Aerial photography small target detection method and system based on deep neural network

Abstract

The invention discloses an aerial photography small target detection method and system based on a deep neural network. The method comprises the following steps: acquiring aerial photography training set data; inputting the aerial photography training set data into a pre-constructed deep neural network for training, and generating an aerial photography small target detection model; obtaining a to-be-detected aerial photo; and inputting a to-be-detected aerial photo into the aerial photo small target detection model, and outputting a target classification result. According to the invention, through the aerial photography small target detection model, the target features in the aerial photography picture of the unmanned aerial vehicle are better extracted, and the detection precision of the small target is improved.

Description

Technical Field

[0001] The present invention relates to the field of big data intelligent analysis, and in particular to a method for detecting small aerial targets based on a deep neural network. Background Art

[0002] Camera-equipped drones or general-purpose drones have been rapidly deployed in a wide range of applications, including agriculture, aerial photography, fast delivery, and surveillance. In the field of object detection, one-stage models are more suitable for the needs of real-time detection. Most existing small object detection models do not perform well in real-world scenarios. There are some problems with small object detection in drone aerial photography scenarios: 1) The altitude of the drone is constantly changing during flight, and the scale of the target in the image will fluctuate dramatically; 2) When the drone flies quickly, the captured image is blurred, resulting in the loss of features of small targets; 3) The drone has a wider viewing angle, and the distribution of small objects in the image is prone to imbalance.

[0003] In the prior art, such as (Chinese patent document CN118736452A), although a variety of deep learning-based small target detection methods for aerial images have been proposed, there are still several shortcomings. First, the target detection network in the existing methods usually relies on the backbone network for multiple downsampling, which easily leads to the loss or weakening of small target features, thereby affecting the detection accuracy, especially in aerial images with complex backgrounds and large scale differences. In addition, the processing of features in the prior art often lacks an effective feature fusion mechanism, especially when processing multi-scale targets, it is difficult to effectively integrate feature information from different scales. Although traditional convolutional neural networks (CNNs) have been improved by increasing the network depth or introducing new convolution operations, these methods often lead to a significant increase in the amount of calculation and parameters, affecting the efficiency and real-time performance of the model. Summary of the invention

[0004] In view of the above technical deficiencies, the present invention provides an aerial photography small target detection method based on deep neural network. The aerial photography small target detection model extracts the features of the target in the image through the backbone network; the aerial photography image to be detected is input into the aerial photography small target detection model, and the target classification result is output.

[0005] The present invention focuses on how to improve the small target detection performance in the drone capture scene, where the small target refers to a target that is less than 80 pixels in a 256×256 image. The present invention provides a method for detecting weak small targets in aerial images based on feature reorganization and allocation, which significantly improves the detection accuracy and efficiency of small targets. Specifically, the present invention extracts multi-scale features through a downsampling backbone network, and refines redundant feature information in combination with a feature reorganization module to generate reorganized features carrying key redundant features. This method can effectively avoid the problem of small target feature loss caused by excessive downsampling in traditional methods. In addition, the combination of the feature reorganization module and the feature allocation module can effectively fuse features from different sources, ensuring the accurate fusion of multi-scale features, thereby improving the performance of the model in complex background and multi-scale target detection. The present invention ensures the improvement of detection accuracy by optimizing the feature fusion method, while reducing the amount of calculation, making the detection process more efficient and adapting to the needs of real-time detection.

[0006] The purpose of the present invention is achieved by at least one of the following technical solutions.

[0007] A method for detecting small aerial targets based on a deep neural network comprises the following steps: S1. Obtain aerial photography training set data; S2, input the aerial photography training set data into the pre-built deep neural network for training to generate an aerial photography small target detection model; S3, obtaining the aerial image to be detected; S4. Input the aerial image to be detected into the aerial small target detection model and output the target classification result.

