Aerial target detection method based on cross-weighted pixel reconstruction

Abstract

This invention discloses a method for aerial target detection based on cross-weighted pixel reconstruction. The method comprises obtaining an image dataset containing aerial targets and preparing training samples; constructing an aerial target detection network model comprising a CPR module and an FS-PANet network module using the Yolov10 network as a base network; inputting the training samples into the aerial target detection network model for training, and saving the trained network model parameters; and inputting the image to be detected into the trained aerial target detection network model for aerial target detection. The method has the following significant effects: by mining the height, width, and channel direction of the input feature map, it can capture directional and positional information, enabling the model to more accurately locate and identify targets of interest. Furthermore, through an average pooling operation, it can effectively remove noise contained in low-resolution feature maps, reducing interference and strengthening target features. This improves feature fusion efficiency, thereby increasing target detection accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of aerial target detection, and in particular to an aerial target detection method based on cross-weighted pixel reconstruction. Background Art

[0002] In recent years, drones have become increasingly popular in urban management, precision agriculture, environmental monitoring, traffic control, military operations, and disaster relief due to their flexibility, versatility, low cost, and ease of operation. However, the rapid adoption of drones has also raised numerous concerns, such as the use of drones for illegal attacks, interference, or surveillance, which seriously impacts social security and public privacy. Therefore, effective drone detection is crucial and has become a hot topic of research. Currently, mainstream drone detection methods include radar, audio, and optoelectronic visual detection. Compared with other detection technologies, optoelectronic visual detection offers more intuitive results, higher accuracy, and a wider range of applications, making it the mainstream method for drone detection.

[0003] However, drones are becoming increasingly smaller, especially when they are farther from the sensor. The imaging scale of drones in optical sensors is significantly reduced, and they suffer from problems such as reduced shape texture and degraded surface features. Furthermore, drones have weak contrast with the background, a low signal-to-clutter ratio, and are susceptible to interference from cloud clutter and birds. These unfavorable factors pose serious challenges to image detection. Traditional target detection methods often rely on hand-crafted features, which struggle to accurately represent the complex and ever-changing target morphology and appearance. Furthermore, these methods typically employ fixed feature extractors and classifiers, often exhibiting poor detection performance in complex scenes, making them difficult to effectively apply to drone detection.

[0004] In recent years, with the rapid advancement of GPU parallel computing technology, deep learning algorithms have developed rapidly and have been widely used in the field of object detection. Currently, mainstream deep learning-based object detection algorithms can be divided into two main categories. One is the two-stage detection algorithm, such as R-CNN (Convolutional Neural Network, CNN), Fast R-CNN, Faster R-CNN, and Mask R-CNN. These algorithms first generate region proposals on the input image and then perform feature extraction and classification. While they offer high detection accuracy, they suffer from large model parameter sizes and slow detection speed, making them difficult to meet the requirements of high-speed real-time detection. The other is the single-stage algorithm, which mainly includes SSD (Single-Shot MultiBox Detector), RetinaNet, and YOLO (You Only Look Once). Single-stage algorithms directly transform object bounding box localization into a regression problem. Without generating proposals, they significantly increase detection speed, but their detection accuracy is inferior to that of two-stage detection algorithms. Therefore, to maintain high detection speed and accuracy for drones, improvements to deep learning-based detection algorithms are needed.

[0005] To this end, a research on anti-UAV target detection algorithm based on YOLOv5s-AntiUAV[J].(Tan Liang, Zhao Liangjun, Zheng Liping, Xiao Bo, Electro-Optics and Control, 2024, 31(5):40-45.) discloses a detection method for UAVs. This method first introduces the Slim Neck paradigm combined with deep hyperparameter convolution to enhance the algorithm's feature extraction and representation learning capabilities; secondly, the SPD-Conv module is introduced in the backbone and neck networks respectively to avoid information loss caused by strided convolution and pooling, thereby improving the algorithm's detection performance of small targets in low-resolution images; finally, the loss function is optimized, and Alpha-CIoULoss is used to replace the CIOU Loss in YOLOv5s to enhance the algorithm's versatility.

[0006] However, this method still has the following disadvantages:

[0007] 1. Weak feature extraction capabilities affect target detection stability. This technical solution calculates feature correlations across the channel dimension and assigns weights to different channels. However, it only learns the degree of correlation between channels and fails to capture the spatial information of the image. This makes it difficult to effectively extract targets in complex backgrounds, significantly affecting the stability of target detection.

