CN115147745A - Small target detection method based on urban unmanned aerial vehicle image - Google Patents

Abstract

The invention discloses a small-scale target detection method based on an urban unmanned aerial vehicle image, and belongs to the technical field of computer vision technology and target detection. The invention carries out data enhancement; extracting features through a composite backbone network; combining a bidirectional characteristic pyramid network to perform multi-scale characteristic fusion; and according to the multi-scale characteristic information, combining a space pyramid attention mechanism to carry out target positioning and classification. The invention solves the problems of missing detection and false detection of small-scale targets in unmanned aerial vehicle target detection; improving the accuracy of model detection; and the detection performance of the small-scale target is ensured.

Description

A small target detection method based on urban UAV images

technical field

The invention belongs to the neighborhood of computer vision technology and target detection technology, and in particular relates to a small-scale target detection implementation method based on urban drone images.

Background technique

Camera-equipped drones have received a lot of attention in recent years. Drones equipped with embedded devices can analyze the captured data and give rise to a variety of new application scenarios, such as: drones can provide valuable insights to help farmers or ranchers optimize agricultural operations, monitor crop growth, maintain Herd safety and more. Use drones to extract aerial imagery instead of expensive cranes and helicopters. Drones can effectively deliver packages such as medical supplies, food or other goods to designated locations. Drones can provide real-time visualizations of security threats and emergencies by monitoring large areas. Drones are very useful in helping search for missing persons, fugitives or rescue survivors, as well as drop supplies in difficult terrain and harsh conditions. Therefore, it becomes very important to automatically understand the visual data collected from drones, which makes computer vision more and more closely related to drones. As two fundamental problems of computer vision, object detection and object tracking for UAV videos are also being extensively studied.

In UAV-supported urban aerial surveillance and visual object recognition scenarios, UAVs usually need to detect objects in near real-time by processing video images captured by cameras mounted on the UAV. The drone is able to follow the target and actively change its direction and detection area based on the visual feedback to optimize detection performance. Since these small UAVs are limited by their own size and power capabilities, it becomes infeasible to process massive video streams with high inference accuracy. In this context, machine learning, especially deep learning, is becoming more and more popular in many computer vision-based drone vision applications. Since traditional feature engineering is not always well suited for UAV aerial tracking in complex environments, deep neural networks, especially convolutional neural networks, are used in image classification and recognition to extract key features from captured images. Therefore, the use of deep learning technology for target detection in UAV aerial images has important research value and significance.

At present, the target detection methods based on deep learning can be mainly divided into two categories. One is the two-stage method based on the target candidate area. This method achieves the classification of the detection result by extracting the candidate area and performing deep learning on the corresponding area, so the accuracy of this method is high. The other is the one-stage method based on deep learning regression, which can directly calculate the category probability and position coordinate value of the object. After a single detection, the final detection result can be directly obtained, which greatly improves the detection speed. The rise of the deep convolutional network CNN in 2012 divided the target detection technology into two stages: the traditional target detection period and the detection period based on deep learning, and led the rapid development of the target detection field. Among them, compared with the two-stage series of candidate frame extraction and classification, YOLO unifies the target detection problem into a regression problem, and directly achieves target classification and target position regression through one detection; YOLOv2 continues to maintain processing speed. Train the target detector on the detection data set and the classification data set, use the data of the detection data set to learn the exact position of the object, and use the data of the classification data set to increase the amount of classification categories and improve the robustness of the model; YOLOv3 combines a single neural network It is applied to multiple positions and scales of the image, divides the image into different regions, predicts the bounding box and probability of each region, and regards the bounding box and classification probability with a higher score as the detection result; YOLOv4 combines a large number of research technologies, Combining them and making appropriate innovations achieves a perfect balance between speed and accuracy; in the training phase of the YOLOv5 algorithm, techniques such as Mosaic data enhancement, adaptive anchor box calculation, and adaptive image scaling are introduced, and the Focus structure and The CSP structure, the FPN+PAN structure is added to the Neck network, and the GIOU_Loss loss function and DIOU_nms are used in the Head output layer, which greatly improves the speed and accuracy. Although the YOLO series algorithm model improves the detection speed, the detection effect of small targets is not good by directly predicting the target position and classification through regression.