[0008] Furthermore, the aerial photography small target detection model includes a backbone network, a fusion network and a prediction network; The input of the backbone network is the original aerial image. The backbone network includes a first dynamic area detection module and a convolutional layer connected in sequence. The convolutional layer of the backbone network outputs a low-level feature map to the fusion network. The fusion network includes a second dynamic region detection module, a dense region attention module and a transformer encoding module, and the fusion network generates a high-level feature map; The prediction network includes a transformer prediction head module. The high-level feature map is input into the prediction network. The transformer prediction head module generates detection results, including the location of different categories of objects, confidence, and classification results of target categories.

[0009] Further, the first dynamic area detection module and the second dynamic area detection module include a dynamic area-aware convolution module, a batch normalization module and an activation module connected in sequence; The dynamic area-aware convolution module extracts small target features from the input data through dynamic-aware convolution. The batch normalization module normalizes the input small target features by calculating the mean and variance of each channel, and outputs the normalized small target features. The activation module receives the normalized small target features, applies the activation function, performs nonlinear transformation, and outputs the enhanced small target features.

[0010] Furthermore, in the dynamic region-aware convolution module, the input data is convolved to obtain a feature map ; Dynamic-aware convolution includes standard convolution and local convolution; In the standard convolution stage, the feature map pass dimensional convolution kernel to extract coarse-grained features and generate guiding features ,pass Function Generation A guide mask; In the local convolution stage, the feature map is transformed into Divide into Shared Areas , and through the filter generation module of the dynamic area-aware convolution module Generate corresponding Shared area Convolutional filters ; Filter generation module The corresponding convolutional filters are automatically learned and generated based on the spatial features of each shared area. Specifically, The weight of each filter is dynamically adjusted according to the feature information in the region, so that each filter All of them can adapt to the specific semantic and spatial feature requirements of the area. This generation process ensures that the feature extraction of each area is highly targeted and adaptive; the feature maps in the shared area will share the same convolution filter for processing, thereby ensuring that features with similar semantics can receive consistent convolution operations in the same shared area; the role of the guided mask module is to enable the spatial dimension of the feature map to allocate appropriate convolution filters according to the characteristics of different shared areas, thereby improving the perception of small targets in the shared area; Shared Area The features inside the convolution filter are input Perform convolution operation; each convolution filter The convolution kernel will be based on the corresponding shared area The semantic information of is dynamically adjusted to ensure that the convolution operation can adapt to the feature requirements of different shared areas, so as to better extract the local features of small targets and generate local feature maps; Shared Areas The local feature map and Guided masks for dynamic convolution Operation, generate small target features.

[0011] Furthermore, in the filter generation module, the input feature map is transformed into Downsample to size , the feature map after downsampling After two convolutional layers, we get convolutional filters, where the first convolutional layer uses As an activation function, no activation function is used in the second convolutional layer.

[0012] Furthermore, the dense region attention module includes a channel attention module and a spatial attention module; In the dense area attention module, the channel attention module receives the input feature map and generates a channel attention map; the input feature map and the channel attention map are weighted by element-level multiplication operations, and the channel weighted feature map is output; The spatial attention module receives the channel weighted feature map and generates a spatial attention map; the channel weighted feature map and the spatial attention map are weighted through element-level multiplication operations to output a dense area feature map.

[0013] Further, the transformer encoding module includes a multi-head attention module, a feed-forward module and a first degradation layer; The multi-head attention module receives the input data of the transformer encoding module, the input and output of the multi-head attention module are residually connected and input to the feedforward module, the input and output of the feedforward module are residually connected and input to the first degradation layer, the first degradation layer outputs the output data of the transformer encoding module, and inputs it to the transformer prediction head module of the prediction network; The multi-head attention module includes a first normalization layer, a multi-head attention layer, and a second degradation layer connected sequentially; The feedforward module includes a second normalization layer and a plurality of stacked feedforward layers, the second normalization layer receives input data of the feedforward module, and the feedforward layer receives output data of the second normalization layer; The main function of the normalization layer is to normalize the features by calculating the mean and variance of each channel to enhance the stability during training. The normalization process can effectively avoid the problem of gradient vanishing or gradient exploding, accelerate the convergence of the model, and maintain the stability of network training. The degradation layer is used to avoid overfitting problems during network training.

