normalizing ground truth bounding boxes and adjusting the input image size to 640×640×3. Furthermore, nine different sizes of prior boxes are set to achieve multi-scale feature map bounding box regression.

[0073] Step S2: The backbone network, composed of coordinate attention multi-scale stacked modules, performs preliminary feature extraction on the image.

[0074] The backbone network, composed of multi-scale stacked coordinate attention modules, performs preliminary feature extraction on the image, such as... Figure 1 As shown, the backbone network contains a series of coordinate attention multi-scale stacked modules and downsampling transition modules to obtain preliminary feature representations with location information. The downsampling transition module includes two 1×1 convolutional layers, one 3×3 convolutional layer, three batch normalization layers, and three SiLU activation function layers. See [link to documentation]. Figure 3 The coordinate attention multi-scale stacked module consists of three 1×1 convolutional layers, four 3×3 convolutional layers, seven batch normalization layers, eight SiLU activation layers, and one coordinate attention layer. The coordinate attention layer is as follows: Figure 2 As shown, it includes a global average pooling layer in the X direction, a global average pooling layer in the Y direction, three 1×1 convolutional layers, a batch normalization layer, and three sigmoid activation layers. The calculation process is expressed as follows:

[0075]

[0076]

[0077] f=σ(F1([Z h Z w ])) (3)

[0078] g h =σ(F h (f h (4)

[0079] g w =σ(F w (f w (5)

[0080]

[0081] Among them, X c It is an input feature with c channels, Y cThese are the output features, and W and H are the width and height of the input features. and This is the output of global average pooling of the input features along the horizontal and vertical directions, where i and j determine the location of the feature points. σ represents the Sigmoid activation function, F1, F2... h F w Let f represent a 1×1 convolution transformation function, where f is an intermediate feature. h and f w Let f be divided into two feature representations along the horizontal and vertical dimensions, and g h and g w The output of the two-dimensional feature representation is processed by a 1×1 convolution transformation function and an activation function.

[0082] Step S3: Use the cross-stage local void spatial pyramid pooling module to obtain features of different receptive fields and perform channel aggregation to obtain deep feature representations.

[0083] See Figure 5 The cross-stage local dilated spatial pyramid pooling module divides the input features into two parts by channel. One part passes through a 1×1 convolutional layer to retain more detailed information, while the other part passes through a dilated spatial pyramid pooling layer to obtain different receptive fields. The cross-stage local dilated spatial pyramid pooling module includes three 1×1 convolutional layers, one 3×3 convolutional layer, four batch normalization layers, four SiLU activation function layers, and a dilated spatial pyramid pooling layer. The dilated spatial pyramid pooling layer itself contains three 1×1 convolutional layers, three 3×3 convolutional layers, one pooling layer, one batch normalization layer, one ReLU activation layer, and one upsampling layer. Figure 4 This is a traditional diagram of a hollow space pyramid pooling module.

[0084] The expression for calculating the size of the dilated convolution feature map is as follows:

[0085] n=k+(k-1)*(d-1) (7)

[0086]

[0087] Where d is the introduced hyperparameter, and (d-1) is the number of spaces in the input feature map. k is the original convolution kernel size, n is the convolution kernel size after adding spaces, i is the size of the input feature map, p is the padding value, s is the stride, and o is the size of the final output feature map.

[0088] Step S4: Input the deep features obtained in step S3 into the path aggregation network module for feature fusion, and use the coordinate attention multi-scale stacking module to extract richer spatial location information to obtain feature representations at different scales.

[0089] The path aggregation network module consists of four 1×1 convolutional layers, four coordinate attention multi-scale stacked modules, two downsampling transition modules, and two upsampling layers. See [link to relevant documentation]. Figure 3 The structures of the coordinate attention multi-scale stacking module and the downsampling transition module are the same as those described in step S2. The entire path aggregation network is implemented by upsampling deep features and continuously fusing them with shallow features. Then, the fused shallow feature map is continuously fused with deep features through the coordinate attention multi-scale stacking module and the downsampling transition module, thereby obtaining three feature representations with different scales and rich semantic information. The specific calculation process is expressed as follows:

[0090] P′4=F ca ([F1(P4),F up (F1(P5))]) (9)

[0091] P″3=F ca ([F1(P3),F up (F1(P′4))]) (10)

[0092] P″4=F ca ([F dp (P″3),P′4]) (11)

[0093] P″5=F ca ([F dp (P″4),F1(P5)]) (12)

[0094] Wherein, P3 and P4 are feature maps of different scales extracted by the backbone network, P5 is a deep feature map after passing through the cross-stage local void spatial pyramid pooling module, P′4 is an intermediate feature map obtained during the upsampling fusion process, and F ca For coordinate attention multi-scale stacked module transformation function, F up and F dp F1 is the transformation function for the upsampling and downsampling transition modules, P″3, P″4 and P″5 are the transformation functions for the 1×1 convolution, and P″3, P″4 and P″5 are the output feature representations of the three different scales after final fusion.