SUMMARY OF THE INVENTION

In order to overcome the above-mentioned deficiencies of the prior art, the present invention provides a small-scale target detection method based on urban UAV images. The model is improved to a detection method suitable for small targets in urban UAV images; thus, the problem of missed detection and false detection of small targets in UAV image detection is solved, and the detection accuracy of small targets is improved.

The technical terms of the present invention are first explained below.

IoU (Intersection over Union): IoU is the ratio of the intersection and union of the prior box and the predicted box. The value range of IoU is [0, 1], IoU=0 when there is no intersecting area between the real frame and the prior target frame, and IoU=1 when the two frames are completely coincident. The larger the IoU value, the more similar the shape and position of the ground-truth box and the prior target box. The metric value is negatively correlated with the similarity, that is, the smaller the metric value is, the greater the similarity is, and d _iou =1-IOU is obtained, that is, the distance metric between the prior frame and the prediction frame.

K-means clustering: It is an iterative solution clustering analysis algorithm. The steps in the present invention are: according to the height and width of the initially given bounding box, first randomly select 9 initial cluster centers, and then calculate the distance between the height and width of each initial given bounding box from the 9 cluster centers Happening. The sample points composed of the width and height of the bounding box are assigned to the cluster centers with the shortest distance. After completing the allocation of 9 sample points, the cluster center will be recalculated according to the existing sample points near the center point to obtain a new cluster center point. The above process is repeated continuously until the termination condition is satisfied, that is, no sample points are reassigned to different cluster centers or the cluster centers no longer change.

Non-Maxium Suppression (NMS): Search for local maxima and suppress maxima. According to the threshold, duplicate detection frames are filtered out by traversing, sorting, etc. It has been widely used in computer vision, such as edge detection, object detection, etc.

The technical scheme provided by the present invention is:

A small-scale target detection method based on urban UAV images, comprising the following steps:

1) Initialize the original image:

Obtain the storage path of the UAV image, the type of detection target, the position information of the detection target in the image, and obtain the target position and classification result of the original image according to the position and length and width of the target bounding box;

2) Image data enhancement:

By means of Mosaic data enhancement, the original images are spliced by random scaling, random cropping, and random arrangement, and then a new image containing the target bounding box is obtained;

3) Image feature extraction:

3-1) For different UAV image sizes, set a fixed initial anchor frame size in advance, then cluster the frame size, compare the obtained result with the real bounding box size, and calculate the difference between the two. The gap between the two, select the appropriate border size to reversely update the set initial border size, and adjust the border size adaptively according to the target size;

3-2) Using the composite backbone network superimposed by the double-layer CSPDarknet53 backbone network, the feature maps of the same scale are repeatedly sampled for many times to enhance the feature extraction capability of the backbone network for small targets. The processing flow is shown in the following formula:

表示在第l阶段主骨干网络的操作，

表示在第l阶段辅助骨干网络输出的特征图，

表示在第l阶段主骨干网络输出的特征图，

表示在第l-1阶段主骨干网络输出的特征图，G(·)表示辅助骨干网络和主骨干网络之间的1×1的卷积和归一化操作。Among them, it is assumed that each backbone network has L stages,

represents the operation of the backbone network in phase l,

represents the feature map output by the auxiliary backbone network in stage l,

represents the feature map output by the main backbone network in the lth stage,

represents the feature map output by the main backbone network at stage l-1, G( ) represents the 1×1 convolution and normalization operations between the auxiliary backbone network and the main backbone network.

4) Multi-scale feature fusion:

Use a bidirectional feature pyramid network with top-down and bottom-up bidirectional paths to fuse multi-scale image feature information bidirectionally multiple times;

5) Object classification and location prediction:

Using spatial pyramid attention mechanism to achieve object localization and classification.

Further, the bounding box of each object along with its classification name and confidence is drawn with different colors on the original image.

Compared with the prior art, the beneficial effects of the present invention are:

The invention provides a small target detection method based on unmanned aerial vehicle images. On the basis of the original yolov5, multiple composite backbone networks are used to strengthen the feature extraction capability of small-scale targets and improve the feature extraction capability of the target detection network; The feature pyramid network is used to strengthen the fusion ability of the network to multi-scale features, which is used for multi-scale learning of small-scale target features in the training process; the spatial pyramid attention mechanism is used to strengthen the network model's detection ability of small-scale targets, and further strengthen The detection performance of the model for small objects.

