An aerial image target detection system and method for point labeling

Abstract

The application belongs to the technical field of computer vision, and discloses an aerial image target detection system and method for point labeling, which comprises an image preprocessing module, a feature extraction module, a multiscale attention expansion module and a detection module; the image preprocessing module is used for covering the aerial image with multi-category template pictures, and performing random flipping, random scaling and random rotation processing to obtain a preprocessed image; the feature extraction module is used for constructing a feature extraction network, and performing feature extraction on the preprocessed image by using the feature extraction network to obtain an initial feature map; the multiscale attention expansion module is used for deep processing of the initial feature map to obtain a multiscale attention feature map; and the detection module is used for predicting target classification scores, dividing candidate points according to the classification scores as indexes, and predicting target coordinates; the application can effectively improve the accuracy of aerial image target detection for point labeling.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of computer vision, and particularly relates to an aerial image target detection system and method for point labeling. BACKGROUND

[0002] With the rapid development of deep convolutional neural networks, deep learning research has become a hot spot. Unlike traditional methods, deep learning algorithms do not require manual feature extraction, but automatically extract image features through convolution operations. Deep network architecture can capture more comprehensive and detailed features, so many researchers have applied deep learning algorithms to the field of aerial image detection. The targets in aerial images are numerous and have various orientations, and the cost of directional box labeling is high. Point labeling greatly reduces the labeling cost, but since point labeling does not provide the size and direction information of the target, it affects the detection performance of the model. SUMMARY

[0003] To solve the problems existing in the prior art, the present application provides an aerial image target detection system and method for point labeling, aiming to realize high-precision positioning function by combining ultra-wideband (UWB) positioning technology with smart wearable devices.

[0004] To achieve the above-mentioned purpose, the present application provides the following solutions:

[0005] An aerial image target detection system for point labeling, comprising an image preprocessing module, a feature extraction module, a multi-scale attention expansion module and a detection module.

[0006] The image preprocessing module is used for data enhancement processing of the aerial image to obtain a preprocessed image.

[0007] The feature extraction module is used for constructing a feature extraction network and extracting features of the preprocessed image by using the feature extraction network to obtain an initial feature map.

[0008] The multi-scale attention expansion module is used for deep processing of the initial feature map to obtain a multi-scale attention feature map.

[0009] The detection module is used for predicting target classification scores by using the classification branch of the network detection head, and dividing and detecting the multi-scale attention feature map by using the classification scores as indicators, and predicting target coordinates by using the regression branch of the network detection head.

[0010] Preferably, the feature extraction module comprises a network construction unit and a feature extraction unit.

[0011] The network construction unit is used for constructing a backbone network by using ResNet50.

[0012] The feature extraction unit is configured to extract features of the preprocessed image by using the backbone network to obtain an initial feature map.

[0013] Preferably, the backbone network comprises a C1 layer, a C2 layer, a C3 layer, a C4 layer and a C5 layer.

[0014] The C1 layer is composed of 1 7x7 convolution, 1 BN layer, 1 ReLU activation function and 1 max pooling layer; the C2 layer is composed of 6 1x1 convolutions, 3 3x3 convolutions, 3 BN layers and 3 ReLU activation functions; the C3 layer is composed of 8 1x1 convolutions, 4 3x3 convolutions, 4 BN layers and 4 ReLU activation functions; the C4 layer is composed of 12 1x1 convolutions, 6 3x3 convolutions, 6 BN layers and 6 ReLU activation functions; and the C5 layer is composed of 6 1x1 convolutions, 3 3x3 convolutions, 3 BN layers and 3 ReLU activation functions.

[0015] Preferably, the multi-scale attention expansion module first undergoes 32-group grouping processing, and then undergoes four branches, including a receptive field expansion branch, an X-channel attention branch, a Y-channel attention branch and a global attention branch.

[0016] The receptive field expansion branch is composed of 1 3x3 expansion convolution with an expansion rate of 2, 1 global average pooling layer and 1 Softmax activation function, and simultaneously transmits information to the X-channel attention branch and the Y-channel attention branch.