[0095] Step S5: Use convolutional modules and SiLU nonlinear activation functions to perform target classification and bounding box regression on feature representations at different scales.

[0096] The feature representations obtained in step S4 at three different scales are decoded and predicted using a convolutional module and a nonlinear activation function. This involves generating bounding boxes on the feature map, performing target classification and bounding box regression using a loss function, and selecting the highest-scoring predicted box using non-maximum suppression, thus achieving accurate target detection. The bounding box regression calculation process is expressed as follows:

[0097] b x =σ(t) x )×2-0.5+c x (13)

[0098] b y =σ(t) y )×2-0.5+c y (14)

[0099] b w =p w (σ(t w )×2) 2 (15)

[0100] b h =p h (σ(t h )×2) 2 (16)

[0101] Among them, c x and c y t represents the coordinates of the top-left corner of the grid, used to determine the grid position. x t y t w and t h p is the offset relative to the grid. w and p h To separate the width and height of the bounding box relative to the grid, b x b y b w and b h Let be the coordinates of the predicted point and the width and height of the predicted bounding box, and σ be the Sigmoid activation function.

[0102] As an optional implementation, this embodiment uses a weighted sum of classification loss, localization loss, and confidence loss as the loss function during the training phase. The Adam optimizer is used to update network parameters, and a multinomial descent learning rate strategy is employed to update the learning rate. Training is iterated for 300 epochs, with testing performed on a validation set each time, and the model weights with the smallest loss function value on the validation set are retained. During the testing phase, the preprocessed test dataset is input into the optimal model for testing, and the detection results are output.

[0103] This embodiment also provides a target detection device for drone aerial images, including:

[0104] At least one processor;

[0105] At least one memory for storing at least one program;

[0106] When the at least one program is executed by the at least one processor, the at least one processor implements Figure 6 The method shown.

[0107] This embodiment of the UAV aerial image target detection device can execute the UAV aerial image target detection method provided in the method embodiment of the present invention, and can execute any combination of the implementation steps of the method embodiment, and has the corresponding functions and beneficial effects of the method.

[0108] This application also discloses a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium and execute the computer instructions, causing the computer device to perform... Figure 6 The method shown.

[0109] This embodiment also provides a storage medium storing instructions or programs that can execute the UAV aerial image target detection method provided in the method embodiment of the present invention. When the instructions or programs are run, any combination of implementation steps of the method embodiment can be executed, and the method has the corresponding functions and beneficial effects.

[0110] In some alternative embodiments, the functions / operations mentioned in the block diagrams may not occur in the order shown in the operation diagrams. For example, depending on the functions / operations involved, two consecutively shown blocks may actually be executed substantially simultaneously, or the blocks may sometimes be executed in reverse order. Furthermore, the embodiments presented and described in the flowcharts of this invention are provided by way of example to provide a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logic flows presented herein. Alternative embodiments are contemplated in which the order of various operations is altered and sub-operations described as part of a larger operation are executed independently.

[0111] Furthermore, although the invention has been described in the context of functional modules, it should be understood that, unless otherwise stated, one or more of the described functions and / or features may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in a separate physical device or software module. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding the invention. Rather, given the properties, functions, and internal relationships of the various functional modules in the apparatus disclosed herein, the actual implementation of the module will be understood within the scope of conventional skill of an engineer. Therefore, those skilled in the art can implement the invention as set forth in the claims using ordinary techniques without excessive experimentation. It is also understood that the specific concepts disclosed are merely illustrative and not intended to limit the scope of the invention, which is determined by the full scope of the appended claims and their equivalents.

[0112] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0113] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0114] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0115] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0116] In the foregoing description of this specification, references to terms such as "one embodiment," "another embodiment," or "some embodiments" indicate that a specific feature, structure, material, or characteristic described in connection with an embodiment or example is included in at least one embodiment or example of the present invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0117] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