[0008] 2. Weak feature fusion capabilities affect target detection accuracy. This technical solution's feature fusion method effectively utilizes feature maps from different hierarchies, facilitating multi-scale target detection. However, this feature fusion method also dilutes the semantic information of features to a certain extent, thereby losing some important information within the feature map and reducing target detection accuracy. Summary of the Invention

[0009] In view of the shortcomings of the existing technology, the purpose of the present invention is to provide an aerial target detection method based on cross-weighted pixel reconstruction. By constructing a cross-weighted pixel reconstruction module to fully mine the spatial information and channel information of the feature map, and by designing a guided feature selection path aggregation network module to improve the quality of the low-level feature map, the problems of low target detection stability and low target detection accuracy existing in the background technology can be effectively solved.

[0010] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0011] A method for detecting aerial targets based on cross-weighted pixel reconstruction includes the following steps:

[0012] Step 1: Obtain an image dataset containing aerial targets and create training samples;

[0013] Step 2: Based on the Yolov10 network as the basic network, a network model for aerial target detection is constructed, which includes the CPR module and the FS-PANet network module.

[0014] The CPR module is used to capture different dimensional information along the height, width and channel directions of the feature map after two convolutions in the backbone of the Yolov10 network, and generate cross weights to reconstruct the pixels of the feature map;

[0015] The FS-PANet network module is used to filter lower-level features from top to bottom using adjacent higher-level features for feature maps of different scales output by the backbone of the Yolov10 network, and fuse the filtered lower-level features with the higher-level features. The fused features are then integrated from bottom to top into features of different scales and output to the head of the Yolov10 network.

[0016] Step 3: input the training samples into the aerial target detection network model for training, and save the trained network model parameters;

[0017] Step 4: Input the image to be detected into the trained aerial target detection network model for aerial target detection.

[0018] Furthermore, the process of preparing the training samples is as follows:

[0019] Step 1.1: Select a data source and obtain a sample set from it.

[0020] Step 1.2: Label the images in the sample set;

[0021] Step 1.3: Cut the marked image to the required size;

[0022] Step 1.4: Convert the label format to the required format to obtain training samples.

[0023] Furthermore, the training samples are composed of a training set and a test set, the training set is used to train the constructed aerial target detection network model, and the test set is used to evaluate the performance of the aerial target detection network model after training.

[0024] Furthermore, the CPR module processes the input feature map as follows:

[0025] Calculate the channel-wise feature weights of the input feature map;

[0026] For the input feature map, features are aggregated in the height and width directions to obtain feature maps in the height and width directions respectively;

[0027] Apply the channel-wise weights to the feature maps in the height and width directions, and calculate the feature weights in the height and width directions respectively;

[0028] The feature weights in the height and width directions are crossed to obtain cross weights in multiple dimensions, and the cross weights are multiplied with the input feature map, and the input feature map is recalibrated to obtain the output feature map.

[0029] Furthermore, the CPR module includes a global average pooling unit, a height average pooling unit, a width average pooling unit, three BConv units, three Sigmoid function units, and four Hadamard product units. The input feature maps are respectively input into the input ends of the global average pooling unit, the height average pooling unit, and the width average pooling unit. The output end of the global average pooling unit is connected to the input end of the first Sigmoid function unit via the first BConv unit, the output end of the height average pooling unit is connected to the first input end of the first Hadamard product unit via the second BConv unit, and the output end of the width average pooling unit is connected to the first input end of the second Hadamard product unit via the third BConv unit. The input end, the two output ends of the first Sigmoid function unit are respectively connected to the second input end of the first Hadamard product unit and the second input end of the second Hadamard product unit, the output end of the first Hadamard product unit is connected to the first input end of the third Hadamard product unit via the second Sigmoid function unit, the output end of the second Hadamard product unit is connected to the second input end of the third Hadamard product unit via the third Sigmoid function unit, the output end of the third Hadamard product unit is connected to the first input end of the fourth Hadamard product unit, the second input end of the fourth Hadamard product unit inputs the input feature map, and the output end of the fourth Hadamard product unit outputs the output feature map, wherein:

[0030] The global average pooling unit is used to aggregate channel-wise features in the input feature map;

[0031] The height average pooling unit is used to aggregate the height-direction features in the input feature map;

[0032] The width average pooling unit is used to aggregate the features in the width direction of the input feature map;

[0033] Three BConv units are used to learn parameters;

[0034] The three Sigmoid function units are used to calculate the feature weights in the channel, height, and width directions respectively;

[0035] The first Hadamard product unit is used to apply the feature weight of the channel direction to the feature map in the height direction;

[0036] The second Hadamard product unit is used to apply the feature weights in the channel direction to the feature map in the width direction;

[0037] The third Hadamard product unit is used to cross the feature weights in the height and width directions to obtain cross weights in multiple dimensions;

[0038] The fourth Hadamard product unit is used to apply the cross weights to the input feature map to obtain the output feature map.