[0017] The X-channel attention branch is composed of 1 X-channel pooling layer.

[0018] The Y-channel attention branch is composed of 1 Y-channel pooling layer.

[0019] The global attention branch is composed of 1 global average pooling layer, 1 1x1 convolution and a dimension expansion layer.

[0020] The feature map after grouping is spliced after undergoing the X-channel pooling layer and the Y-channel pooling layer, and is restored back to the two branches before splicing by 1 1x1 convolution, is multiplied by the feature map after grouping by a Sigmod activation function, is multiplied by the feature map after grouping by a Sigmod activation function after being multiplied by the information transmitted by the receptive field expansion branch through the group normalization layer, the global average pooling layer and the Softmax activation function, and is multiplied by the feature map after the expansion convolution of the receptive field expansion branch, and is added to the information transmitted by the receptive field expansion branch, is multiplied by the feature map after grouping by a Sigmod activation function, to obtain an attention feature map; and the global attention branch is added to the attention feature map to obtain a multi-scale attention expansion feature map.

[0021] Preferably, the detection module comprises a detection head construction unit, a classification unit, a candidate point matching unit and a regression unit.

[0022] The detection head construction unit is configured to construct the network detection head, wherein the network detection head comprises a classification branch and a regression branch.

[0023] The classification unit is configured to detect the multi-scale attention expansion feature map by using the classification branch to obtain the predicted target classification score.

[0024] The candidate point matching unit is configured to match candidate points for each real target by using the classification score as an index.

[0025] The regression unit is configured to detect the multi-scale attention expansion feature map by using the regression branch to obtain the predicted target coordinates.

[0026] Preferably, the classification branch is composed of 4 3*3 convolutions, and the regression branch is composed of 4 3*3 convolutions.

[0027] Preferably, the candidate point matching unit comprises a generation module, a screening module, a calculation module and a filtering module.

[0028] The generation module is configured to generate a loss matrix for each labeled real point according to the classification score obtained by the classification branch, wherein the value of the loss matrix is the classification loss, i.e., 1-classification score, and select 25 points around the anchor point step length and anchor box size as the initial candidate points.

[0029] The screening module is configured to adopt a dynamic estimation method for the 25 candidate points, first generate a score matrix, wherein the value of the score matrix is the classification score predicted in the classification branch, and then select the top K coarse-grained candidate points with the highest scores in the score matrix for each real target, sum the classification scores and take the integer part to further select Q fine-grained candidate points.

[0030] The calculation module is configured to calculate the candidate point matching matrix of the real point according to the screened Q candidate points.

[0031] The filtering module is configured to filter the situation that a single candidate point is labeled as a candidate point by multiple real points at the same time, and select the candidate point with the smallest loss value in the loss matrix of the multiple real points as the candidate point of the real point.

[0032] The application also provides an aerial image target detection method for point labeling, which is realized by the system.

[0033] The aerial image is subjected to multi-class template picture overlay operation, and is subjected to random flipping, random scaling and random rotation processing to obtain a preprocessed image.

[0034] construct a feature extraction network, and perform feature extraction on the preprocessed image by using the feature extraction network to obtain an initial feature map;

[0035] construct a multi-scale attention expansion network, including four multi-scale attention expansion modules composed of dilated convolution with expansion rates of 2, 4, 6 and 8 connected in series, and perform deep processing on the initial feature map to obtain a multi-scale attention expansion feature map;

[0036] construct a network detection head, and perform detection on the multi-scale attention expansion feature map by using a classification branch of the network detection head to obtain a predicted target classification score; perform matching of candidate points by using a candidate point matching unit; and perform detection on the multi-scale attention expansion feature map by using a regression branch to obtain a predicted target coordinate.