[0118] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the above embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A method for detecting a target in a UAV aerial image, characterized in that, The method comprises the following steps: obtaining an unmanned aerial vehicle aerial image, performing data enhancement and preprocessing on the unmanned aerial vehicle aerial image, and setting a prior box; extracting features of the unmanned aerial vehicle aerial image by a backbone network composed of a coordinate attention multi-scale stacking module to obtain preliminary features; obtaining features of different receptive fields by using a cross-stage local dilated spatial pyramid pooling module according to the preliminary features, and performing channel aggregation to obtain deep features; inputting the obtained deep features into a path aggregation network module for feature fusion, and extracting more rich spatial position information by using the coordinate attention multi-scale stacking module to obtain feature representations of different scales; performing target classification and bounding box regression on the feature representations of different scales through a convolution module and a nonlinear activation function; the cross-stage local dilated spatial pyramid pooling module includes three 1×1 convolution layers, one 3×3 convolution layer, four batch normalization layers, four SiLU activation function layers and a dilated spatial pyramid pooling layer; the dilated spatial pyramid pooling layer includes three 1×1 convolution layers, three 3×3 convolution layers, one pooling layer, one batch normalization layer, one Relu activation layer and one up-sampling layer; the expression of the calculation process of the dilated convolution feature map size is as follows: performing data enhancement on the unmanned aerial vehicle aerial image, comprising: at least one of the following three methods: Mosaic method, Mixup method or random flip is used to perform data enhancement processing on the unmanned aerial vehicle aerial image; wherein, is an introduced hyper-parameter, then is the number of empty spaces of the input feature map; is the original convolution kernel size, is the convolution kernel size after adding empty spaces, is the size of the input feature map, is the padding value, is the step size, is the size of the final output feature map. 2.The UAV aerial image target detection method of claim 1, wherein, in the Mosaic method, a plurality of pictures are randomly cropped and then spliced into one picture; in the Mixup method, new samples and label data are generated by adding two samples and label data in proportion. the backbone network includes a series of coordinate attention multi-scale stacking modules and down-sampling transition modules, so as to obtain feature representations with position information; the down-sampling transition module includes two 1×1 convolution layers, one 3×3 convolution layer, three batch normalization layers and three SiLU activation function layers; 3.The UAV aerial image target detection method of claim 1, wherein, the coordinate attention multi-scale stacking module includes three 1×1 convolution layers, four 3×3 convolution layers, seven batch normalization layers, eight SiLU activation layers and a coordinate attention layer; the coordinate attention layer includes one global average pooling layer in the X direction, one global average pooling layer in the Y direction, three 1×1 convolution layers, one batch normalization layer and three Sigmoid activation layers. the calculation process expression of the coordinate attention layer is as follows: ​ 4.The UAV aerial image target detection method of claim 3, wherein, ​ wherein, is an input feature having channels, is an output feature, and are width and height of the input feature, and are outputs of global average pooling of the input feature along horizontal and vertical directions, and are used to determine the position of the feature point; denotes an activation function, , , denotes a 1x1 convolutional transformation function, is an intermediate feature, and are split into two feature representations along horizontal and vertical dimensions, and are outputs of the two-dimensional feature representations after the 1x1 convolutional transformation function and the activation function.

5. The unmanned aerial vehicle aerial image target detection method according to claim 1, characterized in that, The path aggregation network module includes four 1 1 convolutional layer, four coordinate attention multi-scale stacking modules, two down-sampling transition modules, and two up-sampling layers; The implementation process of the whole path aggregation network is to up-sample deep features and constantly fuse them with shallow features, and then constantly fuse the fused shallow feature maps with deep features through a coordinate attention multi-scale stacking module and a down-sampling transition module, so as to obtain three kinds of feature representations with different scales and rich semantic information. 6.The UAV aerial image target detection method of claim 5, wherein, The calculation process expression of the path aggregation network module is as follows: wherein, are feature maps of different scales extracted from the backbone network respectively, are deep feature maps that have passed through the cross-stage local empty spatial pyramid pooling module, are intermediate feature maps obtained in the up-sampling fusion process, are coordinate attention multi-scale stacking module transformation functions, and are transformation functions of the up-sampling and down-sampling transition modules respectively, is a transformation function of a 1x1 convolution, , and are three different scale output feature representations after final fusion.

7. The unmanned aerial vehicle aerial image target detection method according to claim 1, wherein, The target classification and the bounding box regression of the feature representations with different scales through the convolution module and the nonlinear activation function include: Generating a bounding box on the feature map, performing target classification and bounding box regression through a loss function, and screening the highest scoring prediction box through non-maximum suppression, so as to realize accurate detection of the target; The bounding box regression calculation process expression is as follows: where, and is the top-left corner coordinate of the grid, which is used to determine the grid position; , , and is the offset of the relative grid, and is the width and height of the bounding box relative to the grid, respectively, , , and is the predicted point coordinate and the predicted bounding box width and height, is the activation function.

8. An unmanned aerial vehicle aerial image target detection device, characterized in that, including: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements the method of any one of claims 1-7.

9. A computer readable storage medium having stored therein a program which is executable by a processor, characterized in that, The program executable by the processor, when executed by the processor, is used to perform the method of any one of claims 1-7.

Images