Description of drawings

FIG. 1 is a flowchart of a small target detection method based on an urban UAV image according to an embodiment of the present invention.

FIG. 2 is a schematic diagram showing the relationship of each stage of the target detection method according to the embodiment of the present invention.

FIG. 3 is a flowchart of a bidirectional feature pyramid network provided by an embodiment of the present invention.

Detailed ways

In order to make it easier to understand the purpose, technical solutions and embodiments of the present invention, the present invention will be further described below with reference to the accompanying drawings and embodiments. This embodiment is only used to explain the present invention, and does not limit the present invention.

The invention is based on a small-scale target detection method based on urban UAV images. By improving the target detection method YOLOv5, a detection model is constructed based on a composite backbone network, a bidirectional feature pyramid network and a spatial pyramid attention mechanism, and a multi-scale feature layer is obtained according to the model. The prediction results are used to realize the classification and location prediction of the target. The method specifically includes 6 execution steps, as shown in Figure 1, including:

1) Initialize the original image: The data set used in the embodiment of the present invention is 8629 pictures taken by drones, and each picture has a corresponding standard file, which contains the type of the target in the picture, and the bounding box. The coordinates of the upper left corner and other information. Obtain the storage path of the UAV image, the type of detection target, and the position information of the detection target in the image through the standard file, and obtain the target position and classification result in the original image according to the upper left corner coordinates and length and width of the marked target bounding box. , shuffle the storage order of image data, and randomly generate training set and test set;

2) Data enhancement of the image: Through the Mosaic data enhancement method, the four original images are spliced by random scaling, random cropping, and random arrangement, and then a new image containing the target bounding box is obtained. The new images are input into the model for learning, which improves the detection effect of small targets. Four pictures are randomly read from the data image each time, and the four pictures are flipped left and right, the size is scaled, and the brightness, saturation, and hue are changed. After the operation is completed, the processed original pictures are processed according to The first picture is placed in the upper left, the second picture is placed in the lower left, the third picture is placed in the lower right, and the fourth picture is placed in the upper right four directions. After that, the combination of pictures and the combination of boxes are carried out. After the placement of the four pictures is completed, the fixed area of the four pictures is intercepted by means of a matrix, and then they are spliced together to form a new picture, a new picture It contains a series of content such as boxes. For a new picture obtained after the stitching is completed, when the frame of the picture or the picture itself exceeds the edge dividing line between the two pictures, the part of the frame or the picture that exceeds the dividing line is processed, and edge processing is performed.

3) Perform feature extraction on the target in the image, as shown in Figure 2:

3-1) First, according to different UAV image sizes, the initial anchor frame sizes are set in advance for three types of targets, large, medium and small, on three scales of 80×80, 40×40 and 20×20. Then, on the basis of this value, the K-means clustering algorithm is used to cluster the width and height of all target bounding boxes in the training set, and as a result, 9 width and height combinations are obtained, that is, 9 target anchor boxes are obtained. In this process, in order to avoid the appearance of errors, the IoU distance is used to measure the sample distance of the K-means algorithm. Among them, the value range of IoU is [0, 1], IoU=0 when there is no intersecting area between the real bounding box and the target anchor box, and IoU=1 when the two boxes are completely coincident. The larger the IoU value, the more similar the shape and position of the ground-truth bounding box and the target anchor box. In order to make the metric value negatively correlated with the similarity, that is, the smaller the metric value, the greater the similarity, take a negative value for the IoU value and add 1 to get: d _iou = 1-IOU, that is, use the K-means algorithm to cluster anchor boxes The distance metric to use for width and height. By comparing the target anchor box results obtained by the K-means clustering algorithm with the real bounding box size, calculating the gap between the two to update the network in reverse, iterating the network parameters, and continuously adjusting the anchor box adaptively according to the target size size.

3-2) Multi-sampling the feature maps of the same scale by superimposing the two-layer CSPDarknet53 backbone network to enhance the feature extraction capability of the backbone network for small targets.