[0014] Each feedforward layer includes linear transformation and nonlinear transformation; the feedforward layer first performs linear transformation on the input data through the fully connected layer, and then performs nonlinear transformation through the activation function to enhance the expressiveness of the features; the stacking of multiple feedforward layers can perform more complex feature transformations based on the input features, allowing the model to capture deeper feature information.

[0015] Furthermore, the backbone network includes a first dynamic area detection module and a The first convolutional layer, The first convolutional layer is a shallow convolutional layer, The first convolutional layer is the middle convolutional layer, The first convolution layer is a deep convolution layer; the shallow convolution layer, the middle convolution layer and the deep convolution layer output shallow features, middle features and deep features to the fusion network respectively; In the fusion network, the deep features are passed through the transformer encoding module and the second dynamic area detection module to obtain the first data; The first data is passed through the dense area attention module, and is concatenated with the intermediate layer features processed by the second dynamic area detection module and the dense area attention module. The obtained result is passed through the second convolution layer to obtain the second data; After the second data passes through the second dynamic area detection module and the dense area attention module, the second data is spliced ​​with the shallow features processed by the second dynamic area detection module and the dense area attention module to obtain third data; The third data passes through the second convolution layer and the dense area attention module, and is concatenated with the second data to obtain the fourth data; The fourth data passes through the second convolution layer and the dense area attention module, and is concatenated with the first data to obtain the fifth data; The third data, the fourth data and the fifth data are respectively input into the prediction network after passing through the second convolutional layer and the transformer encoding module.

[0016] Furthermore, in step S2, the hyperparameters of the pre-built deep neural network are set, and the hyperparameters include the learning rate and the number of training rounds ; The Adam optimizer is used to train the deep neural network with set hyperparameters.

[0017] A small target detection system for aerial photography based on a deep neural network, the system specifically comprising: The training data acquisition module is used to obtain the aerial photography training set data; The model training module is used to input the aerial photography training set data into the pre-built deep neural network for training and generate an aerial photography small target detection model; The module for acquiring the data to be detected is used to obtain the aerial pictures to be detected; The prediction module is used to input the aerial pictures to be detected into the aerial small target detection model and output the target classification results.

[0018] The present invention has the following beneficial effects and advantages: 1. The method of the present invention is directed to a method for detecting small aerial targets, and improves the ability of detecting small aerial targets by introducing a dynamic region detection module. The dynamic region detection module includes a dynamic region-aware convolution module, a batch normalization module, and an activation module. The dynamic region-aware convolution module flexibly adjusts the convolution operation so that the network can focus on the small target area in the image; the batch normalization module accelerates training and improves stability, helping the network overcome the gradient vanishing problem and optimize the learning effect; the activation block enhances the nonlinear feature response of the network so that it can more sensitively capture the details of small targets. Through the combination of these three, the network exhibits stronger feature extraction and target recognition capabilities when processing small targets.

[0019] 2. The method of the present invention improves the detection performance of small targets at the expense of a small detection speed through a transformer encoding module. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 Schematic diagram of the structure of the aerial photography small target detection model in an embodiment of the present invention.

[0021] Figure 2 It is a flow chart of a method for detecting small aerial targets based on a deep neural network in an embodiment of the present invention.

[0022] Figure 3 Schematic diagram of the structure of the transformer encoding module in an embodiment of the present invention.

[0023] Figure 4 Schematic diagram of the structure of the dense area attention module in an embodiment of the present invention.

[0024] Figure 5 Schematic diagram of the structure of the dynamic area detection module in the embodiment of the present invention.

[0025] Figure 6 This is a diagram of the implementation effect of an embodiment of the present invention on the VisDrone2021 dataset. DETAILED DESCRIPTION

[0026] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the specific implementation of the present invention is described in detail below with reference to the accompanying drawings and examples.

[0027] Example: Small objects are defined as objects with pixels smaller than goal.