[0039] Furthermore, the calculation formula of the feature weight in the channel direction is:

[0040] ω ch =σ(conv 1×1 (ReLU(conv 1×1 (GAP(F)))))

[0041] Among them, ω ch is the weight in the channel direction, GAP(·) represents the global average pooling with a pooling size of 1×H×W, conv 1×1 (·) represents 1×1 convolution, ReLU(·) represents activation function, and σ(·) represents sigmoid activation function;

[0042] The calculation formula of the feature weight in the height direction is:

[0043] ω height =σ(F height ⊙expand(ω ch ))

[0044] The calculation formula for the feature weight in the width direction is:

[0045] ω width =σ(F width ⊙expand(ω ch ))

[0046] Here, expand(·) represents the dimension expansion operation, and ⊙ represents the Hadamard product.

[0047] Furthermore, the FS-PANet network module includes two AFS modules, three C2f modules, a convolution module, two Concat modules, an SCDown downsampling module and a C2fCIB module. The two input ends of the first AFS module are respectively input with the high-level feature map and the middle-level feature map extracted by the backbone of the Yolov10 network. The output end of the first AFS module is connected to the input end of the first C2f module. The two output ends of the first C2f module are respectively connected to an input end of the second AFS module and an input end of the first Concat module. The other input end of the second AFS module is input with the low-level feature map extracted by the backbone of the Yolov10 network. The output ends of the two AFS modules are connected to the input end of the second C2f module, one output end of the second C2f module is connected to the other input end of the first Concat module via the convolution module, the output end of the first Concat module is connected to one input end of the second Concat module via the third C2f module and the SCDown downsampling module, the other input end of the second Concat module inputs the high-level feature map, the output end of the second Concat module is connected to the input end of the C2fCIB module, and the other output end of the second C2f module, the other output end of the third C2f module, and the output end of the C2fCIB module are all connected to the head part of the Yolov10 network.

[0048] Furthermore, the AFS module includes a transposed convolution unit, a CA attention unit, a multiplication unit, and a feature fusion unit. The transposed convolution unit is used to perform transposed convolution on the input higher-level feature map to obtain an intermediate feature map of the same size as the lower-level feature map. The CA attention unit is used to calculate the weight of each feature channel. The multiplication unit is used to multiply the channel weight of the obtained higher-level feature map with the lower-level feature map to enhance the useful features in the lower-level feature map. The feature fusion unit is used to perform feature fusion on the higher-level feature map and the enhanced lower-level feature map and output them.

[0049] Furthermore, the AFS module calculates the weight of each feature channel using the following formula:

[0050] ω=σ(BConv(GAP 1×2H×2W (C' high ))+BConv(GMP 1×2H×2W (C' high )))

[0051] Among them, ω represents the weight of each feature channel, σ(·) represents the sigmoid activation function, GAP 1×2H×2W (·), GMP 1×2H×2W(·) represent global average pooling and global maximum pooling with pooling size of 1×2H×2W, respectively.

[0052] Furthermore, the loss function of the aerial target detection network model is:

[0053]

[0054] Among them, L inner-MPDIoU is the loss function, IoU inner is the inner intersection and union ratio, d1 is the distance between the upper left corner coordinates of the true bounding box and the predicted bounding box, d2 is the distance between the lower right corner coordinates of the true bounding box and the predicted bounding box, and w and h are the width and height of the input feature map respectively.

[0055] The remarkable effects of the present invention are:

[0056] 1. Improved target detection stability. This method mines information on the height, width, and channel direction of the input feature map. On the one hand, it can capture directional and positional information, enabling the model to more accurately locate and identify targets of interest. On the other hand, the average pooling operation can effectively remove noise contained in low-resolution feature maps, reducing interference and enhancing target features.

[0057] 2. Improved target detection accuracy. This invention fuses adjacent feature maps, uses higher-level feature maps to filter important features from adjacent lower-level feature maps, then fuses the filtered lower-level features with the higher-level features, further integrating them into features of different scales. This improves the efficiency of feature fusion and thus improves target detection accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 is a flow chart of the method of the present invention;

[0059] Figure 2 It is the overall structure diagram of the aerial target detection network model of the present invention;

[0060] Figure 3 This is the structural diagram of the CPR module;

[0061] Figure 4 This is the structural diagram of the FS-PANet network module. DETAILED DESCRIPTION

[0062] The specific implementation manner and working principle of the present invention will be further described in detail below with reference to the accompanying drawings.