Composite backbone network based on CSPDarknet53: Based on the CSPDarknet53 backbone network in yolov5, combined with the composite backbone network, the proposed network structure to enhance feature extraction of small objects. In the embodiment of the present invention, an image with a size of 640×640 is used, and the feature image obtained after convolution processing in the auxiliary backbone network is input into the main backbone network of the same scale, and then superimposed and fused with the feature image of the main backbone network. , and then input it as input into the backbone network model. The same operations are taken in the backbone network stages of different scales to strengthen the feature extraction of the backbone network for objects of different scales. Specific steps include:

之后，进行相似的操作，将特征图

通过3×3的卷积、归一化、激活函数操作以及残差特征学习模块之后，得到80×80×128的特征图

再将特征图

进行3×3的卷积操作和残差学习之后，得到40×40×256的特征图

最后，将特征图

通过3×3的卷积操作得到的20×20×512的特征图，通过空间金字塔池化模块，将5×5，9×9，13×13的最大池化操作得到的结果和模块输入结果相融合，再进行残差特征学习后得到辅助骨干网络的最终输出的20×20×512特征图

First, the original image with a size of 640×640×3 is sliced by Focus to obtain a feature map of 320×320×12, and then through a 3×3 convolution operation, the obtained 320×320×32 feature map is defined as x0. In the auxiliary backbone network, after the feature map x ⁰ is operated by 3×3 convolution, normalization, and activation function, the expansion channel obtains 160×160×64 image features, and then the residual feature learning module will use 3 1×1 convolution, normalization, and activation function operation feature maps and the feature maps input to the module are merged to obtain a 160×160×64 feature map containing more feature information.

After that, a similar operation is performed to convert the feature map

After 3×3 convolution, normalization, activation function operation and residual feature learning module, a feature map of 80×80×128 is obtained

Then the feature map

After 3×3 convolution operation and residual learning, a 40×40×256 feature map is obtained

Finally, the feature map

The feature map of 20 × 20 × 512 obtained by the 3 × 3 convolution operation, through the spatial pyramid pooling module, the results obtained by the 5 × 5, 9 × 9, and 13 × 13 max pooling operations and the module input results After fusion, the residual feature learning is performed to obtain a 20×20×512 feature map of the final output of the auxiliary backbone network.

通过1×1的卷积、归一化操作和经过切片处理后得到的特征图x⁰相叠加得到320×320×32的特征图定义为

之后，将特征图

通过3×3的卷积、归一化、激活函数操作和残差特征学习模块之后的得到160×160×64特征图

将特征图

通过卷积和归一化操作后和特征图

叠加之后的结果，再进行相似的操作，通过卷积归一化操作以及残差特征学习模块之后，得到80×80×128的特征图

之后，同样将经过特征图

通过卷积归一化操作后和特征图

叠加，将其作为输入特征输入到主骨干网络中，再次通过卷积和残差学习得到40×40×256的特征图

最后，将特征图

通过卷积归一化后和特征图

叠加后，输入到网络中，进行卷积操作、空间金字塔池化和残差学习之后，得到主骨干网络的最终输出的20×20×512特征图

其中通过主骨干网络得到的特征图

和

将作为骨干网络在80×80、40×40和20×20三个尺度上提取到的特征输出到模型的下一阶段进行处理。In the main backbone network, the feature map obtained through the auxiliary backbone network is firstly

The feature map of 320 × 320 × 32 is obtained by superimposing the feature map x ⁰ obtained after 1 × 1 convolution, normalization and slicing, which is defined as

After that, the feature map

The 160×160×64 feature map is obtained after 3×3 convolution, normalization, activation function operation and residual feature learning module

feature map

After operation and feature map through convolution and normalization

After superimposing the results, similar operations are performed. After the convolution normalization operation and the residual feature learning module, a feature map of 80×80×128 is obtained.

After that, the same will go through the feature map

After normalization operation and feature map by convolution

Superimpose, input it into the main backbone network as an input feature, and obtain a feature map of 40×40×256 through convolution and residual learning again

Finally, the feature map

After normalization and feature map by convolution

After superposition, input into the network, after convolution operation, spatial pyramid pooling and residual learning, the 20×20×512 feature map of the final output of the main backbone network is obtained

The feature map obtained through the main backbone network

and

The features extracted as the backbone network at three scales of 80×80, 40×40 and 20×20 are output to the next stage of the model for processing.