[0028] A method for detecting small aerial targets based on deep neural networks, such as Figure 2 As shown, the following steps are included: S1. Obtain aerial photography training set data; In one embodiment, the aerial photography training set data is the VisDrone2021 dataset and the DOTA-v1.0 dataset; The VisDrone2021 dataset contains 10 categories: pedestrian, person, car, van, bus, truck, motorcycle, bicycle, awning tricycle and tricycle. There are 6471 training images, 548 validation sets and 3190 test sets.

[0029] The DOTA-v1.0 dataset contains 2,806 aerial images and a total of 188,282 instances. DOTA-V1.0 contains 15 categories: airplane (PL), baseball field (BD), bridge (BR), farmland (GTF), small vehicle (SV), large vehicle (LV), ship (SH), tennis court (TC), basketball court (BC), storage tank (ST), football field (SBF), roundabout (RA), port (HA), swimming pool (SP), helicopter (HC).

[0030] S2, input the aerial photography training set data into the pre-built deep neural network for training to generate an aerial photography small target detection model; like Figure 1 As shown, the aerial photography small target detection model includes a backbone network (which can be based on the backbone network of YOLOv7), a fusion network and a prediction network; The input of the backbone network is the original aerial image. The backbone network includes a first dynamic area detection module and a convolutional layer connected in sequence. The convolutional layer of the backbone network outputs a low-level feature map to the fusion network. The fusion network includes a second dynamic region detection module, a dense region attention module and a transformer encoding module, and the fusion network generates a high-level feature map; The prediction network includes a transformer prediction head module (see Alexey Dosovitskiy, et al. "AnImage is Worth 16x16 Words: Transformers for Image Recognition at Scale." InProceedings of the IEEE International Conference on Computer Vision (ICCV), 2021.), where the high-level feature map is input into the prediction network and the transformer prediction head module generates detection results, including the location of different categories of objects, confidence, and classification results of the target categories.

[0031] like Figure 5 As shown, the first dynamic area detection module and the second dynamic area detection module include a dynamic area-aware convolution module, a batch normalization module and an activation module connected in sequence; The dynamic area-aware convolution module extracts small target features from the input data through dynamic-aware convolution. The batch normalization module normalizes the input small target features by calculating the mean and variance of each channel, and outputs the normalized small target features. The activation module receives the normalized small target features, applies the activation function, performs nonlinear transformation, and outputs the enhanced small target features.

[0032] In one embodiment, the activation module uses function; like Figure 5 As shown in the dynamic area perception convolution module, the input data is convolved to obtain the feature map ; Dynamic-aware convolution includes standard convolution and local convolution; In one embodiment, during the standard convolution stage, the feature map pass The convolution kernel is used to extract coarse-grained features and generate guiding features ,pass Function Generation A guide mask; In the local convolution stage, the feature map is transformed into Divide into Shared Areas , and through the filter generation module of the dynamic area-aware convolution module Generate corresponding Shared area Convolutional filters ; Filter generation module The corresponding convolutional filters are automatically learned and generated based on the spatial features of each shared area. Specifically, The weight of each filter is dynamically adjusted according to the feature information in the region, so that each filter All of them can adapt to the specific semantic and spatial feature requirements of the area. This generation process ensures that the feature extraction of each area is highly targeted and adaptive; the feature maps in the shared area will share the same convolution filter for processing, thereby ensuring that features with similar semantics can receive consistent convolution operations in the same shared area; the role of the guided mask module is to enable the spatial dimension of the feature map to allocate appropriate convolution filters according to the characteristics of different shared areas, thereby improving the perception of small targets in the shared area; Shared Area The features inside the convolution filter are input Perform convolution operation; each convolution filter The convolution kernel will be based on the corresponding shared area The semantic information of is dynamically adjusted to ensure that the convolution operation can adapt to the feature requirements of different shared areas, so as to better extract the local features of small targets and generate local feature maps; Shared Areas The local feature map and Guided masks for dynamic convolution Operation (see Chen J, Wang X, Guo Z, et al. Dynamic region-aware convolution[C] / / Proceedings of the IEEE / CVFconference on computer vision and pattern recognition. 2021: 8064-8073.), generate small target features, and obtain feature maps .

