Aerial image target detection method based on cross-level deformable feature aggregation network

Abstract

An aerial image target detection method based on a cross-level deformable feature aggregation network, characterized in that it comprises the following steps: 1: an image acquisition module acquires aerial image data a in real time; 2: a preprocessing module performs a preprocessing operation on the aerial image data a to obtain standard graphic data b; 3: a feature extraction module in a cross-level deformable feature aggregation network CDNet performs a feature extraction operation on the standard graphic data b to obtain feature data c; 4: a feature fusion module performs a feature fusion operation on the feature data c to obtain fusion data d; 5: an image detection module performs an image detection operation on the fusion data d and outputs a detection result e. Effect: The method can effectively extract high-resolution small object detail features, aggregate cross-level feature information and spatial detail information of a target, realize more accurate detection performance, improve the sensitivity of the network to irregular target boundaries, and enhance the modeling capability of the model.

Description

Technical Field

[0001] This invention relates to the field of image target detection technology, and in particular to an aerial image target detection method based on a cross-level deformable feature aggregation network. Background Technology

[0002] Target detection in drone aerial photography is increasingly becoming a key application in various fields, such as intelligent transportation, drone reconnaissance, remote sensing mapping, environmental monitoring, and search and rescue. With the continuous improvement of hardware and computing power, deep learning-based target detection methods have made significant progress and breakthroughs in recent years, and are widely used in various fields. Based on their detection processes, target detection methods can be divided into two main categories: two-stage detectors and single-stage detectors. Two-stage detectors first generate candidate regions containing targets in the image to be detected; then, they use convolutional neural networks to regress and classify these candidate regions. Representative algorithms include R-CNN, Fast R-CNN, Faster R-CNN, SPPNet, Mask R-CNN, and FPN. In contrast, single-stage detection algorithms directly regress and classify targets in the image. Representative algorithms include YOLOX, DETR, CornerNet, SSD, YOLOv6, FCOS, EfficientNet, and PP-YOLO. Single-stage algorithms, by balancing accuracy and speed, have wide applicability in practical applications.

[0003] However, these excellent detection methods cannot be well transferred to drone aerial photography scenarios, as the specific characteristics of drone aerial data limit the performance of these detectors. From a practical application perspective, drone aerial images have the following characteristics:

[0004] 1) The drone's view contains a large number of densely packed, tiny objects with low resolution and limited features;

[0005] 2) Due to the diversity of drone aerial photography perspectives and altitudes, objects of the same category exhibit significant scale variations and diverse shape boundaries in different scenarios, and this difference is further amplified for different object categories. This poses a significant challenge to target detection in aerial photography scenarios.

[0006] Some scholars have used methods such as high-resolution refined modeling to improve the detection performance of small objects in aerial photography, but these methods suffer from drawbacks such as complex models and a large number of parameters. To handle extremely small object instances, Huang et al. proposed merging sub-regions clustered from the output of a coarse detector to suppress the background and performing separate inference on the sub-regions, significantly improving detection accuracy. Lu et al. used CSWin Transformer as the backbone to build a hybrid patch embedding module, achieving multi-scale object feature extraction and utilizing low-level information such as image edges and corners. A paper introduces a Semantic Embedding Density Adaptive Network (SDANet), which first performs pixel-level prediction to generate object clustering proposals and centers, then performs clustering partitioning and recursive center matching to obtain each object. Yin et al. proposed SDPDnet, which generates a dynamic number of proposal boxes on feature maps at different scales through scale-separated activation pyramids to improve learning for various object distributions and scales. While these methods enhance small object recognition capabilities to some extent, the large parameter overhead makes the networks unsuitable for real-time lightweight inference and detection on UAV platforms.

[0007] Although object detectors derived from the YOLO algorithm have demonstrated excellent detection speed and overall accuracy in recent years, their ability to accurately capture the complex features of small and irregularly shaped objects deserves further improvement.

[0008] Zeng designed a multi-level feature fusion structure and a hybrid coordinate attention module to distinguish foreground and background features, enriching shallow feature semantics and enhancing the ability to extract features from small objects. Ye et al. constructed a lightweight feature extraction module to control computational cost and parameters, and used convolutional projection instead of linear positional projection to reduce computational overhead and achieve lightweight real-time detection of UAV aerial images. Song designed a novel multi-scale hybrid attention (MHA) structure that can extract multi-dimensional channel information and capture similarity relationships between objects of different sizes to enhance the feature representation of small objects. Xu introduced DSAA-YOLO, which employs a dense residual super-resolution module and an information-aligned feature enhancement module to effectively improve the resolution of small objects whose quality has been degraded due to transformation, and maximizes the extraction of features from the original image. However, these methods cannot effectively adapt to irregular objects with large scale variations, and their dense small object detection performance remains limited.

[0009] Limitations of existing technologies: Existing image target detection methods are not suitable for real-time lightweight inference and detection on UAV platforms, cannot accurately capture the complex features of small and irregularly shaped objects, and cannot accurately detect dense small targets as well as significant scale and shape changes between targets. Summary of the Invention

[0010] This invention provides an aerial image target detection method based on a cross-level deformable feature aggregation network, which can achieve accurate detection of aerial image targets.

[0011] To achieve the above objectives, this invention provides an aerial image target detection method based on a cross-level deformable feature aggregation network, the key of which includes the following steps:

[0012] Step 1: The image acquisition module acquires aerial image data a in real time and transmits the aerial image data a to the preprocessing module;

[0013] Step 2: The preprocessing module performs preprocessing operations on the aerial image data a to obtain standard graphic data b, and then passes it to the cross-level deformable feature aggregation network CDNet;

[0014] Step 3: The feature extraction module in the cross-level deformable feature aggregation network CDNet performs feature extraction on the standard graphic data b to obtain feature data c, and then passes it to the feature fusion module;

[0015] Step 4: The feature fusion module performs feature fusion operation on the feature data c to obtain fused data d, and then transmits it to the image detection module;

[0016] Step 5: The image detection module performs image detection on the fused data d and outputs the detection result e.

[0017] In step 2, the preprocessing operation is to modify the image so that it meets the input requirements of the cross-level deformable feature aggregation network CDNet.

[0018] Through the above design, the feature extraction module efficiently extracts rich feature information from standard graphic data b, and the extracted feature data information is aggregated into high-resolution shallow feature maps and reorganized into multi-level feature information through the feature fusion module, and the deep feature extraction structure is optimized to improve the positioning accuracy of small targets.

[0019] Since targets in aerial photography are almost always small to medium in size, information about small targets is severely lost in deep feature maps, resulting in limited contribution to the detection of tiny targets. To enhance the model's efficient use of shallow, high-resolution feature maps, the 20×20 deep 32x downsampled feature map and its corresponding prediction head were eliminated in the feature fusion and image detection modules. This mitigates the problem of low accuracy in detecting small to medium-sized targets due to the model's overemphasis on easily detectable large targets, and also reduces the number of model parameters.