[0037] Compared with the prior art, the beneficial effects of the present application are:

[0038] The present application designs a multi-scale attention expansion module on the basis of extracting features by using a ResNet50 model, which makes up for the deficiency that point labeling does not provide target size information and cannot use a multi-scale detection feature pyramid network, and also provides a larger receptive field of the model and fine feature representation with attention. In addition, a candidate point distribution strategy based on classification score as an evaluation index is adopted, and K coarse-grained candidate points and Q fine-grained candidate points of a real target are dynamically designed, which solves the problem that point labeling does not provide target size information and cannot use a candidate point distribution strategy based on intersection over union as an evaluation index, and effectively improves the performance of point labeling in aerial image target detection. BRIEF DESCRIPTION OF DRAWINGS

[0039] In order to more clearly illustrate the technical solutions of the present application, the following briefly introduces the drawings needed in the embodiments. Obviously, the drawings described in the following embodiments are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0040] Figure 1 It is a structure schematic diagram of an aerial image target detection system for point labeling according to an embodiment of the present application.

[0041] Figure 2 It is a backbone network schematic diagram according to an embodiment of the present application.

[0042] Figure 3 It is a multi-scale attention expansion module schematic diagram according to an embodiment of the present application.

[0043] Figure 4 It is a coarse-grained and fine-grained candidate point screening schematic diagram according to an embodiment of the present application.

[0044] Figure 5 This is a schematic diagram illustrating the matching matrix calculation according to an embodiment of the present invention;

[0045] Figure 6 This is a schematic diagram of shared candidate points for filtering according to an embodiment of the present invention. Detailed Implementation

[0046] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0048] Example 1

[0049] In this embodiment, as Figure 1 As shown, an aerial image target detection system for point annotation includes: an image preprocessing module, a feature extraction module, a multi-scale attention expansion module, and a detection module.

[0050] The image preprocessing module is used to perform data augmentation on aerial images to obtain preprocessed images.

[0051] The image preprocessing module's workflow includes: cropping a template image for each category from the dataset to be detected; during training, overlaying multiple template images of different categories onto the aerial image and performing random flipping, scaling, and rotation to obtain several preprocessed images with different orientations and sizes. Using the category template images as pseudo-labels to train the model to learn target size information, while simultaneously increasing the diversity of data representation, helps the system learn more robust feature representations.

[0052] The feature extraction module is used to construct a feature extraction network and use the feature extraction network to extract features from the preprocessed image to obtain an initial feature map.

[0053] The feature extraction module includes a network construction unit and a feature extraction unit; the network construction unit uses ResNet50 as the backbone network; the feature extraction unit uses the backbone network to extract features from the preprocessed image to obtain an initial feature map.

[0054] like Figure 2As shown, the backbone network includes layers C1, C2, C3, C4, and C5. Layer C1 consists of one 7×7 convolution, one batch normalization (BN) layer, one ReLU activation function, and one max pooling layer. The input to layer C1 is a preprocessed image of 3×1024×1024 pixels. After passing through one 7×7 convolution, one BN layer, one ReLU activation function, and one max pooling layer, the output is a 64×256×256 first feature map. Layer C2 consists of six 1×1 convolutions, three 3×3 convolutions, three BN layers, and three ReLU activation functions. The input to layer C2 is the 64×256×256 first feature map, and the output is a 256×256×256 second feature map. Layer C3 consists of eight 1×1 convolutions, four 3×3 convolutions, and four batch normalization (BN) layers. The C3 layer consists of 12 1×1 convolutions, 6 3×3 convolutions, 6 BN layers, and 6 ReLU activation functions. The input of the C3 layer is a second feature map of 256×256×256, and the output is a third feature map of 512×128×128. The C4 layer consists of 12 1×1 convolutions, 6 3×3 convolutions, 6 BN layers, and 6 ReLU activation functions. The input of the C4 layer is a third feature map of 512×128×128, and the output is a fourth feature map of 1024×64×64. The C5 layer consists of 6 1×1 convolutions, 3 3×3 convolutions, 3 BN layers, and 3 ReLU activation functions. The input of the C5 layer is a fourth feature map of 1024×64×64, and the output is an initial feature map of 2048×32×32.

[0055] The multi-scale attention expansion module is used to extract fine feature representations from the initial feature map to obtain a multi-scale attention expansion feature map.