[0063] This embodiment aims to solve the problem that drones are small in size and easily interfered by similar-shaped targets such as flying birds in drone detection tasks. A Cross-weighted Pixel Reconstructionand Feature Selection Network CFS-YOLO is proposed based on the Yolov10n network. First, through the CPR module added to the backbone, it is possible to fully mine the information of different dimensions of the feature map and generate cross weights to reconstruct the pixels of the feature map to reduce background noise interference and improve the feature expression ability of the network. Then, a selective feature fusion mechanism is designed to replace the neck part in the traditional Yolov10n. By using the rich semantic information of the high-level feature map to guide the low-level feature map to screen important features, and fully fuse the screened features in a top-down and bottom-up bidirectional manner, it effectively combines the detailed information of the feature maps of different layers to improve the performance of the model when dealing with small targets, similar targets and occluded targets. The specific embodiments are described as follows:

[0064] This embodiment provides an aerial target detection method based on cross-weighted pixel reconstruction, such as Figure 1 The specific steps are as follows:

[0065] Step 1: Obtain an image dataset containing aerial targets and create training samples;

[0066] Step 1.1: Select a data source and obtain a sample set from it.

[0067] Step 1.2: Label the images in the sample set;

[0068] Step 1.3: Cut the marked image to the required size;

[0069] Step 1.4: Convert the label format to the required format to obtain training samples.

[0070] In this patent, the training samples consist of a training set and a test set, and the ratio of the training set to the test set is 8:2. The training set is used to train the constructed aerial target detection network model, and the test set is used to evaluate the performance of the aerial target detection network model after training.

[0071] Step 2: Using the Yolov10n network as the basic network, build an aerial target detection network model including the CPR module and the FS-PANet network module;

[0072] In this example, the aerial target detection network model is constructed as follows Figure 2As shown in the figure, the CPR module is added to the backbone of the Yolov10n network to mine information on the height, width, and channel direction of the input feature map, and the FS-PANet network module is used to replace the neck part of the Yolov10n network to fuse adjacent feature maps. Specifically:

[0073] The CPR (Cross-weighted Pixel Reconstruction) module is used to capture different dimensional information along the height, width, and channel directions of the feature map after two convolutions in the backbone of the Yolov10 network. The average pooling operation is used to reduce information redundancy and suppress background noise along the height and width directions, thereby highlighting the features of the target of interest. The global average pooling operation is used to aggregate channel information, and weights are assigned to channels according to the importance of different channels, guiding the model to use limited resources to learn important channel features. Further, the parameters of the different dimensional information after the pooling operation are learned using a BConv module composed of pointwise convolution and ReLU activation function. The attention weights of different dimensions are generated by the Sigmoid function, and the different weights are crossed and applied to the original feature map. Each pixel of the original feature map is reconstructed, reducing background interference while improving the model's ability to recognize targets of interest.

[0074] like Figure 3 As shown, the CPR module includes a global average pooling unit, a height average pooling unit, a width average pooling unit, three BConv units, three Sigmoid function units, and four Hadamard product units. The input feature maps are respectively input into the input ends of the global average pooling unit, the height average pooling unit, and the width average pooling unit. The output end of the global average pooling unit is connected to the input end of the first Sigmoid function unit via the first BConv unit, the output end of the height average pooling unit is connected to the first input end of the first Hadamard product unit via the second BConv unit, and the output end of the width average pooling unit is connected to the first input end of the second Hadamard product unit via the third BConv unit. The two output ends of the first Sigmoid function unit are respectively connected to the second input end of the first Hadamard product unit and the second input end of the second Hadamard product unit, the output end of the first Hadamard product unit is connected to the first input end of the third Hadamard product unit via the second Sigmoid function unit, the output end of the second Hadamard product unit is connected to the second input end of the third Hadamard product unit via the third Sigmoid function unit, the output end of the third Hadamard product unit is connected to the first input end of the fourth Hadamard product unit, the second input end of the fourth Hadamard product unit inputs the input feature map, and the output end of the fourth Hadamard product unit outputs the output feature map, wherein:

[0075] The global average pooling unit is used to aggregate channel-wise features in the input feature map;

[0076] The height average pooling unit is used to aggregate the height-direction features in the input feature map;

[0077] The width average pooling unit is used to aggregate the features in the width direction of the input feature map;

[0078] Three BConv units are used to learn parameters;

[0079] The three Sigmoid function units are used to calculate the feature weights in the channel, height, and width directions respectively;

[0080] The first Hadamard product unit is used to apply the feature weight of the channel direction to the feature map in the height direction;

[0081] The second Hadamard product unit is used to apply the feature weights in the channel direction to the feature map in the width direction;

[0082] The third Hadamard product unit is used to cross the feature weights in the height and width directions to obtain cross weights in multiple dimensions;

[0083] The fourth Hadamard product unit is used to apply the cross weights to the input feature map to obtain the output feature map.