[0020] Preferably, in step 3, the feature extraction module performs feature extraction on the standard graphic data b, as follows:

[0021] Step A1: The first downsampling module in the feature extraction module acquires the standard graphic data b, performs a downsampling operation on the standard graphic data b to obtain the first downsampling data b1, and passes it to the second downsampling module;

[0022] Step A2: The second downsampling module performs a downsampling operation on the first downsampling data b1 to obtain the second downsampling data b2, and then passes it to the first image fusion module and the feature fusion module;

[0023] Step A3: The first graph fusion module performs graph fusion operation on the second downsampled data b2 to obtain the first graph fusion data b3, and then passes it to the third downsampled module and the feature fusion module;

[0024] Step A4: The third downsampling module performs a downsampling operation on the first image fusion data b3 to obtain the third downsampling data b4, and then transmits it to the second image fusion module and the feature fusion module;

[0025] Step A5: The second graph fusion module performs graph fusion operation on the third downsampled data b4 to obtain the second graph fused data b5, and then passes it to the fourth downsampled module and the feature fusion module;

[0026] Step A6: The fourth downsampling module performs a downsampling operation on the second image fusion data b5 to obtain the fourth downsampling data b6, and then transmits it to the third image fusion module and the feature fusion module;

[0027] Step A7: The third graph fusion module performs graph fusion operation on the fourth downsampled data b6 to obtain the third graph fusion data b7, and then passes it to the fifth downsampled module and the feature fusion module;

[0028] Step A8: The fifth downsampling module performs a downsampling operation on the third image fusion data b7 to obtain the fifth downsampling data b8, and then transmits it to the fourth image fusion module;

[0029] Step A9: The fourth graph fusion module performs graph fusion operation on the fifth downsampled data b8 to obtain the fourth graph fusion data b9, and then passes it to the spatial pyramid pooling module;

[0030] Step A10: The spatial pyramid pooling module performs multi-scale pooling operation on the fourth graph fusion data b9 to obtain feature data c, and then passes it to the feature fusion module.

[0031] Through the above design, the downsampling module is used to perform convolutional downsampling on the image, reducing the resolution while extracting features and semantic information. The graph fusion module is used to effectively fuse feature maps from different levels, enabling the model to obtain more contextual and high-resolution information. The spatial pyramid pooling module can not only pool feature maps at different scales to extract richer feature information and adapt to targets of different sizes, but also increase the model's receptive field and enhance its target detection capability without increasing computational cost.

[0032] Preferably, the first image fusion module, the second image fusion module, the third image fusion module, and the fourth image fusion module have the same structure, and each is provided with an eighth downsampling module, a first Split segmentation function layer, a bottleneck structure, a seventh channel splicing layer, and a ninth downsampling module connected in sequence; the bottleneck structure is composed of K bottleneck layers with the same structure connected in sequence, where K≥2;

[0033] The output of each bottleneck layer is connected to the input of the seventh channel splicing layer; each bottleneck layer is provided with a sixth convolutional layer, a seventh convolutional layer and a residual connection layer in sequence;

[0034] Let F be the input data of the i-th graph fusion module. The input data F can be either the second downsampled data b2, the third downsampled data b4, the fourth downsampled data b6, or the fifth downsampled data b8. Then, the i-th graph fusion module performs graph fusion operation on the input data F, and the steps are as follows:

[0035] Step D1: The eighth downsampling module performs a downsampling operation on the input data F to obtain the eighth downsampled data F1, and passes it to the first Split segmentation function layer;

[0036] Step D2: The first Split function layer divides the eighth downsampled data F1 into first segmented data F2 and second segmented data F3. The first Split function layer passes the first segmented data F2 to the bottleneck structure and the seventh channel splicing layer, and passes the second segmented data F3 to the seventh channel splicing layer.

[0037] Step D3: The first bottleneck layer in the bottleneck structure performs feature extraction and compression on the first segmented data F2 to obtain the first bottleneck data K1, and then passes it to the second bottleneck layer and the seventh channel splicing layer.

[0038] The j-th bottleneck layer corresponds to the (j-1)-th bottleneck data K. j-1 Perform feature extraction and compression operations to obtain the j-th bottleneck data K. j This is then passed to the (j+1)th bottleneck layer and the seventh channel splicing layer; where j is a positive integer less than K.

[0039] The Kth bottleneck layer corresponds to the (K-1)th bottleneck data K. K-1 Perform feature extraction and compression operations to obtain the Kth bottleneck data K. K The output data of the K bottleneck layers is then passed to the seventh channel splicing layer; the output data of the K bottleneck layers together form the bottleneck output data F4 of the bottleneck structure.

[0040] Step D4: The seventh channel stitching layer performs a channel-dimensional stitching operation on the first segmented data F2, the second segmented data F3, and the bottleneck output data F4 to obtain the seventh stitched data F5, which is then passed to the ninth downsampling module.

[0041] Step D5: The ninth downsampling module performs a downsampling operation on the seventh stitched data F5 to obtain the fused data of the i-th image. 1≤i≤4.

[0042] The above design, by stitching together the output data of K bottleneck layers along the channel dimension through the channel stitching layer, can increase the information interaction between different channels and achieve richer feature expression.

[0043] Preferably, in step A10, the spatial pyramid pooling module includes a tenth downsampling module, an eleventh downsampling module, a first max pooling layer, a second max pooling layer, a third max pooling layer, and an eighth channel stitching layer. The spatial pyramid pooling module performs multi-scale pooling operations on the fourth image fused data b9, as follows:

[0044] Step B1: The tenth downsampling module acquires the fourth image fusion data b9 and performs downsampling operation to obtain the tenth downsampling data c1, which is then passed to the first max pooling layer and the eighth channel splicing layer.

[0045] Step B2: The first max pooling layer performs max pooling on the tenth downsampled data c1 to obtain the first max pooled data c2, and then passes it to the second max pooling layer and the eighth channel splicing layer.

[0046] Step B3: The second max pooling layer performs max pooling on the first max pooling data c2 to obtain the second max pooling data c3, and then passes it to the third max pooling layer and the eighth channel splicing layer.

[0047] Step B4: The third max pooling layer performs max pooling on the second max pooling data c3 to obtain the third max pooling data c4, and then passes it to the eighth channel splicing layer.

[0048] Step B5: The eighth channel splicing layer performs a channel-dimensional splicing operation on the tenth downsampled data c1, the first max pooled data c2, the second max pooled data c3, and the third max pooled data c4 to obtain the eighth spliced ​​data c5, which is then passed to the eleventh downsampled module.

[0049] Step B6: The eleventh downsampling module performs a downsampling operation on the eighth spliced ​​data c5 to obtain feature data c.

[0050] The above design uses multiple max pooling layers to pool data at different scales, thus adapting to targets of different sizes.

[0051] Preferably, the first downsampling module, the second downsampling module, the third downsampling module, the fourth downsampling module, the fifth downsampling module, the sixth downsampling module, the seventh downsampling module, the eighth downsampling module, the ninth downsampling module, the tenth downsampling module, the eleventh downsampling module, the twelfth downsampling module, the thirteenth downsampling module, and the fourteenth downsampling module have the same structure, and each is provided with a first two-dimensional convolutional layer, a batch normalization layer, and a SiLU activation function layer connected in sequence.

[0052] The downsampling module performs convolutional downsampling on the image, reducing the resolution while extracting features and semantic information.