[0056] like Figure 3 As shown, the multi-scale attention expansion module processes the input feature map in groups, including: receptive field expansion branch, X-channel attention branch, Y-channel attention branch and global attention branch.

[0057] The multi-scale attention expansion module takes an initial feature map of 2048×32×32 as input before grouping the input feature map, divides it into 32 groups, and outputs a feature map of 64×32×32 for each group.

[0058] The receptive field expansion branch consists of a 3×3 dilated convolution with a dilation rate of 2, a global average pooling layer, and a softmax activation function, simultaneously passing the feature map to the X-channel attention branch and the Y-channel attention branch. The input to the dilated convolution is a 64×32×32 feature map, and the output is a 64×32×32 feature map. The input to the average pooling layer is a 64×32×32 feature map, and the output is a 64×1×1 feature map. The input to the softmax activation function is a 64×1×1 feature map, and the output is a 1×64 feature map.

[0059] The X-channel attention branch includes an X-channel average pooling layer, and information is fused with the Y-channel attention branch; the input of the X-channel average pooling layer is a feature map of 64x32x32, and the output is a feature map of 64x1x32.

[0060] The Y-channel attention branch includes a Y-channel average pooling layer, and information is fused with the X-channel attention branch; the input of the Y-channel average pooling layer is a feature map of 64x32x32, and the output is a feature map of 64x32x1.

[0061] After the feature map X-channel average pooling layer and the Y-channel average pooling layer, splicing and a 1x1 convolution are performed, the input is a feature map of 64x1x32 and a feature map of 64x32x1, and the output is a feature map of 64x32x32; X-branch Sigmod activation function and Y-branch Sigmod activation function are performed respectively, the input of the X-branch Sigmod activation function is a feature map of 64x32x32, and the output is a feature map of 64x1x32; the input of the Y-branch Sigmod activation function is a feature map of 64x32x32, and the output is a feature map of 64x32x1; multiplication is performed on the feature map after grouping the initial feature map, the input is a feature map of 64x1x32, a feature map of 64x32x1 and a feature map of 64x32x32, and the output is a feature map of 64x32x32; group normalization is performed, the input is a feature map of 64x32x32, and the output is a feature map of 64x32x32; an average pooling layer is performed, the input is a feature map of 64x32x32, and the output is a feature map of 64x1x1; a Softmax activation function is performed, the input is a feature map of 64x1x1, and the output is a feature map of 1x64; multiplication is performed on the feature map after the dilated convolution of the receptive field expansion branch, and addition is performed on the feature map transmitted by the receptive field expansion branch, and then a Softmax activation function is performed, the input is a feature map of 1x64 and a feature map of 64x32x32, and the output is a feature map of 1x32x32; multiplication is performed on the feature map after grouping the initial feature map, the input is a feature map of 1x32x32, and the output is a feature map of 2048x32x32.

[0062] The global attention branch is composed of a global average pooling layer, a 1*1 convolution and a dimension expansion layer, and the feature map with attention is obtained by adding the feature map after fusing the X channel and the Y channel; the input of the global average pooling layer is the initial feature map of 2048*32*32, and the output is the feature map of 2048*1*1; the input of the 1*1 convolution is the feature map of 2048*1*1, and the output is the feature map of 2048*1*1; the input of the dimension expansion layer is the feature map of 2048*1*1, and the output is the feature map of 2048*32*32; the input of adding the global attention branch feature map and the feature map after fusing the X channel and the Y channel is the feature map of 2048*32*32 and the feature map of 2048*32*32, and the output is the multi-scale attention expansion feature map of 2048*32*32.

[0063] The detection module is used to construct a network detection head, and the network detection head is used to detect the multi-scale attention expansion feature map to obtain a predicted target classification score and a predicted target coordinate.

[0064] The detection module comprises a detection head construction unit, a classification unit, a candidate point matching unit and a regression unit.

[0065] The detection head construction unit is used to construct a network detection head, and the network detection head comprises a classification branch and a regression branch; the classification branch is composed of four 3*3 convolutions; and the regression branch is composed of four 3*3 convolutions.