[0084] In this example, the CPR module is processing the input feature map The processing steps are as follows:

[0085] Step A1: The input image is convolved twice on the backbone to obtain the input feature map. Calculate input feature map The feature weight of the channel direction is calculated as follows:

[0086] ω ch =σ(conv 1×1 (ReLU(conv 1×1 (GAP(F))))) (1)

[0087] Among them, ω ch is the weight in the channel direction, GAP(·) represents the global average pooling with a pooling size of 1×H×W, conv 1×1 (·) represents 1×1 convolution, ReLU(·) represents the activation function, and σ(·) represents the sigmoid activation function.

[0088] Step A2: Use the average pooling operation and BConv unit to aggregate the features of the input feature map in the height and width directions to obtain feature maps in the height and width directions respectively. The calculation process is as follows:

[0089] F height =conv 1×1 (ReLU(conv 1×1 (AvgPool 1×H×1 (F)))) (2)

[0090] F width =conv 1×1 (ReLU(conv 1×1 (AvgPool 1×1×W (F)))) (3)

[0091] Among them, F height 、F width They are feature maps in height and width directions, AvgPool 1×H×1 (·) represents the average pooling with a pooling size of 1×H×1, AvgPool 1×1×W (·) is 1×1×W average pooling.

[0092] Step A3: Apply the channel-wise weights to the feature maps in the height and width directions, and calculate the feature weights in the height and width directions respectively;

[0093] Among them, the calculation formula of the feature weight in the height direction is:

[0094] ω height =σ(F height ⊙expand(ω ch )) (4)

[0095] Among them, the calculation formula of the feature weight in the width direction is:

[0096] ω width =σ(F width ⊙expand(ω ch )) (5)

[0097] In the formula, expand(·) represents the dimension expansion operation, and v represents the Hadamard product.

[0098] Step A4: Cross the feature weights in the height and width directions to obtain cross weights in multiple dimensions, perform dot multiplication on the cross weights and the input feature map, and recalibrate the input feature map to obtain the output feature map;

[0099] The calculation formula for the output feature map is as follows:

[0100] F'=Fv(expand(ω width )vexpand(ω height )) (6)

[0101] In this example, to improve the efficiency of feature fusion at different layers, a FS-PANet (Feature Selection Path Aggregation Network) network module is proposed to replace the neck part of the traditional Yolov10n network. First, an Adjacent Feature Selection (AFS) module is constructed to filter low-level features from the backbone of the Yolov10n network using adjacent high-level features from top to bottom. The filtered low-level features are then fused with the high-level features. The fused features are further integrated from bottom to top to output features of different scales to the head part of the Yolov10n network.

[0102] Specifically, the structure of the FS-PANet network module is as follows Figure 4 As shown, it includes two AFS modules, three C2f modules, a convolution module, two Concat modules, a SCDown downsampling module and a C2fCIB module. The two input ends of the first AFS module are respectively input into the high-level feature map and the middle-level feature map extracted from the backbone of the Yolov10 network. The output end of the first AFS module is connected to the input end of the first C2f module. The two output ends of the first C2f module are respectively connected to an input end of the second AFS module and an input end of the first Concat module. The other input end of the second AFS module is input into the low-level feature map extracted from the backbone of the Yolov10 network. The output end is connected to the input end of the second C2f module, one output end of the second C2f module is connected to the other input end of the first Concat module via the convolution module, the output end of the first Concat module is connected to an input end of the second Concat module via the third C2f module and the SCDown downsampling module, the other input end of the second Concat module inputs the high-level feature map, the output end of the second Concat module is connected to the input end of the C2fCIB module, and the other output end of the second C2f module, the other output end of the third C2f module, and the output end of the C2fCIB module are all connected to the head part of the Yolov10 network.