[0053] Preferably, in step 4, the feature fusion module performs a feature fusion operation on the feature data c, as follows:

[0054] Step C1: The first convolutional layer in the feature fusion module obtains the first image fusion data b3, performs a convolution operation on the first image fusion data b3 to obtain the first convolutional data d1, and passes it to the third channel stitching layer and the fourth channel stitching layer;

[0055] The second convolutional layer obtains the second image fusion data b5, performs a convolution operation on the second image fusion data b5 to obtain the second convolutional data d2, and passes it to the second channel stitching layer and the fifth channel stitching layer;

[0056] The third convolutional layer obtains the third image fusion data b7, performs a convolution operation on the third image fusion data b7 to obtain the third convolutional data d3, and passes it to the first channel stitching layer and the sixth channel stitching layer;

[0057] The fourth convolutional layer acquires feature data c and performs a convolution operation on the feature data c to obtain the fourth convolutional data d4, which is then passed to the first upsampling module.

[0058] Step C2: The first upsampling module performs an upsampling operation on the fourth convolutional data d4 to obtain the first upsampled data d5, and then passes it to the first channel splicing layer;

[0059] Step C3: The first channel stitching layer performs a channel-dimensional stitching operation on the third convolutional data d3 and the first upsampled data d5 to obtain the first stitched data d6, and then passes it to the first deformable attention bottleneck module;

[0060] Step C4: The first deformable attention bottleneck module efficiently extracts the boundary information of the first spliced ​​data d6 to obtain the first deformable attention data d7, and passes it to the second upsampling module and the sixth channel splicing layer;

[0061] Step C5: The second upsampling module upsamples the first deformable attention data d7 to obtain the second upsampled data d8, and then transmits it to the second channel splicing layer;

[0062] Step C6: The second channel splicing layer performs a channel-dimensional splicing operation on the second convolutional data d2 and the second upsampled data d8 to obtain the second spliced ​​data d9, which is then passed to the second deformable attention bottleneck module.

[0063] Step C7: The second deformable attention bottleneck module efficiently extracts the boundary information of the second spliced ​​data d9 to obtain the second deformable attention data d10, and passes it to the third upsampling module and the fifth channel splicing layer;

[0064] Step C8: The third upsampling module performs an upsampling operation on the second deformable attention data d10 to obtain the third upsampled data d11, and transmits it to the third channel splicing layer;

[0065] Step C9: The third channel splicing layer performs a channel-dimensional splicing operation on the first convolutional data d1 and the third upsampled data d11 to obtain the third spliced ​​data d12, and passes it to the third deformable attention bottleneck module.

[0066] Step C10: The third deformable attention bottleneck module efficiently extracts the boundary information of the third splicing data d12 to obtain the third deformable attention data d13, and passes it to the fourth channel splicing layer;

[0067] Step C11: The fourth channel stitching layer performs a channel-dimensional stitching operation on the first convolutional data d1, the second downsampled data b2, and the third deformable attention data d13 to obtain the fourth stitched data d14.

[0068] Step C12: The fourth deformable attention bottleneck module efficiently extracts the boundary information of the fourth spliced ​​data d14 to obtain the fourth deformable attention data d15, and passes it to the fifth convolutional layer;

[0069] Step C13: The fifth convolutional layer performs a convolution operation on the fourth deformable attention data d15 to obtain the fifth convolutional data d16, and passes it to the sixth downsampling module and the image detection module;

[0070] Step C14: The sixth downsampling module performs a downsampling operation on the fifth convolutional data d16 to obtain the sixth downsampled data d17, and passes it to the fifth channel splicing layer;

[0071] Step C15: The fifth channel splicing layer performs a channel-dimensional splicing operation on the second convolutional data d2, the third downsampled data b4, the second deformable attention data d10, and the sixth downsampled data d17 to obtain the fifth spliced ​​data d18, which is then passed to the fifth deformable attention bottleneck module.

[0072] Step C16: The fifth deformable attention bottleneck module efficiently extracts the boundary information of the fifth spliced ​​data d18 to obtain the fifth deformable attention data d19, and passes it to the seventh downsampling module and the image detection module;

[0073] Step C17: The seventh downsampling module performs a downsampling operation on the fifth deformable attention data d19 to obtain the seventh downsampled data d20, and transmits it to the sixth channel splicing layer;

[0074] Step C18: The sixth channel stitching layer performs a channel-dimensional stitching operation on the third convolutional data d3, the fourth downsampled data b6, the first stitched data d6, and the seventh downsampled data d20 to obtain the sixth stitched data d21, which is then passed to the sixth deformable attention bottleneck module.

[0075] Step C19: The sixth deformable attention bottleneck module efficiently extracts the boundary information of the sixth stitched data d21 to obtain fused data d, and transmits it to the image detection module.

[0076] Through the above design, as the network downsamples, the feature map shrinks continuously, and the target size collapses continuously. The shallow high-resolution 160×160 image has more detailed information about the location of small targets, realizing feature extraction of the detailed information of small targets in the shallow layer, enhancing the localization ability of small targets, and removing the low-resolution 20×20 feature extraction structure in the deep layer that does not contribute much to the localization of small targets, reducing the number of redundant parameters in the network.

[0077] Furthermore, by designing a cross-level spindle-shaped feature fusion structure, the spatial information representation of dense small targets is enriched, and cross-level multi-scale feature information is effectively aggregated to achieve more comprehensive semantic and spatial information integration, thereby improving the detection accuracy of multi-scale and small targets in aerial photography scenarios.

[0078] Preferably, the first deformable attention bottleneck module, the second deformable attention bottleneck module, the third deformable attention bottleneck module, the fourth deformable attention bottleneck module, the fifth deformable attention bottleneck module and the sixth deformable attention bottleneck module have the same structure, and each is provided with a CA attention mechanism module, a twelfth downsampling module, a second split function layer, a deformable convolutional structure, a ninth channel splicing layer and a thirteenth downsampling module.

[0079] The deformable convolutional structure is composed of m deformable convolutional blocks with identical structures connected sequentially, where m ≥ 2; the output of each deformable convolutional block is connected to the input of the ninth channel splicing layer.

[0080] Each of the deformable convolutional blocks is sequentially provided with a fourteenth downsampling module and a deformable convolutional layer.

[0081] Through the above design, the deformable convolutional layer can select a central location on the feature map and then sample outwards. For each sampling location, a learnable parameter offset is added, enabling the convolutional layer to perform a translation transformation. Furthermore, to avoid introducing useless regional information into the deformable convolution calculation and interfering with feature extraction, a weight coefficient Δm is introduced. k By assigning smaller coefficients to sampling points in regions of less interest, the sampling regions better conform to the shape of irregular targets, thus improving the network's modeling ability. The calculation method for deformable convolutional layers is as follows:

[0082]

[0083] R={(-1,-1),(-1,0),…,(0,1),(1,1)}

[0084] Where p0 is the current pixel, n represents the total number of sampling positions, and R represents the predetermined sampling position region; p n Δp represents the nth predefined network sampling position. n It is the offset corresponding to the nth sampling position, Δm k Represents the weighting coefficient, Δm k ∈(0,1); x represents the input feature map, and y(p0) represents the value of the output feature map at position p0. ω(p n ) is p n The learnable weights corresponding to the positions.

[0085] The CA attention mechanism module adopts the CA (Coordinate attention) attention mechanism, which takes into account both channel and spatial information, and can capture remote dependencies while retaining accurate location information.

[0086] The deformable attention bottleneck module has good adaptability to position changes. The embedded CA attention mechanism makes the network pay more attention to the information of the target itself and suppress the interference information of irrelevant areas. The deformable convolution makes the network more closely resemble the shape of irregular targets in aerial photography scenes during sampling, and realizes efficient extraction of boundary information of various deformable objects.