[0033] In the filter generation module, the input feature map is transformed into Downsample to size , the feature map after downsampling After two convolutional layers, we get convolutional filters, where the first convolutional layer uses As an activation function, no activation function is used in the second convolutional layer.

[0034] like Figure 4 As shown, the dense region attention module includes a channel attention module and a spatial attention module; In the dense area attention module, the channel attention module receives the input feature map and generates a channel attention map; the input feature map and the channel attention map are weighted by element-level multiplication operations, and the channel weighted feature map is output; The spatial attention module receives the channel weighted feature map and generates a spatial attention map; the channel weighted feature map and the spatial attention map are weighted through element-level multiplication operations to output a dense area feature map.

[0035] In one embodiment, the channel attention module receives an input feature map of size ,in and are the height and width of the image, is the number of channels of the feature map; The channel attention module calculates the global maximum pooling map and average pooling map through maximum pooling and average pooling operations respectively; the global maximum pooling map and average pooling map are integrated and passed through a fully connected layer to generate a channel attention map with a size of , indicating the importance of each channel; The spatial attention module receives channel-weighted feature maps of size , a two-dimensional spatial attention map is generated through channel aggregation operation; in the spatial attention module, the maximum pooling map and the average pooling map are generated in the spatial dimension through the maximum pooling and average pooling operations respectively, and the maximum pooling map and the average pooling map are spliced ​​along the channel axis into a size of The spliced ​​feature map is The convolutional layer is processed to generate the final spatial attention map, which is The spatial attention map represents the degree of attention paid to each location in the image, especially the enhancement of small object regions.

[0036] like Figure 3 As shown, the transformer encoding module includes a multi-head attention module, a feedforward module and a first degradation layer; The multi-head attention module receives the input data of the transformer encoding module, the input and output of the multi-head attention module are residually connected and input to the feedforward module, the input and output of the feedforward module are residually connected and input to the first degradation layer, the first degradation layer outputs the output data of the transformer encoding module, and inputs it to the transformer prediction head module of the prediction network; The multi-head attention module consists of a first normalization layer, a multi-head attention layer (Vaswani A.Attention is all you need[J]. Advances in Neural Information Processing Systems, 2017.) and a second degradation layer connected sequentially; The feedforward module includes a second normalization layer and a plurality of stacked feedforward layers, the second normalization layer receives input data of the feedforward module, and the feedforward layer receives output data of the second normalization layer; The main function of the normalization layer is to normalize the features by calculating the mean and variance of each channel to enhance the stability during training. The normalization process can effectively avoid the problem of gradient vanishing or gradient exploding, accelerate the convergence of the model, and maintain the stability of network training. The degradation layer is used to avoid overfitting problems during network training.

[0037] Each feedforward layer includes linear transformation and nonlinear transformation; the feedforward layer first performs linear transformation on the input data through the fully connected layer, and then performs nonlinear transformation through the activation function to enhance the expressiveness of the features; the stacking of multiple feedforward layers can perform more complex feature transformations based on the input features, allowing the model to capture deeper feature information.

[0038] The backbone network includes a first dynamic area detection module and a The first convolutional layer, The first convolutional layer is a shallow convolutional layer, The first convolutional layer is the middle convolutional layer, The first convolution layer is a deep convolution layer; the shallow convolution layer, the middle convolution layer and the deep convolution layer output shallow features, middle features and deep features to the fusion network respectively; In one embodiment, Take 4, Take 2.

[0039] In another embodiment, Take 3, Take 1.