[0103] exist Figure 4In the FS-PANet network module, C2, C3, C4, and C5 are feature maps extracted from the backbone. As the scale of the feature map decreases, the receptive field of the feature map on the input image gradually increases. Therefore, the high-level feature map contains more semantic information than the low-level feature map. Using adjacent high-level feature maps to filter low-level feature maps can improve the quality of low-level features, which is conducive to distinguishing similar target features and improving the fusion efficiency of features at different layers. The main part of the FS-PANet network module that implements adjacent feature filtering and then fusion is the AFS module. The AFS module takes two adjacent layers of feature maps as input, and the high-level feature map is adjusted to the same scale as the low-level feature map through 3×3 transposed convolution, that is, Figure 4 The feature maps M3, M4, and M5 in the image are constructed. The CA (Channel attention) module is used to measure the importance of each feature channel and output the channel weight. The obtained high-level channel weight is multiplied by the low-level feature map to enhance the useful features in the low-level feature map and suppress the useless features. The feature maps P3, P4, and P5 after feature fusion are obtained, thereby improving the efficiency of feature fusion, making full use of semantic information, and reducing information redundancy in feature fusion.

[0104] from Figure 4 It can also be seen that the AFS module includes a transposed convolution unit, a CA attention unit, a multiplication unit, and a feature fusion unit. The transposed convolution unit is used to perform transposed convolution on the input higher-level feature map to obtain an intermediate feature map of the same size as the lower-level feature map. The CA attention unit is used to calculate the weight of each feature channel. The multiplication unit is used to multiply the channel weight of the obtained higher-level feature map with the lower-level feature map to enhance the useful features in the lower-level feature map. The feature fusion unit is used to fuse the higher-level feature map and the enhanced lower-level feature map and output them.

[0105] Based on the above structure, the main processing process of the AFS module is as follows:

[0106] Step B1: Take two adjacent feature maps As input, the high-level feature map C high Perform transposed convolution to obtain the intermediate feature map C' high , C' high and low-level feature map C low The dimensions are the same, and the calculation formula is:

[0107] C′ high =TransConv(C high ) (7)

[0108] Among them, TransConv(·) is a 3×3 transposed convolution with a stride of 2.

[0109] Step B2, intermediate feature map C' high After the CA attention unit performs global average pooling (GAP) and global maximum pooling (GMP) to extract weight features, the calculation formula is as follows:

[0110] ω=σ(BConv(GAP 1×2H×2W (C' high ))+BConv(GMP 1×2H×2W (C' high ))) (8)

[0111] Among them, ω represents the weight of each feature channel, σ(·) represents the sigmoid activation function, GAP 1×2H×2W (·), GMP 1×2H×2W (·) represent global average pooling and global maximum pooling with pooling size of 1×2H×2W, respectively.

[0112] The extracted weight features are then added together after the excitation operation, and then converted into a weight of 0 to 1 through the sigmoid function.

[0113] Step B3: The multiplication unit multiplies the obtained channel weight of the higher-level feature map by the lower-level feature map to enhance the useful features in the lower-level feature map;

[0114] Step B4: The feature fusion unit fuses the higher-level feature map and the enhanced lower-level feature map and outputs the result.

[0115] In this example, MPDIoU loss is used in conjunction with Inner-IoU based on auxiliary bounding boxes to calculate the bounding box regression loss. Compared to MPDIoU, Inner-MPDIoU uses auxiliary bounding boxes of different scales to calculate the loss, focusing more on the core part of the bounding box, enabling more accurate judgment of overlapping areas, and effectively accelerating the bounding box regression process. The calculation process of the Inner-MPDIoU loss function is as follows:

[0116]

[0117]

[0118] Among them, L inner-MPDIoU is the loss function, IoU inner is the inner intersection over union (Inner-IoU), d1 is the true bounding box Β gt and the predicted bounding box Β prd The distance between the upper left corner coordinates, d2 is the real bounding box Βgt and the predicted bounding box Β prd The distance between the lower right corner coordinates of the predicted bounding box Β prd The width and height are w prd 、h prd , the coordinates of its upper left and lower right corners are The coordinates of its center point are The ground-truth bounding box Β gt The width and height are w gt 、h gt , the coordinates of its upper left and lower right corners are The coordinates of its center point are w and h are the width and height of the input feature map respectively.

[0119] In this embodiment, the calculation formula of the inner intersection over union (Inner-IoU) is as follows:

[0120]

[0121] uinon=(w gt *h gt )*ratio 2 +(w prd *h prd )*ratio 2 -inter (12)

[0122]

[0123]

[0124] Among them, ratio is the scale factor, and its value range is [0.5, 1.5].

[0125] Step 3: input the training samples into the aerial target detection network model for training, and save the trained network model parameters;

[0126] This example uses Windows 10, 16GB of RAM, an Intel(R) Core(TM) i7-13700KF CPU, and Python 3.11.9, PyTorch 2.0.1, and CUDA 11.8. All models are trained and tested on an NVIDIA RTX 4080 GPU.