[0087] Preferably, in step 5, the image detection module is provided with a first detection head, a second detection head, and a third detection head;

[0088] The first detection head acquires the fifth convolutional data d12 and performs a detection operation on it to obtain the first detection data e1; the second detection head acquires the fifth deformable attentional data d14 and performs a detection operation on it to obtain the second detection data e2; the third detection head acquires the fused data d and performs a detection operation on it to obtain the third detection data e3; the image detection module merges the first detection data e1, the second detection data e2, and the third detection data e3 and outputs the detection result e.

[0089] By setting up three detection heads to detect images of different resolutions and point targets of different sizes, all detection results are finally merged together. The final output, 'e', ​​is a detected image labeled with the detected targets, their categories, and locations.

[0090] Preferably, the first detection head, the second detection head, and the third detection head have the same structure, and each is provided with an eighth convolutional layer, a ninth convolutional layer, a tenth convolutional layer, an eleventh convolutional layer, a second two-dimensional convolutional layer, a third two-dimensional convolutional layer, a bounding box loss function layer, and a binary classification cross-entropy loss function layer.

[0091] The input of the eighth convolutional layer is connected to the feature fusion module, and the output is connected to the bounding box loss function layer via the ninth convolutional layer and the second two-dimensional convolutional layer in sequence; the input of the tenth convolutional layer is connected to the feature fusion module, and the output is connected to the binary classification cross-entropy loss function layer via the eleventh convolutional layer and the third two-dimensional convolutional layer in sequence.

[0092] The above design allows for better calculation of the bounding box loss function value and the binary cross-entropy loss function value.

[0093] Preferably, the bounding box loss function layer applies the WIoU loss function to the Inner-IoU loss function based on the auxiliary bounding box, resulting in the Inner-WIoU bounding box loss function. The calculation formula for the Inner-IoU loss function based on the auxiliary bounding box is as follows:

[0094]

[0095]

[0096]

[0097]

[0098]

[0099] uniob=(w*h)*(br) 2 +(w gt *h gt )*(br) 2 -inter

[0100]

[0101] Among them, (x gt ,y gt (x, y) represents the center coordinates of the ground truth bounding box, and (x, y) represents the center coordinates of the predicted bounding box. gt h represents the actual frame width. gt `w` represents the height of the true bounding box, `h` represents the width of the predicted bounding box, and `br` represents the scaling factor, ranging from [0.5, 1.5]; `L` represents the height of the predicted bounding box. Inner-IoU This represents the IoU loss function value based on the auxiliary bounding box; (x r ,y u () indicates the coordinates of the top-left corner of the prediction box. This represents the coordinates of the top-left corner of the true bounding box, (x l ,y d The ) indicates the coordinates of the bottom right corner of the prediction box. This represents the coordinates of the bottom right corner of the actual bounding box;

[0102] The formula for calculating the bounding box loss function Inner-WIoU is as follows:

[0103] L IoU =1-IoU

[0104]

[0105]

[0106]

[0107] in, Indicates that L is currently IoU The mean of S, where IoU represents the intersection-union ratio of the predicted bounding box and the ground truth bounding box. WIoU Used to measure the center distance between the predicted bounding box and the ground truth bounding box; exp represents the natural exponential function, i.e., e raised to the power of n; W g H represents the width of the smallest bounding rectangle between the predicted bounding box and the ground truth bounding box. g This represents the height of the smallest bounding rectangle between the predicted bounding box and the ground truth bounding box; The superscript * in W indicates that W g and H g Separate from the computation graph;

[0108] L Inner-WIoU This represents the bounding box loss function value. α represents the monotonic focus factor; β represents the outlier of the anchor box, and the higher the outlier, the worse the quality of the predicted box; r represents the non-monotonic frequency modulation factor, which is controlled by the outlier β and α; α and δ represent the model parameters, with α taking the value of 1.9 and δ taking the value of 3.

[0109] Based on the above design, an auxiliary bounding box dynamic non-monotonic gradient gain loss function, namely the Inner-WIoU bounding box loss function, was designed. This Inner-WIoU bounding box loss function designs an auxiliary box proportionally to the target bounding box. The auxiliary box of the predicted box is used to fit the auxiliary box of the ground truth bounding box, and the fitting process employs dynamic non-monotonic frequency tuning. The design of the auxiliary box can adjust the weight of the target penalty coefficient, accelerating network convergence, while dynamic non-monotonic frequency tuning can increase the network's attention to ordinary samples, improving the model's generalization performance.

[0110] The bounding box loss function Inner-WIoU consists of Inner-IoU and WIoU. Inner-IoU designs auxiliary bounding boxes with equal scale for the ground truth box and the predicted box, and uses the scaled bounding boxes to calculate IOU, thereby improving the training gain of the network for high IOU samples.

[0111] The proposed bounding box loss function, Inner-WIoU, can effectively suppress the competitiveness of high-quality anchor boxes, reduce the harmful gradients contributed by low-quality samples, and enhance the contribution weight of target samples by adjusting the scaling factor, thereby accelerating network convergence and improving model detection performance.

[0112] The beneficial effects of this invention are:

[0113] 1. It can effectively extract detailed features of high-resolution small objects, aggregate cross-level feature information and spatial detail information of the target, and achieve more accurate detection performance.

[0114] 2. By designing a deformable attention bottleneck module, the network's sensitivity to irregular target boundaries is improved, effectively enhancing the model's modeling ability and bringing benefits to target detection tasks under the dynamic perspective of UAVs.

[0115] 3. By designing the bounding box loss function Inner-WIoU, harmful gradients contributed by extreme samples are effectively avoided, focusing on ordinary samples and improving detection accuracy and network generalization performance.

[0116] 4. With fewer parameters, it can better adapt to real-time target detection tasks in dynamic perspective scenarios of real drones. Attached Figure Description

[0117] Figure 1 This is a schematic diagram of the process of the present invention;

[0118] Figure 2 This is a schematic diagram of the structure of CDNet, the cross-level deformable feature aggregation network in this invention;

[0119] Figure 3 This is a schematic diagram of the structure of the downsampling module CBS and the graph fusion module in the embodiment;

[0120] Figure 4 This is a schematic diagram of the spatial pyramid pooling module in the embodiment;

[0121] Figure 5 This is a schematic diagram of the detection head in the embodiment;

[0122] Figure 6 This is a structural schematic diagram of the deformable bottleneck module in the embodiment;

[0123] Figure 7 This is a schematic diagram of the Inner-WIoU bounding box loss function in the embodiment;

[0124] Figure 8 This is a simplified structural diagram of the cross-level deformable feature aggregation network CDNet in the embodiment. Detailed Implementation

[0125] The present invention will be further described in detail below with reference to the accompanying drawings and specific examples. The following embodiments or drawings are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0126] like Figure 1 As shown: An aerial image target detection method based on a cross-level deformable feature aggregation network, the key of which includes the following steps:

[0127] Step 1: The image acquisition module acquires aerial image data a in real time and transmits the aerial image data a to the preprocessing module;

[0128] Step 2: The preprocessing module performs preprocessing operations on the aerial image data a to obtain standard graphic data b, and then passes it to the cross-level deformable feature aggregation network CDNet;

[0129] Step 3: The feature extraction module in the cross-level deformable feature aggregation network CDNet performs feature extraction on the standard graphic data b to obtain feature data c, and then passes it to the feature fusion module;

[0130] Step 4: The feature fusion module performs feature fusion operation on the feature data c to obtain fused data d, and then transmits it to the image detection module;

[0131] Step 5: The image detection module performs image detection on the fused data d and outputs the detection result e.