[0066] The classification unit detects the multi-scale attention expansion feature map by using the classification branch, extracts features of the multi-scale attention expansion feature map through three 3*3 convolutions connected in series, and then predicts a target classification score through a 3*3 convolution.

[0067] The candidate point matching unit divides candidate points according to classification scores as evaluation indexes.

[0068] The regression unit detects the multi-scale attention expansion feature map by using the regression branch, extracts features of the multi-scale attention expansion feature map through three 3*3 convolutions connected in series, and then predicts a target coordinate through a 3*3 convolution.

[0069] The working process of the candidate point matching unit comprises a generation module, a screening module, a calculation module and a filtering module.

[0070] The generation module is used to generate a loss matrix according to the classification scores obtained by the classification branch for each labeled real point (assuming that there are three real targets in an image, and the model predicts 1000 prediction boxes, a loss matrix with a dimension of 3*1000 needs to be generated), and the value of the loss matrix is a classification loss, that is, 1-classification score, and 25 points around the anchor point step and anchor box size are selected as initial candidate points.

[0071] Innovation: The workflow of the candidate point matching unit is proposed by other papers in target detection methods under the annotation of rotating boxes or horizontal boxes. The present application applies it to point annotation target detection, and removes the positioning loss calculation in the loss matrix according to the characteristics of point annotation that cannot provide IoU, and only retains the classification loss calculation. In addition, the present application first selects the surrounding 25 initial candidate points before performing subsequent operations. (Generally, the target detection method (rotating box or horizontal box annotation) usually only uses IoU as the evaluation index. Point annotation cannot calculate IoU, and can directly use classification loss as the evaluation index, but the effect is usually poor. Although this method is equivalent to using classification loss as the evaluation index, the candidate points are refined step by step, which can improve the detection performance).

[0072] The screening module is used to screen the 25 candidate points by using a dynamic estimation method. First, a score matrix is generated, and the value of the score matrix is the classification score predicted in the classification branch. Then, the top K score maximum coarse-grained candidate points are selected for each real target in the score matrix (experimentally, K = 10 is the best), and the classification scores are summed and rounded down to further screen Q fine-grained candidate points. Figure 4

[0073] The calculation module is used to calculate the candidate point matching matrix of the real point according to the Q fine-grained candidate points screened above, that is, assigning 1 to the corresponding position of the Q fine-grained candidate points with the lowest loss value in the loss matrix for each real point, and assigning 0 to the remaining positions. Figure 5

[0074] The filtering module is used to filter the situation where a single candidate point is marked as a candidate point by multiple real points. The mutual matching matrix is calculated, that is, the values in the matching matrix are summed by column. If there is a column value greater than 1 (a single candidate point is marked as a candidate point by multiple real points) after summation, a unique real point needs to be assigned to the candidate point. Compare the values in the loss matrix, select the candidate point corresponding to the real point with a smaller loss matrix value as the final candidate point assigned to the real point, and assign 1 to the corresponding position in the mutual matching matrix and 0 to the remaining positions. The mutual matching matrix is the final candidate point assignment matrix for the real point. Figure 6

[0075] Embodiment Two

[0076] In this embodiment, a point annotation aerial image target detection method includes the following steps:

[0077] S1. Perform multi-class template picture overlay operation on the aerial image, and perform data enhancement processing to obtain a preprocessed image.

[0078] ​​​The method for obtaining the preprocessed image comprises: randomly flipping, randomly scaling and randomly rotating the aerial image to obtain a plurality of preprocessed images with different orientations and sizes.

[0079] S2. Constructing a feature extraction network and extracting features of the preprocessed image by using the feature extraction network to obtain an initial feature map.

[0080] The feature extraction network is constructed by using ResNet50 as a backbone network; and the initial feature map is obtained by using the backbone network to extract features of the preprocessed image.