[0040] In the fusion network, the deep features are passed through the transformer encoding module and the second dynamic area detection module to obtain the first data; The first data is passed through the dense area attention module, and is concatenated with the intermediate layer features processed by the second dynamic area detection module and the dense area attention module. The obtained result is passed through the second convolution layer to obtain the second data; After the second data passes through the second dynamic area detection module and the dense area attention module, the second data is spliced ​​with the shallow features processed by the second dynamic area detection module and the dense area attention module to obtain third data; The third data passes through the second convolution layer and the dense area attention module, and is concatenated with the second data to obtain the fourth data; The fourth data passes through the second convolution layer and the dense area attention module, and is concatenated with the first data to obtain the fifth data; The third data, the fourth data and the fifth data are respectively input into the prediction network after passing through the second convolutional layer and the transformer encoding module.

[0041] Set the hyperparameters of a pre-built deep neural network, including the learning rate and the number of training rounds ; The Adam optimizer is used to train the deep neural network with set hyperparameters.

[0042] In one embodiment, the model is trained for 100 epochs and the initial learning rate is set to 1e-3. In the experiment, the evaluation indicators are precision, recall, and For DOTA-v1.0, the initial learning rate of Adam is .

[0043] S3, obtaining the aerial image to be detected; S4. Input the aerial image to be detected into the aerial small target detection model and output the target classification result.

[0044] In one embodiment, Figure 6 As shown in the figure, the model was trained for 100 batches and gradually converged after 50 batches. In the end, the accuracy of the result was 67.5%, the recall rate was 58.0%, and the result was The result is 60.5%. It is 39.22%.

[0045] In one embodiment, the aerial photography small target detection model of the present invention is compared with TPH-YOLOv5 and YOLOv5 series architectures. The comparison results are shown in Table 1. The Tiny-YOLOv7 model performs better than the other models, with an improvement of 0.92% in accuracy, 1.5% in recall, 1.87% in mAP0.5, and 1.6% in CNN. The score has increased by 1.07%. Compared with other YOLO series models, the Tiny-YOLOv7 model performs well on the VisDrone2021 dataset. In addition, this model was experimented on the dataset DOTA-v1.0. As shown in Table 2, the second row of Table 2 is a two-stage model, the third row of Table 2 is a one-stage model, and the aerial small target detection model of the present invention is in the last row. As a one-stage model, although Tiny-YOLOv7 did not achieve the optimal performance of the two-stage model, it showed good generalization ability in the experiment. For the one-stage model, Tiny-YOLOv7 achieved the best performance. The result is 80.9%, which is 0.9% higher than the GDW method.

[0046] Table 1 Implementation effect of the method of the present invention on the VisDrone2021 dataset

[0047] Table 2 Implementation effect of the method of the present invention on the DOTA-v1.0 dataset

[0048] Tables 1 and 2 show the experimental results of the model on the VisDrone2021 and DOTA-v1.0 datasets, respectively.

[0049] The preferred embodiments of the present application disclosed above are only used to help understand the present invention and its core ideas. For those skilled in the art, according to the ideas of the present invention, there will be changes in specific application scenarios and implementation operations, and this specification should not be interpreted as limiting the present invention. The present invention is only limited by the claims and their full scope and equivalents.

Claims

1. A method for detecting small aerial targets based on deep neural networks, characterized in that: The following steps are involved: S1. Obtain aerial photography training set data; S2. Input the aerial photography training set data into a pre-built deep neural network for training to generate an aerial photography small target detection model; the aerial photography small target detection model includes a backbone network, a fusion network and a prediction network; S3, obtaining the aerial image to be detected; S4. Input the aerial image to be detected into the aerial small target detection model and output the target classification result.

2. The method for detecting small aerial targets based on a deep neural network according to claim 1, characterized in that: In step S2, The input of the backbone network is the original aerial image, the backbone network includes a first dynamic area detection module and a convolution layer connected in sequence, and the convolution layer of the backbone network outputs a low-level feature map to the fusion network; The fusion network includes a second dynamic region detection module, a dense region attention module and a transformer encoding module, and the fusion network generates a high-level feature map; The prediction network includes a transformer prediction head module, the high-level feature map is input into the prediction network, and the transformer prediction head module generates detection results, including target locations, confidence levels, and classification results of target categories of different categories.