[0132] like Figure 2 As shown: In step 3, the feature extraction module performs feature extraction on the standard graphic data b, and the steps are as follows:

[0133] Step A1: The first downsampling module in the feature extraction module acquires the standard graphic data b, performs a downsampling operation on the standard graphic data b to obtain the first downsampling data b1, and passes it to the second downsampling module;

[0134] Step A2: The second downsampling module performs a downsampling operation on the first downsampling data b1 to obtain the second downsampling data b2, and then passes it to the first image fusion module and the feature fusion module;

[0135] Step A3: The first graph fusion module performs graph fusion operation on the second downsampled data b2 to obtain the first graph fusion data b3, and then passes it to the third downsampled module and the feature fusion module;

[0136] Step A4: The third downsampling module performs a downsampling operation on the first image fusion data b3 to obtain the third downsampling data b4, and then transmits it to the second image fusion module and the feature fusion module;

[0137] Step A5: The second graph fusion module performs graph fusion operation on the third downsampled data b4 to obtain the second graph fused data b5, and then passes it to the fourth downsampled module and the feature fusion module;

[0138] Step A6: The fourth downsampling module performs a downsampling operation on the second image fusion data b5 to obtain the fourth downsampling data b6, and then transmits it to the third image fusion module and the feature fusion module;

[0139] Step A7: The third graph fusion module performs graph fusion operation on the fourth downsampled data b6 to obtain the third graph fusion data b7, and then passes it to the fifth downsampled module and the feature fusion module;

[0140] Step A8: The fifth downsampling module performs a downsampling operation on the third image fusion data b7 to obtain the fifth downsampling data b8, and then transmits it to the fourth image fusion module;

[0141] Step A9: The fourth graph fusion module performs graph fusion operation on the fifth downsampled data b8 to obtain the fourth graph fusion data b9, and then passes it to the spatial pyramid pooling module;

[0142] Step A10: The spatial pyramid pooling module performs multi-scale pooling operation on the fourth graph fusion data b9 to obtain feature data c, and then passes it to the feature fusion module.

[0143] In step 4, the feature fusion module performs a feature fusion operation on the feature data c, as follows:

[0144] Step C1: The first convolutional layer in the feature fusion module obtains the first image fusion data b3, performs a convolution operation on the first image fusion data b3 to obtain the first convolutional data d1, and passes it to the third channel stitching layer and the fourth channel stitching layer;

[0145] The second convolutional layer obtains the second image fusion data b5, performs a convolution operation on the second image fusion data b5 to obtain the second convolutional data d2, and passes it to the second channel stitching layer and the fifth channel stitching layer;

[0146] The third convolutional layer obtains the third image fusion data b7, performs a convolution operation on the third image fusion data b7 to obtain the third convolutional data d3, and passes it to the first channel stitching layer and the sixth channel stitching layer;

[0147] The fourth convolutional layer acquires feature data c and performs a convolution operation on the feature data c to obtain the fourth convolutional data d4, which is then passed to the first upsampling module.

[0148] Step C2: The first upsampling module performs an upsampling operation on the fourth convolutional data d4 to obtain the first upsampled data d5, and then passes it to the first channel splicing layer;

[0149] Step C3: The first channel stitching layer performs a channel-dimensional stitching operation on the third convolutional data d3 and the first upsampled data d5 to obtain the first stitched data d6, and then passes it to the first deformable attention bottleneck module;

[0150] Step C4: The first deformable attention bottleneck module efficiently extracts the boundary information of the first spliced ​​data d6 to obtain the first deformable attention data d7, and passes it to the second upsampling module and the sixth channel splicing layer;

[0151] Step C5: The second upsampling module upsamples the first deformable attention data d7 to obtain the second upsampled data d8, and then transmits it to the second channel splicing layer;

[0152] Step C6: The second channel splicing layer performs a channel-dimensional splicing operation on the second convolutional data d2 and the second upsampled data d8 to obtain the second spliced ​​data d9, which is then passed to the second deformable attention bottleneck module.

[0153] Step C7: The second deformable attention bottleneck module efficiently extracts the boundary information of the second spliced ​​data d9 to obtain the second deformable attention data d10, and passes it to the third upsampling module and the fifth channel splicing layer;

[0154] Step C8: The third upsampling module performs an upsampling operation on the second deformable attention data d10 to obtain the third upsampled data d11, and transmits it to the third channel splicing layer;

[0155] Step C9: The third channel splicing layer performs a channel-dimensional splicing operation on the first convolutional data d1 and the third upsampled data d11 to obtain the third spliced ​​data d12, and passes it to the third deformable attention bottleneck module.

[0156] Step C10: The third deformable attention bottleneck module efficiently extracts the boundary information of the third splicing data d12 to obtain the third deformable attention data d13, and passes it to the fourth channel splicing layer;

[0157] Step C11: The fourth channel stitching layer performs a channel-dimensional stitching operation on the first convolutional data d1, the second downsampled data b2, and the third deformable attention data d13 to obtain the fourth stitched data d14.

[0158] Step C12: The fourth deformable attention bottleneck module efficiently extracts the boundary information of the fourth spliced ​​data d14 to obtain the fourth deformable attention data d15, and passes it to the fifth convolutional layer;

[0159] Step C13: The fifth convolutional layer performs a convolution operation on the fourth deformable attention data d15 to obtain the fifth convolutional data d16, and passes it to the sixth downsampling module and the image detection module;

[0160] Step C14: The sixth downsampling module performs a downsampling operation on the fifth convolutional data d16 to obtain the sixth downsampled data d17, and passes it to the fifth channel splicing layer;

[0161] Step C15: The fifth channel splicing layer performs a channel-dimensional splicing operation on the second convolutional data d2, the third downsampled data b4, the second deformable attention data d10, and the sixth downsampled data d17 to obtain the fifth spliced ​​data d18, which is then passed to the fifth deformable attention bottleneck module.

[0162] Step C16: The fifth deformable attention bottleneck module efficiently extracts the boundary information of the fifth spliced ​​data d18 to obtain the fifth deformable attention data d19, and passes it to the seventh downsampling module and the image detection module;

[0163] Step C17: The seventh downsampling module performs a downsampling operation on the fifth deformable attention data d19 to obtain the seventh downsampled data d20, and transmits it to the sixth channel splicing layer;

[0164] Step C18: The sixth channel stitching layer performs a channel-dimensional stitching operation on the third convolutional data d3, the fourth downsampled data b6, the first stitched data d6, and the seventh downsampled data d20 to obtain the sixth stitched data d21, which is then passed to the sixth deformable attention bottleneck module.

[0165] Step C19: The sixth deformable attention bottleneck module efficiently extracts the boundary information of the sixth stitched data d21 to obtain fused data d, and transmits it to the image detection module.

[0166] In step 5, the image detection module is equipped with a first detection head, a second detection head, and a third detection head;

[0167] The first detection head acquires the fifth convolutional data d12 and performs a detection operation on it to obtain the first detection data e1; the second detection head acquires the fifth deformable attentional data d14 and performs a detection operation on it to obtain the second detection data e2; the third detection head acquires the fused data d and performs a detection operation on it to obtain the third detection data e3; the image detection module merges the first detection data e1, the second detection data e2, and the third detection data e3 and outputs the detection result e.