[0127] Training and evaluation process:

[0128] ①Use the conda create-n yolov10 python=3.11.9 command to create a virtual environment named yolov10 and activate the environment;

[0129] ②Then install PyTorch and related dependent libraries;

[0130] ③ According to the specific situation of the dataset, modify the corresponding yolov10.yaml configuration file to specify the number of categories, the paths of the training set and the test set, and other information;

[0131] ④ Use the training command to start training. After training is completed, the model is evaluated on the test set and the evaluation indicators such as accuracy, recall rate, mAP and other test results are output.

[0132] Step 4: Input the image to be detected into the trained aerial target detection network model for aerial target detection.

[0133] To sum up, in order to solve the problems of low target detection stability and low target detection accuracy in the background technology, the present invention, firstly, mines information on the height, width, and channel direction of the input feature map through the CPR module. On the one hand, it can capture direction perception and position perception information, so that the model can more accurately locate and identify the target of interest. On the other hand, the average pooling operation can well remove the noise contained in the low-resolution feature map, reduce interference and enhance the target features; secondly, the FS-PANet network module is used to fuse adjacent feature maps, and the higher-level feature map is used to filter the important features of the adjacent lower-level feature maps. Then, the filtered lower-level features are fused with the higher-level features, and further integrated into features of different scales, thereby improving the feature fusion efficiency and thus improving the target detection accuracy.

[0134] The technical solution provided by the present invention is introduced in detail above. Specific examples are used herein to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of the present invention, the present invention can also be improved and modified in a number of ways, and these improvements and modifications also fall within the scope of protection of the claims of the present invention.

Claims

1. A method for aerial target detection based on cross-weighted pixel reconstruction, characterized in that: The steps include: Step 1: Obtain an image dataset containing aerial targets and create training samples; Step 2: Based on the Yolov10 network as the basic network, a network model for aerial target detection is constructed, which includes the CPR module and the FS-PANet network module. The CPR module is used to capture different dimensional information along the height, width and channel directions of the feature map after two convolutions in the backbone of the Yolov10 network, and generate cross weights to reconstruct the pixels of the feature map; The CPR module processes the input feature map as follows: Calculate the channel-wise feature weights of the input feature map; For the input feature map, features are aggregated in the height and width directions to obtain feature maps in the height and width directions respectively; Apply the channel-wise weights to the feature maps in the height and width directions, and calculate the feature weights in the height and width directions respectively; The feature weights in the height and width directions are crossed to obtain cross weights in multiple dimensions, and the cross weights are multiplied by the input feature map, and the input feature map is recalibrated to obtain the output feature map; The FS-PANet network module is used to filter lower-level features from top to bottom using adjacent higher-level features for feature maps of different scales output by the backbone of the Yolov10 network, and fuse the filtered lower-level features with the higher-level features. The fused features are then integrated from bottom to top into features of different scales and output to the head of the Yolov10 network. Step 3: input the training samples into the aerial target detection network model for training, and save the trained network model parameters; Step 4: Input the image to be detected into the trained aerial target detection network model for aerial target detection.

2. The aerial target detection method based on cross-weighted pixel reconstruction according to claim 1, characterized in that: The process of making the training samples is as follows: Step 1.1: Select a data source and obtain a sample set from it. Step 1.2: Label the images in the sample set; Step 1.3: Cut the marked image to the required size; Step 1.4: Convert the label format to the required format to obtain training samples.

3. The aerial target detection method based on cross-weighted pixel reconstruction according to claim 2, characterized in that: The training samples consist of a training set and a test set, the training set is used to train the constructed aerial target detection network model, and the test set is used to evaluate the performance of the aerial target detection network model after training.