[0081] The backbone network comprises a C1 layer, a C2 layer, a C3 layer, a C4 layer and a C5 layer; the C1 layer is composed of one 7x7 convolution, one BN layer, one ReLU activation function and one maximum pooling layer; the C2 layer is composed of six 1x1 convolutions, three 3x3 convolutions, three BN layers and three ReLU activation functions; the C3 layer is composed of eight 1x1 convolutions, four 3x3 convolutions, four BN layers and four ReLU activation functions; the C4 layer is composed of twelve 1x1 convolutions, six 3x3 convolutions, six BN layers and six ReLU activation functions; and the C5 layer is composed of six 1x1 convolutions, three 3x3 convolutions, three BN layers and three ReLU activation functions.

[0082] S3. Constructing a multi-scale attention expansion network comprising four multi-scale attention expansion modules composed of expansion convolutions with expansion rates of 2, 4, 6 and 8 connected in series, and performing deep processing on the initial feature map to obtain a multi-scale attention expansion feature map.

[0083] S4. Constructing a network detection head and detecting the multi-scale attention expansion feature map by using the network detection head to obtain a predicted target classification score and a predicted target coordinate.

[0084] The network detection head comprises a classification branch, a candidate point matching unit and a regression branch; the classification branch is used to detect the multi-scale attention expansion feature map to obtain the predicted target classification score; the candidate point matching unit is used to match the candidate points; and the regression branch is used to detect the multi-scale attention expansion feature map to obtain the predicted target coordinate.

[0085] The classification branch is composed of four 3x3 convolutions;

[0086] The matching method of the candidate points comprises:

[0087] For each labeled real point, 25 points around it are selected according to the anchor point step and the anchor box size, and a loss matrix is generated according to the classification score; the above 25 candidate points are used to dynamically estimate the top K score coarse-grained candidate points of each real point, the sum of the scores is calculated and further filtered to obtain Q fine-grained candidate points; the candidate point matching matrix of the real point is calculated for the above screened Q fine-grained candidate points; the situation that a single candidate point is marked as a candidate point by multiple real points at the same time is filtered, the mutual matching matrix is calculated, and the final candidate point assigned to each real target is obtained;

[0088] The regression branch is composed of 4 3x3 convolutions.

[0089] The above-described embodiments are merely descriptions of the preferred modes of the present application and are not intended to limit the scope of the present application. Various modifications and improvements to the technical solutions of the present application made by those of ordinary skill in the art without departing from the design spirit of the present application shall fall within the protection scope of the present application as defined by the claims.

Claims

1. A target detection system for point-annotated aerial images, characterized in that, The system includes: an image preprocessing module, a feature extraction module, a multi-scale attention expansion module, and a detection module; The image preprocessing module is used to perform data enhancement processing on aerial images to obtain preprocessed images; The feature extraction module is used to construct a feature extraction network and use the feature extraction network to extract features from the preprocessed image to obtain an initial feature map; The multi-scale attention expansion module is used to perform depth processing on the initial feature map to obtain a multi-scale attention expansion feature map. The detection module is used to predict the target classification score using the classification branch of the network detection head, and to perform candidate point division and detection on the multi-scale attention expansion feature map using the classification score as an indicator, and to predict the target coordinates using the regression branch of the network detection head. The multi-scale attention expansion module first undergoes 32 group processing steps, followed by four branches: receptive field expansion branch, X-channel attention branch, Y-channel attention branch, and global attention branch. The receptive field expansion branch consists of a 3×3 dilated convolution with an expansion rate of 2, a global average pooling layer, and a softmax activation function, and simultaneously transmits information to the X-channel attention branch and the Y-channel attention branch. The X-channel attention branch is composed of one X-channel average pooling layer; The Y-channel attention branch is composed of a single Y-channel average pooling layer; The global attention branch consists of a global average pooling layer, a 1×1 convolutional layer, and a dimension expansion layer; In this process, after the feature maps are grouped, they are concatenated through X-channel and Y-channel average pooling layers and then restored to the original two branches through a 1×1 convolution. Each branch is then multiplied with the grouped feature maps using a Sigmoid activation function, followed by a group normalization layer, a global average pooling layer, and a Softmax activation function. Finally, the convolutional feature maps from the receptive field expansion branch are multiplied with the information passed through the receptive field expansion branch, and then multiplied again with the grouped feature maps using a Sigmoid activation function to obtain an attention-enabled feature map. The global attention branch is then added to the attention-enabled feature map to obtain a multi-scale attention-enhanced feature map.