[0168] like Figure 8As shown in the figure: P2, F2, E2, and T2 represent the cross-level channel aggregation structure for low-resolution feature maps, P3, F3, and T3 represent the cross-level channel aggregation structure for medium-resolution feature maps, P4, F4, and T4 represent the cross-level channel aggregation structure for high-resolution feature maps, and P5 and F5 are used to remove the low-resolution 20×20 feature extraction structure in the deep layer that does not contribute much to the localization of small targets, thereby reducing the number of redundant parameters in the network.

[0169] In the cross-level deformable feature aggregation network CDNet, channel expansion is performed on high-resolution feature maps to enrich feature information and enhance the representation of small target features; channel compression is performed on low-resolution feature maps to enhance information interaction between different channels. A 1×1 convolutional layer is used for channel expansion and compression to save computation. To further improve the representation of target feature information, cross-level feature fusion is performed on the multi-level feature maps after channel enhancement through multiple channel concatenation layers. Feature maps of the same size in the network are concatenated across levels in the channel direction to form the final feature map to be detected. Cross-level fusion of feature maps of different levels and different numbers of channels helps to highlight the target details in the feature maps and improves the overall detection accuracy.

[0170] Specifically, for shallow feature fusion, a separate spindle-shaped skip feature fusion module is used. First, the channels are expanded to highlight the boundary information of small targets. Then, the feature maps spliced ​​across layers are subjected to feature enhancement processing, namely E2 and T2. Finally, at the output, 1×1 convolution is used for channel compression, forming a spindle-shaped structure that is large at both ends and small in the middle. This high-resolution spindle-shaped skip feature fusion structure can greatly improve the network's ability to utilize shallow feature information and enhance its ability to locate small targets.

[0171] like Figure 3 As shown: The first downsampling module, the second downsampling module, the third downsampling module, the fourth downsampling module, the fifth downsampling module, the sixth downsampling module, the seventh downsampling module, the eighth downsampling module, the ninth downsampling module, the tenth downsampling module, the eleventh downsampling module, the twelfth downsampling module, the thirteenth downsampling module, and the fourteenth downsampling module have the same structure, and each is provided with a first two-dimensional convolutional layer, a batch normalization layer, and a SiLU activation function layer connected in sequence.

[0172] The first image fusion module, the second image fusion module, the third image fusion module, and the fourth image fusion module have the same structure. They are all provided with an eighth downsampling module, a first Split segmentation function layer, a bottleneck structure, a seventh channel splicing layer, and a ninth downsampling module connected in sequence. The bottleneck structure is composed of K bottleneck layers with the same structure connected in sequence.

[0173] The output of each bottleneck layer is connected to the input of the seventh channel splicing layer; each bottleneck layer is provided with a sixth convolutional layer, a seventh convolutional layer and a residual connection layer in sequence.

[0174] like Figure 4 As shown: The spatial pyramid pooling module is configured with a tenth downsampling module, an eleventh downsampling module, a first max pooling layer, a second max pooling layer, a third max pooling layer, and an eighth channel splicing layer; the output of the tenth downsampling module is connected to the input of the first max pooling layer and the eighth channel splicing layer, the output of the first max pooling layer is connected to the input of the second max pooling layer and the eighth channel splicing layer, the output of the second max pooling layer is connected to the input of the third max pooling layer and the eighth channel splicing layer, the output of the third max pooling layer is connected to the input of the eighth channel splicing layer, and the output of the eighth channel splicing layer is connected to the input of the eleventh downsampling module.

[0175] like Figure 6 As shown: The first deformable attention bottleneck module, the second deformable attention bottleneck module, the third deformable attention bottleneck module, the fourth deformable attention bottleneck module, the fifth deformable attention bottleneck module and the sixth deformable attention bottleneck module have the same structure. They are all equipped with a CA attention mechanism module, a twelfth downsampling module, a second split function layer, a deformable convolutional structure, a ninth channel splicing layer and a thirteenth downsampling module.

[0176] The deformable convolutional structure is composed of m deformable convolutional blocks with identical structures connected sequentially; the output of each deformable convolutional block is connected to the input of the ninth channel splicing layer.

[0177] Each of the deformable convolutional blocks is sequentially provided with a fourteenth downsampling module and a deformable convolutional layer.

[0178] like Figure 5 As shown: The first detection head, the second detection head, and the third detection head have the same structure, and each is equipped with an eighth convolutional layer, a ninth convolutional layer, a tenth convolutional layer, an eleventh convolutional layer, a second two-dimensional convolutional layer, a third two-dimensional convolutional layer, a bounding box loss function layer, and a binary classification cross-entropy loss function layer;

[0179] The input of the eighth convolutional layer is connected to the feature fusion module, and the output is connected to the bounding box loss function layer via the ninth convolutional layer and the second two-dimensional convolutional layer in sequence; the input of the tenth convolutional layer is connected to the feature fusion module, and the output is connected to the binary classification cross-entropy loss function layer via the eleventh convolutional layer and the third two-dimensional convolutional layer in sequence.

[0180] Figure 7 This is a schematic diagram of the Inner-WIoU bounding box loss function. Figure 7 China B gtB represents the ground truth bounding box, and B represents the predicted bounding box. The bounding box loss function layer applies the WIoU loss function to the Inner-IoU IoU loss function based on the auxiliary bounding box, resulting in the Inner-WIoU bounding box loss function.

[0181] To further verify the detection performance of the cross-level deformable feature aggregation network CDNet in this invention, on the VisDrone dataset, CDNet was compared with popular methods such as RetinaNet, CenterNet, Faster-RCNN, CornerNet, Cascade ADPN, Cascade R-CNN, QuerDet, and the YOLO series, and the mAP was measured. 50 mAP 50:95 The comparisons were made based on recall and the number of parameters. The results are shown in Table 1.

[0182]

[0183] Table 1

[0184] As shown in Table 1, the cross-level deformable feature aggregation network CDNet in this invention achieves the best detection performance with fewer parameters. Compared to the baseline, CDNet achieves a significantly higher mAP with a parameter reduction of 2.02M. 50:95 It increased by 7.1%; compared to yolov5m, CDNet's mAP... 50 It improved performance by 11.8% and reduced the number of parameters by 56.4%; compared with the typical one-stage detection method YOLOv8l, CDNet achieved higher mAP. 50 mAP 50:95 It improved recall by 5.9%, 3.4%, and 4.3% respectively, and had only 20.9% of the number of parameters of YOLOv8l.

[0185] To further validate the effectiveness of the cross-level deformable feature aggregation network CDNet in target detection of UAV aerial images, CDNet was compared with popular methods such as CornerNet, YOLOv5, YOLOv7, YOLOv8, and YOLOv9 on the TinyPerson dataset, achieving higher mAP. 50 mAP 50:95 The comparison results of recall and parameter count are shown in Table 2:

[0186]

[0187] Table 2

[0188] As shown in Table 2, the cross-level deformable feature aggregation network CDNet achieves the best performance on the TinyPerson dataset with a smaller number of parameters. Compared with the baseline, CDNet's mAP is significantly higher. 50 Increased by 3.6%, mAP 50:95 It improved by 1.2%; compared to the latest YOLOv9 network, CDNet achieved better results with only 17.97% of the parameters, and its mAP was also higher. 50 and mAP 50:95 It achieves the highest performance among all the advanced algorithms currently being compared.