4. The aerial target detection method based on cross-weighted pixel reconstruction according to claim 1, characterized in that: The CPR module includes a global average pooling unit, a height average pooling unit, a width average pooling unit, three BConv units, three Sigmoid function units, and four Hadamard product units. The input feature map is input into the input ends of the global average pooling unit, the height average pooling unit, and the width average pooling unit respectively. The output end of the global average pooling unit is connected to the input end of the first Sigmoid function unit via the first BConv unit, the output end of the height average pooling unit is connected to the first input end of the first Hadamard product unit via the second BConv unit, and the output end of the width average pooling unit is connected to the first input end of the second Hadamard product unit via the third BConv unit. , the two output ends of the first Sigmoid function unit are respectively connected to the second input end of the first Hadamard product unit and the second input end of the second Hadamard product unit, the output end of the first Hadamard product unit is connected to the first input end of the third Hadamard product unit via the second Sigmoid function unit, the output end of the second Hadamard product unit is connected to the second input end of the third Hadamard product unit via the third Sigmoid function unit, the output end of the third Hadamard product unit is connected to the first input end of the fourth Hadamard product unit, the second input end of the fourth Hadamard product unit inputs the input feature map, and the output end of the fourth Hadamard product unit outputs the output feature map, wherein: The global average pooling unit is used to aggregate channel-wise features in the input feature map; The height average pooling unit is used to aggregate the height-direction features in the input feature map; The width average pooling unit is used to aggregate the features in the width direction of the input feature map; Three BConv units are used to learn parameters; The three Sigmoid function units are used to calculate the feature weights in the channel, height, and width directions respectively; The first Hadamard product unit is used to apply the feature weight of the channel direction to the feature map in the height direction; The second Hadamard product unit is used to apply the feature weights in the channel direction to the feature map in the width direction; The third Hadamard product unit is used to cross the feature weights in the height and width directions to obtain cross weights in multiple dimensions; The fourth Hadamard product unit is used to apply the cross weights to the input feature map to obtain the output feature map.

5. The aerial target detection method based on cross-weighted pixel reconstruction according to claim 4, characterized in that: The calculation formula of the feature weight in the channel direction is: ω ch =σ(conv 1×1 (ReLU(conv 1×1 (GAP(F))))); The calculation formula of the feature weight in the height direction is: oh height =σ(F height ⊙expand(ω ch )) The calculation formula for the feature weight in the width direction is: oh width =σ(F width ⊙expand(ω ch )) Here, expand(·) represents the dimension expansion operation, and ⊙ represents the Hadamard product.

6. The aerial target detection method based on cross-weighted pixel reconstruction according to claim 1, characterized in that: The FS-PANet network module includes two AFS modules, three C2f modules, a convolution module, two Concat modules, an SCDown downsampling module and a C2fCIB module. The two input ends of the first AFS module are respectively input with the high-level feature map and the middle-level feature map extracted from the backbone of the Yolov10 network. The output end of the first AFS module is connected to the input end of the first C2f module. The two output ends of the first C2f module are respectively connected to an input end of the second AFS module and an input end of the first Concat module. The other input end of the second AFS module is input with the low-level feature map extracted from the backbone of the Yolov10 network. The output end of the FS module is connected to the input end of the second C2f module, one output end of the second C2f module is connected to the other input end of the first Concat module via the convolution module, the output end of the first Concat module is connected to one input end of the second Concat module via the third C2f module and the SCDown downsampling module, the other input end of the second Concat module inputs the high-level feature map, the output end of the second Concat module is connected to the input end of the C2fCIB module, and the other output end of the second C2f module, the other output end of the third C2f module, and the output end of the C2fCIB module are all connected to the head part of the Yolov10 network.

7. The aerial target detection method based on cross-weighted pixel reconstruction according to claim 6, characterized in that: The AFS module includes a transposed convolution unit, a CA attention unit, a multiplication unit, and a feature fusion unit. The transposed convolution unit is used to perform transposed convolution on the input higher-level feature map to obtain an intermediate feature map of the same size as the lower-level feature map. The CA attention unit is used to calculate the weight of each feature channel. The multiplication unit is used to multiply the channel weight of the obtained higher-level feature map with the lower-level feature map to enhance the useful features in the lower-level feature map. The feature fusion unit is used to perform feature fusion on the higher-level feature map and the enhanced lower-level feature map and output them.

8. The aerial target detection method based on cross-weighted pixel reconstruction according to claim 7, characterized in that: The AFS module calculates the weight of each feature channel using the following formula: ω=σ(BConv(GAP 1×2H×2W (C' high ))+BConv(GMP 1×2H×2W (C' high ))) Among them, ω represents the weight of each feature channel, σ(·) represents the sigmoid activation function, GAP 1×2H×2W (·), GMP 1 ×2H×2W (·) represent global average pooling and global maximum pooling with pooling size of 1×2H×2W, respectively.

9. The aerial target detection method based on cross-weighted pixel reconstruction according to any one of claims 1 to 8, characterized in that: The loss function of the aerial target detection network model is: Among them, L inner-MPDIoU is the loss function, IoU inner is the inner intersection and union ratio, d1 is the distance between the upper left corner coordinates of the true bounding box and the predicted bounding box, d2 is the distance between the lower right corner coordinates of the true bounding box and the predicted bounding box, and w and h are the width and height of the input feature map respectively.

Images