To sum up, the processing flow in the feature extraction stage of the present invention is shown in the following formula:

表示在第l阶段主骨干网络的操作，

表示在第l阶段辅助骨干网络输出的特征图，

表示在第l阶段主骨干网络输出的特征图，

表示在第l-1阶段主骨干网络输出的特征图，G(·)表示辅助骨干网络和主骨干网络之间的1×1的卷积和归一化操作。Among them, it is assumed that each backbone network has L stages,

represents the operation of the backbone network in phase l,

represents the feature map output by the auxiliary backbone network in stage l,

represents the feature map output by the main backbone network in the lth stage,

represents the feature map output by the main backbone network at stage l-1, G( ) represents the 1×1 convolution and normalization operations between the auxiliary backbone network and the main backbone network.

4) Multi-scale feature fusion: In the multi-scale feature fusion stage, by using a bidirectional feature pyramid network with top-down and bottom-up bidirectional paths, multi-scale feature information is fused in multiple directions in both directions to strengthen the feature fusion capability for small targets. . The feature map obtained by feature extraction directly fuses the feature nodes input to the same scale through skip connection, and fuses more features. At the same time, the top-down and bottom-up bidirectional paths are called four times to see the bidirectional feature pyramid network, which achieves higher-level feature fusion.

Bidirectional Feature Pyramid Network: It is an enhanced feature multi-scale fusion structure in the present invention, as shown in FIG. 3 . The PANet structure of yolov5 can fuse multi-scale target feature information, but the fusion process of small-scale targets and large-scale targets is easy to cause missing features, and it takes a lot of time in the training process, which affects the model's detection of small-scale targets capability and efficiency. To this end, the present invention removes the intermediate feature map nodes of 80 × 80 and 20 × 20 scales by introducing a bidirectional feature pyramid network, and only retains the intermediate feature map nodes of 40 × 40 scale, thereby generating a A simple bidirectional network; in addition, adding skip connections between input feature map nodes and output feature map nodes of the same scale can integrate more features without increasing computational cost; finally, it will be able to achieve top-down and The network structure of the bottom-up bidirectional path is regarded as a feature network layer. The feature network layer is repeatedly called four times in the model, and different scales are fused in two directions on three scales of 80×80, 40×40, and 20×20. The target feature information can achieve better fusion of higher-level small targets at multiple scales. Specific steps include:

经过1×1的卷积处理后得到的20×20×256特征图定义为

之后将

通过上采样得到40×40×256特征图，和特征提取到的40×40×256的特征图

叠加后，通过双向特征金字塔网络双向多尺度融合得到40×40×512的特征图，之后通过残差特征学习得到40×40×256的特征图，再经过1×1的卷积处理后得到40×40×128的特征图

将

继续上采样得到80×80×128的特征图，将其和提取到的80×80×128特征图

叠加融合，通过双向特征金字塔网络得到80×80×256的特征图，再经过残差特征学习得到80×80×128的特征图

在这一过程中，网络融合了80×80、40×40和20×20三个尺度上的图像特征，实现了自顶向下的特征融合。First, extract the 20×20×512 feature map from the feature extraction stage

The 20×20×256 feature map obtained after 1×1 convolution processing is defined as

will later

The 40×40×256 feature map is obtained by upsampling, and the 40×40×256 feature map extracted from the feature

After superposition, the feature map of 40×40×512 is obtained by bidirectional multi-scale fusion of the bidirectional feature pyramid network, and then the feature map of 40×40×256 is obtained by residual feature learning, and then 40×40 is obtained after 1×1 convolution processing. ×40×128 feature map

Will

Continue upsampling to obtain a feature map of 80×80×128, and sum it with the extracted 80×80×128 feature map

Stacking and fusion, the feature map of 80×80×256 is obtained through the bidirectional feature pyramid network, and the feature map of 80×80×128 is obtained through residual feature learning.

In this process, the network fuses image features at three scales of 80×80, 40×40 and 20×20 to achieve top-down feature fusion.