[0189] Experimental results show that CDNet can achieve more accurate localization and recognition of small aerial targets with fewer parameters, and has achieved good detection performance in the field of target detection in dynamic viewpoint UAV aerial images.

[0190] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for aerial image target detection based on a cross-level deformable feature aggregation network, characterized in that, Includes the following steps: Step 1: The image acquisition module acquires aerial image data a in real time and transmits the aerial image data a to the preprocessing module; Step 2: The preprocessing module performs preprocessing operations on the aerial image data a to obtain standard graphic data b, and then passes it to the cross-level deformable feature aggregation network CDNet; Step 3: The feature extraction module in the cross-level deformable feature aggregation network CDNet performs feature extraction on the standard graphic data b to obtain feature data c, and then passes it to the feature fusion module; Step 4: The feature fusion module performs feature fusion operation on the feature data c to obtain fused data d, and then transmits it to the image detection module; Step 5: The image detection module performs image detection on the fused data d and outputs the detection result e; In step 4, the feature fusion module performs a feature fusion operation on the feature data c, as follows: Step C1: The first convolutional layer in the feature fusion module obtains the first image fusion data b3, performs a convolution operation on the first image fusion data b3 to obtain the first convolutional data d1, and passes it to the third channel stitching layer and the fourth channel stitching layer; The second convolutional layer obtains the second image fusion data b5, performs a convolution operation on the second image fusion data b5 to obtain the second convolutional data d2, and passes it to the second channel stitching layer and the fifth channel stitching layer; The third convolutional layer obtains the third image fusion data b7, performs a convolution operation on the third image fusion data b7 to obtain the third convolutional data d3, and passes it to the first channel stitching layer and the sixth channel stitching layer; The fourth convolutional layer acquires feature data c and performs a convolution operation on the feature data c to obtain the fourth convolutional data d4, which is then passed to the first upsampling module. Step C2: The first upsampling module performs an upsampling operation on the fourth convolutional data d4 to obtain the first upsampled data d5, and then passes it to the first channel splicing layer; Step C3: The first channel stitching layer performs a channel-dimensional stitching operation on the third convolutional data d3 and the first upsampled data d5 to obtain the first stitched data d6, and then passes it to the first deformable attention bottleneck module; Step C4: The first deformable attention bottleneck module efficiently extracts the boundary information of the first spliced ​​data d6 to obtain the first deformable attention data d7, and passes it to the second upsampling module and the sixth channel splicing layer; Step C5: The second upsampling module upsamples the first deformable attention data d7 to obtain the second upsampled data d8, and then transmits it to the second channel splicing layer; Step C6: The second channel splicing layer performs a channel-dimensional splicing operation on the second convolutional data d2 and the second upsampled data d8 to obtain the second spliced ​​data d9, which is then passed to the second deformable attention bottleneck module. Step C7: The second deformable attention bottleneck module efficiently extracts the boundary information of the second spliced ​​data d9 to obtain the second deformable attention data d10, and passes it to the third upsampling module and the fifth channel splicing layer; Step C8: The third upsampling module performs an upsampling operation on the second deformable attention data d10 to obtain the third upsampled data d11, and transmits it to the third channel splicing layer; Step C9: The third channel splicing layer performs a channel-dimensional splicing operation on the first convolutional data d1 and the third upsampled data d11 to obtain the third spliced ​​data d12, and passes it to the third deformable attention bottleneck module. Step C10: The third deformable attention bottleneck module efficiently extracts the boundary information of the third splicing data d12 to obtain the third deformable attention data d13, and passes it to the fourth channel splicing layer; Step C11: The fourth channel splicing layer performs a channel-dimensional splicing operation on the first convolutional data d1, the second downsampled data b2, and the third deformable attention data d13 to obtain the fourth spliced ​​data d14, which is then passed to the fourth deformable attention bottleneck module. Step C12: The fourth deformable attention bottleneck module efficiently extracts the boundary information of the fourth spliced ​​data d14 to obtain the fourth deformable attention data d15, and passes it to the fifth convolutional layer; Step C13: The fifth convolutional layer performs a convolution operation on the fourth deformable attention data d15 to obtain the fifth convolutional data d16, and passes it to the sixth downsampling module and the image detection module; Step C14: The sixth downsampling module performs a downsampling operation on the fifth convolutional data d16 to obtain the sixth downsampled data d17, and passes it to the fifth channel splicing layer; Step C15: The fifth channel splicing layer performs a channel-dimensional splicing operation on the second convolutional data d2, the third downsampled data b4, the second deformable attention data d10, and the sixth downsampled data d17 to obtain the fifth spliced ​​data d18, which is then passed to the fifth deformable attention bottleneck module. Step C16: The fifth deformable attention bottleneck module efficiently extracts the boundary information of the fifth spliced ​​data d18 to obtain the fifth deformable attention data d19, and passes it to the seventh downsampling module and the image detection module; Step C17: The seventh downsampling module performs a downsampling operation on the fifth deformable attention data d19 to obtain the seventh downsampling data d20, and then transmits it to the sixth channel splicing layer; Step C18: The sixth channel stitching layer performs a channel-dimensional stitching operation on the third convolutional data d3, the fourth downsampled data b6, the first stitched data d6, and the seventh downsampled data d20 to obtain the sixth stitched data d21, which is then passed to the sixth deformable attention bottleneck module. Step C19: The sixth deformable attention bottleneck module efficiently extracts the boundary information of the sixth stitched data d21 to obtain fused data d, and transmits it to the image detection module; The first deformable attention bottleneck module, the second deformable attention bottleneck module, the third deformable attention bottleneck module, the fourth deformable attention bottleneck module, the fifth deformable attention bottleneck module, and the sixth deformable attention bottleneck module have the same structure. They are all equipped with a CA attention mechanism module, a twelfth downsampling module, a second Split segmentation function layer, a deformable convolutional structure, a ninth channel splicing layer, and a thirteenth downsampling module. The deformable convolutional structure is composed of m deformable convolutional blocks with identical structures connected in sequence, and the output end of each deformable convolutional block is connected to the input end of the ninth channel splicing layer. Each of the deformable convolutional blocks is sequentially provided with a fourteenth downsampling module and a deformable convolutional layer.

2. The aerial image target detection method based on a cross-level deformable feature aggregation network according to claim 1, characterized in that: In step 3, the feature extraction module performs feature extraction on the standard graphic data b, as follows: Step A1: The first downsampling module in the feature extraction module acquires the standard graphic data b, performs a downsampling operation on the standard graphic data b to obtain the first downsampling data b1, and passes it to the second downsampling module; Step A2: The second downsampling module performs a downsampling operation on the first downsampling data b1 to obtain the second downsampling data b2, and then passes it to the first image fusion module and the feature fusion module; Step A3: The first graph fusion module performs graph fusion operation on the second downsampled data b2 to obtain the first graph fusion data b3, and then passes it to the third downsampled module and the feature fusion module; Step A4: The third downsampling module performs a downsampling operation on the first image fusion data b3 to obtain the third downsampling data b4, and then transmits it to the second image fusion module and the feature fusion module; Step A5: The second graph fusion module performs graph fusion operation on the third downsampled data b4 to obtain the second graph fused data b5, and then passes it to the fourth downsampled module and the feature fusion module; Step A6: The fourth downsampling module performs a downsampling operation on the second image fusion data b5 to obtain the fourth downsampling data b6, and then transmits it to the third image fusion module and the feature fusion module; Step A7: The third graph fusion module performs graph fusion operation on the fourth downsampled data b6 to obtain the third graph fusion data b7, and then passes it to the fifth downsampled module and the feature fusion module; Step A8: The fifth downsampling module performs a downsampling operation on the third image fusion data b7 to obtain the fifth downsampling data b8, and then transmits it to the fourth image fusion module; Step A9: The fourth graph fusion module performs graph fusion operation on the fifth downsampled data b8 to obtain the fourth graph fusion data b9, and then passes it to the spatial pyramid pooling module; Step A10: The spatial pyramid pooling module performs multi-scale pooling operation on the fourth graph fusion data b9 to obtain feature data c, and then passes it to the feature fusion module.