3. The method for detecting small aerial targets based on a deep neural network according to claim 2, characterized in that: The first dynamic area detection module and the second dynamic area detection module include a dynamic area-aware convolution module, a batch normalization module and an activation module connected in sequence; The dynamic region-aware convolution module extracts small target features from the input data through dynamic-aware convolution. The batch normalization module normalizes the input small target features by calculating the mean and variance of each channel, and outputs the normalized small target features. The activation module receives the normalized small target features, applies the activation function, performs nonlinear transformation, and outputs the enhanced small target features.

4. The method for detecting small aerial targets based on a deep neural network according to claim 3, characterized in that: In the dynamic region-aware convolution module, the input data is convolved to obtain a feature map ; Dynamic-aware convolution includes standard convolution and local convolution; In the standard convolution stage, the feature map pass dimensional convolution kernel to extract features and generate guided features ,pass Function Generation A guide mask; In the local convolution stage, the feature map is transformed into Divide into shared regions, and generates the corresponding Shared area convolutional filters; The convolution filter corresponding to the feature input in the shared area performs convolution operation to generate a local feature map; The local feature maps of the shared region and Guided masks for dynamic convolution Operation, generate small target features.

5. The method for detecting small aerial targets based on a deep neural network according to claim 4, characterized in that: In the filter generation module, the input feature map is transformed into Downsampling, feature map after downsampling After two convolutional layers, we get convolutional filters, where the first convolutional layer uses As an activation function, no activation function is used in the second convolutional layer.

6. The method for detecting small aerial targets based on a deep neural network according to claim 2, characterized in that: The dense region attention module includes a channel attention module and a spatial attention module; In the dense area attention module, the channel attention module receives the input feature map and generates a channel attention map; the input feature map and the channel attention map are weighted by element-level multiplication operations, and the channel weighted feature map is output; The spatial attention module receives the channel weighted feature map and generates a spatial attention map; the channel weighted feature map and the spatial attention map are weighted through element-level multiplication operations to output a dense area feature map.

7. The method for detecting small aerial targets based on a deep neural network according to claim 2, characterized in that: The transformer encoding module includes a multi-head attention module, a feed-forward module, and the first degradation layer; The multi-head attention module receives the input data of the transformer encoding module, the input and output of the multi-head attention module are residually connected and input to the feedforward module, the input and output of the feedforward module are residually connected and input to the first degradation layer, the first degradation layer outputs the output data of the transformer encoding module, and inputs it to the transformer prediction head module of the prediction network; The multi-head attention module includes a first normalization layer, a multi-head attention layer, and a second degradation layer connected sequentially; The feedforward module includes a second normalization layer and a plurality of stacked feedforward layers, the second normalization layer receives input data of the feedforward module, and the feedforward layer receives output data of the second normalization layer.

8. The method for detecting small aerial targets based on a deep neural network according to claim 1, characterized in that: In step S2, the hyperparameters of the pre-built deep neural network are set, including the learning rate and the number of training rounds ; The Adam optimizer is used to train the deep neural network with set hyperparameters.

9. The method for detecting small aerial targets based on a deep neural network according to any one of claims 1 to 8, characterized in that: The backbone network includes a first dynamic area detection module and a The first convolutional layer, The first convolutional layer is a shallow convolutional layer, The first convolutional layer is the middle convolutional layer, The first convolution layer is a deep convolution layer; the shallow convolution layer, the middle convolution layer and the deep convolution layer output shallow features, middle features and deep features to the fusion network respectively.

10. A small target detection system for aerial photography based on deep neural network, characterized in that: The system specifically comprises: The training data acquisition module is used to obtain the aerial photography training set data; The model training module is used to input the aerial photography training set data into the pre-built deep neural network for training and generate an aerial photography small target detection model; The module for acquiring the data to be detected is used to obtain the aerial pictures to be detected; The prediction module is used to input the aerial pictures to be detected into the aerial small target detection model and output the target classification results.

Images