经过3×3的卷积处理来进行下采样之后得到40×40×128的特征图，之后再和40×40×128的特征图

叠加，接着通过双向多尺度特征融合得到40×40×256的特征图，再进行残差特征学习后得到40×40×256的特征图

经过相似的操作，通过3×3卷积来进行下采样得到20×20×256的特征图，之后再和20×20×256的特征图

叠加，通过双向特征融合得到20×20×512的特征图，进行残差特征学习后得到20×20×512的特征图

在这一阶段，网络则实现了自底向上的融合三个尺度上的图像特征。其中，特征图

和

作为模型在80×80、40×40和20×20三个尺度上的特征结果输出到模型的下一阶段进行结果预测。after,

After 3×3 convolution processing for downsampling, the feature map of 40×40×128 is obtained, and then the feature map of 40×40×128 is added.

Overlay, and then obtain a feature map of 40 × 40 × 256 through bidirectional multi-scale feature fusion, and then perform residual feature learning to obtain a feature map of 40 × 40 × 256

After a similar operation, the 20×20×256 feature map is obtained by downsampling through 3×3 convolution, and then the 20×20×256 feature map is added.

Overlay, a feature map of 20 × 20 × 512 is obtained through bidirectional feature fusion, and a feature map of 20 × 20 × 512 is obtained after residual feature learning

At this stage, the network achieves bottom-up fusion of image features at three scales. Among them, the feature map

and

As the feature results of the model at three scales of 80×80, 40×40 and 20×20, it is output to the next stage of the model for result prediction.

5) Perform target classification and location prediction:

The invention utilizes the spatial pyramid attention mechanism to realize target positioning and classification. Spatial pyramid attention mechanism: is the attention mechanism used in the present invention to strengthen the prediction result, as shown in Figure 2. The spatial pyramid attention mechanism is used on the three model output branches, and the input feature maps generate attention maps through the spatial pyramid structure of adaptive average pooling at three scales of 1×1, 2×2, and 4×4. After that, the generated attention map is combined with a fully connected layer and a sigmoid activation layer to generate the attention weight in the corresponding feature map, and the small objects in the original image are more accurately marked. Specifically include the following steps:

和

通过在80×80、40×40和20×20三个尺度上自适应平均池化的空间金字塔结构生成注意力图。其中的1×1的自适应平均池化层的目的是获取特征图中的类别关键信息，2×2池化层用来保存图像中的次重要的关键特征信息，4×4平均池化可以有效的获取特征图中的关键位置信息。5-1) Use the attention mechanism to focus the model's attention on the small target part in the image, extract the key information in the image while ignoring the interference of irrelevant information such as background, and improve the small target localization and classification performance. First, the feature map after feature extraction and multi-scale fusion

and

The attention maps are generated through a spatial pyramid structure with adaptive average pooling at three scales of 80×80, 40×40 and 20×20. The purpose of the 1×1 adaptive average pooling layer is to obtain the key category information in the feature map, the 2×2 pooling layer is used to save the less important key feature information in the image, and the 4×4 average pooling can Effectively obtain key location information in the feature map.

5-2) Pass the generated attention map through a weight composed of a fully connected layer and a sigmoid activation layer to generate the attention weights in the corresponding feature map. Thus, through the attention weight of the attention output, the small objects in the original image are more accurately marked.

The present invention can use the position and score of the predicted bounding box for non-maximum suppression, select the predicted bounding box with the best effect as the target bounding box, and avoid multiple predicted bounding boxes for one target. Sort the predicted bounding box of the image target obtained by the prediction model according to its prediction score from high to low; calculate the predicted bounding box with the largest predicted score, and the IoU value of all other predicted boxes; when the IoU value exceeds the set threshold , remove the predicted bounding box, and then continue to find the highest IoU in the remaining predicted bounding boxes, and then remove the predicted bounding box whose IoU exceeds the threshold, until the predicted bounding box with almost no overlap will be retained until each target. Only one prediction bounding box remains. Then, sort and filter the target types in the predicted bounding box area. The same target type takes the maximum value of its confidence, and selects the type with the highest ranking in the ranking of the maximum confidence value of different target categories, and takes it as Object The kind of object that predicts the bounding box region.