2. The system according to claim 1, characterized in that, The feature extraction module includes: a network construction unit and a feature extraction unit; The network construction unit is used to build a backbone network using ResNet50; The feature extraction unit is used to extract features from the preprocessed image using the backbone network to obtain an initial feature map.

3. The system according to claim 2, characterized in that, The backbone network includes: layer C1, layer C2, layer C3, layer C4, and layer C5; The C1 layer consists of one 7×7 convolution, one batch normalization (BN) layer, one ReLU activation function, and one max pooling layer; the C2 layer consists of six 1×1 convolutions, three 3×3 convolutions, three BN layers, and three ReLU activation functions; the C3 layer consists of eight 1×1 convolutions, four 3×3 convolutions, four BN layers, and four ReLU activation functions; the C4 layer consists of twelve 1×1 convolutions, six 3×3 convolutions, six BN layers, and six ReLU activation functions; and the C5 layer consists of six 1×1 convolutions, three 3×3 convolutions, three BN layers, and three ReLU activation functions.

4. The system according to claim 1, characterized in that, The detection module includes: a detection head construction unit, a classification unit, a candidate point matching unit, and a regression unit; The detection head construction unit is used to construct the network detection head, which includes a classification branch and a regression branch. The classification unit is used to detect the multi-scale attention expansion feature map using the classification branch to obtain the predicted target classification score; The candidate point matching unit is used to match candidate points for each real target using the classification score as an indicator. The regression unit is used to detect the multi-scale attention expansion feature map using the regression branch to obtain the predicted target coordinates.

5. The system according to claim 4, characterized in that, The classification branch consists of four 3×3 convolutions; the regression branch consists of four 3×3 convolutions.

6. The system according to claim 4, characterized in that, The candidate point matching unit includes: a generation module, a filtering module, a calculation module, and a filtering module; The generation module is used to generate a loss matrix for each labeled ground truth point based on the classification score obtained from the classification branch. The value of the loss matrix is ​​the classification loss, i.e., 1 - classification score. The module also selects 25 points around the ground truth point as the initial candidate points based on the anchor step size and the anchor frame size. The filtering module is used to dynamically estimate 25 candidate points. First, a score matrix is ​​generated, and the value of the score matrix is ​​the predicted classification score in the classification branch. Then, the top K coarse-grained candidate points with the largest scores are selected for each true target in the score matrix. The classification scores are summed and rounded down to further filter into Q fine-grained candidate points. The calculation module is used to calculate the candidate point matching matrix of the real points based on the Q selected candidate points; The filtering module is used to filter the case where a single candidate point is simultaneously marked as a candidate point by multiple real points, and selects the real point with the smallest loss value in the loss matrix of multiple real points as the candidate point of that real point.

7. A method for target detection in aerial images with point annotations, said method being implemented using the system described in any one of claims 1-6, characterized in that, The method includes: The aerial images are overlaid with multiple types of template images, and then randomly flipped, scaled, and rotated to obtain the preprocessed images. A feature extraction network is constructed, and the feature extraction network is used to extract features from the preprocessed image to obtain an initial feature map; A multi-scale attention expansion network is constructed, comprising four cascaded multi-scale attention expansion modules consisting of expansion convolutions with expansion rates of 2, 4, 6, and 8, which are used to process the initial feature map depth to obtain a multi-scale attention expansion feature map. A network detection head is constructed, and the classification branch of the network detection head is used to detect the multi-scale attention-expanded feature map to obtain the predicted target classification score; the candidate point matching unit is used to match candidate points; and the regression branch is used to detect the multi-scale attention-expanded feature map to obtain the predicted target coordinates.

Images