6) Visualization: Draw the predicted bounding box of each object, its category category and confidence level on the original image with different colors.

It should be noted that the purpose of publishing the embodiments is to help further understanding of the present invention, but those skilled in the art can understand that various replacements and modifications are possible without departing from the spirit and scope of the present invention and the appended claims of. Therefore, the present invention should not be limited to the contents disclosed in the embodiments, and the scope of protection of the present invention shall be subject to the scope defined by the claims.

Claims (7)

1. a small-scale target detection method based on urban UAV image, is characterized in that, comprises the steps: 1) Initialize the original image: Obtain the storage path of the UAV image, the type of detection target, the position information of the detection target in the image, and obtain the target position and classification result of the original image according to the position and length and width of the target bounding box; 2) Image data enhancement: By means of Mosaic data enhancement, the original images are spliced by random scaling, random cropping, and random arrangement, and then a new image containing the target bounding box is obtained; 3) Image feature extraction: 3-1) Set a fixed initial anchor box size, then cluster the target bounding box size, compare the obtained target anchor box results with the real bounding box size, and select the appropriate one by calculating the gap between the two. The size of the target anchor frame is reversely updated to set the initial anchor frame size, and the anchor frame size is adaptively adjusted according to the target size; 3-2) Using the composite backbone network with the double-layer CSPDarknet53 backbone network superimposed, the feature maps of the same scale are repeatedly sampled for many times, and the processing flow is shown in the following formula:

Among them, it is assumed that each backbone network has L stages,

represents the operation of the backbone network in phase l,

represents the feature map output by the auxiliary backbone network in stage l,

represents the feature map output by the main backbone network in the lth stage,

represents the feature map output by the main backbone network in stage l-1, G( ) represents the 1×1 convolution and normalization operations between the auxiliary backbone network and the main backbone network; 4) Multi-scale feature fusion: Use a bidirectional feature pyramid network with top-down and bottom-up bidirectional paths to fuse multi-scale image feature information bidirectionally multiple times; 5) Object classification and location prediction: Using spatial pyramid attention mechanism to achieve object localization and classification. 2. the small-scale target detection method based on urban UAV image according to claim 1, is characterized in that, after setting initial anchor frame size in step 3-1), use K-means on the basis of this value The clustering algorithm clusters the width and height of all target bounding boxes in the training set, and the result is the target anchor box. The IoU distance is used to measure the sample distance of the K-means algorithm. The value range of IoU is [0, 1]. IoU=0 when there is no intersection area between the frame and the target anchor frame, and IoU=1 when the two frames are completely coincident. 3. the small-scale target detection method based on urban UAV image according to claim 1, is characterized in that, step 5) specifically comprises: 5-1) Extract the feature map of multi-scale fusion, and generate the attention map in the spatial pyramid structure of adaptive average pooling; 5-2) Use the generated attention map to generate the attention weight in the corresponding feature map through a weight composed of a fully connected layer and a sigmoid activation layer. The target is accurately marked. 4 . The small-scale target detection method based on urban UAV images according to claim 3 , wherein the key information in the multi-scale fusion feature map is extracted while ignoring the interference of background irrelevant information. 5 . 5. the small-scale target detection method based on urban UAV image according to claim 1, is characterized in that, step 5-2) utilizes the position and the score of predicted bounding box to carry out non-maximum suppression, namely calculates the maximum predicted score. The predicted bounding box, and the IoU value of all other predicted boxes; when the IoU value exceeds the set threshold, remove the predicted bounding box, and then continue to find the highest IoU in the remaining predicted bounding boxes, and then remove the one with the IoU exceeding the threshold The prediction bounding box of , until the end will keep almost no overlapping boxes, so that only one prediction bounding box is left for each target. 6. The small-scale target detection method based on urban UAV images according to claim 5, characterized in that, sorting and screening the target types in the screened predicted bounding box area, and the same target type is taken as its confidence level. The maximum value is selected, and the category with the highest ranking in the order of the maximum confidence value of different target categories is selected as the target category of the target prediction bounding box area. 7. The small-scale target detection method based on urban UAV images according to claim 1, wherein the predicted bounding box of each target, as well as its category type and confidence level, are drawn with different colors on the original image. .

Images