3. The aerial image target detection method based on a cross-level deformable feature aggregation network according to claim 2, characterized in that: The first image fusion module, the second image fusion module, the third image fusion module, and the fourth image fusion module have the same structure. They are all provided with an eighth downsampling module, a first Split segmentation function layer, a bottleneck structure, a seventh channel splicing layer, and a ninth downsampling module connected in sequence. The bottleneck structure is composed of K bottleneck layers with the same structure connected in sequence. The output of each bottleneck layer is connected to the input of the seventh channel splicing layer; each bottleneck layer is provided with a sixth convolutional layer, a seventh convolutional layer and a residual connection layer in sequence; Let F be the input data of the i-th graph fusion module. Then, the i-th graph fusion module performs graph fusion operation on the input data F, and the steps are as follows: Step D1: The eighth downsampling module performs a downsampling operation on the input data F to obtain the eighth downsampled data F1, and passes it to the first Split segmentation function layer; Step D2: The first Split function layer divides the eighth downsampled data F1 into first segmented data F2 and second segmented data F3. The first Split function layer passes the first segmented data F2 to the bottleneck structure and the seventh channel splicing layer, and passes the second segmented data F3 to the seventh channel splicing layer. Step D3: The first bottleneck layer in the bottleneck structure performs feature extraction and compression operations on the first segmented data F2 to obtain the first bottleneck data. And pass it to the second bottleneck layer and the seventh channel splicing layer; The j-th bottleneck layer for the (j-1)-th bottleneck data Perform feature extraction and compression operations to obtain the j-th bottleneck data. And pass it to the (j+1)th bottleneck layer and the seventh channel splicing layer; The Kth bottleneck layer for the (K-1)th bottleneck data Perform feature extraction and compression operations to obtain the Kth bottleneck data. , and pass it to the seventh channel splicing layer; the output data of K bottleneck layers together form the bottleneck output data F4 of the bottleneck structure; Step D4: The seventh channel stitching layer performs a channel-dimensional stitching operation on the first segmented data F2, the second segmented data F3, and the bottleneck output data F4 to obtain the seventh stitched data F5, which is then passed to the ninth downsampling module. Step D5: The ninth downsampling module performs a downsampling operation on the seventh stitched data F5 to obtain the fused data of the i-th image.

4. The aerial image target detection method based on a cross-level deformable feature aggregation network according to claim 2, characterized in that: In step A10, the spatial pyramid pooling module includes a tenth downsampling module, an eleventh downsampling module, a first max pooling layer, a second max pooling layer, a third max pooling layer, and an eighth channel stitching layer. The spatial pyramid pooling module performs multi-scale pooling operations on the fourth image fused data b9, as follows: Step B1: The tenth downsampling module acquires the fourth image fusion data b9 and performs downsampling operation to obtain the tenth downsampling data c1, which is then passed to the first max pooling layer and the eighth channel splicing layer. Step B2: The first max pooling layer performs max pooling on the tenth downsampled data c1 to obtain the first max pooled data c2, and then passes it to the second max pooling layer and the eighth channel splicing layer. Step B3: The second max pooling layer performs max pooling on the first max pooling data c2 to obtain the second max pooling data c3, and then passes it to the third max pooling layer and the eighth channel splicing layer. Step B4: The third max pooling layer performs max pooling on the second max pooling data c3 to obtain the third max pooling data c4, and then passes it to the eighth channel splicing layer. Step B5: The eighth channel splicing layer performs a channel-dimensional splicing operation on the tenth downsampled data c1, the first max pooled data c2, the second max pooled data c3, and the third max pooled data c4 to obtain the eighth spliced ​​data c5, which is then passed to the eleventh downsampled module. Step B6: The eleventh downsampling module performs a downsampling operation on the eighth spliced ​​data c5 to obtain feature data c.

5. The aerial image target detection method based on a cross-level deformable feature aggregation network according to claim 2, characterized in that: The first, second, third, fourth, fifth, sixth, seventh, eighth, ninth, tenth, eleventh, twelfth, thirteenth and fourteenth downsampling modules have the same structure, and each is provided with a first two-dimensional convolutional layer, a batch normalization layer and a SiLU activation function layer connected in sequence.

6. The aerial image target detection method based on a cross-level deformable feature aggregation network according to claim 1, characterized in that: In step 5, the image detection module is equipped with a first detection head, a second detection head, and a third detection head; The first detection head acquires the fifth convolutional data d12 and performs a detection operation on it to obtain the first detection data e1; the second detection head acquires the fifth deformable attentional data d14 and performs a detection operation on it to obtain the second detection data e2; the third detection head acquires the fused data d and performs a detection operation on it to obtain the third detection data e3; the image detection module merges the first detection data e1, the second detection data e2, and the third detection data e3 and outputs the detection result e.

7. The aerial image target detection method based on a cross-level deformable feature aggregation network according to claim 6, characterized in that: The first, second, and third detection heads have the same structure, and each is equipped with an eighth, ninth, tenth, and eleventh convolutional layer, a second two-dimensional convolutional layer, a third two-dimensional convolutional layer, a bounding box loss function layer, and a binary classification cross-entropy loss function layer. The input of the eighth convolutional layer is connected to the feature fusion module, and the output is connected to the bounding box loss function layer via the ninth convolutional layer and the second two-dimensional convolutional layer in sequence; the input of the tenth convolutional layer is connected to the feature fusion module, and the output is connected to the binary classification cross-entropy loss function layer via the eleventh convolutional layer and the third two-dimensional convolutional layer in sequence.

8. The aerial image target detection method based on a cross-level deformable feature aggregation network according to claim 7, characterized in that: The bounding box loss function layer applies the WIoU loss function to the Inner-IoU loss function based on the auxiliary bounding box, resulting in the Inner-WIoU bounding box loss function. The calculation formula for the Inner-IoU loss function based on the auxiliary bounding box is as follows: ， ； ， ； ， ； ， ； ； ； ； in, This represents the coordinates of the center point of the true bounding box. This indicates the coordinates of the center point of the prediction box. This represents the actual frame width. represents the true bounding box height, 𝑤 represents the predicted bounding box width, ℎ represents the predicted bounding box height; br represents the scaling factor, ranging from [0.5, 1.5]; This represents the IoU loss function value based on the auxiliary bounding box; This indicates the coordinates of the top-left corner of the prediction box. This represents the coordinates of the top-left corner of the true bounding box. This indicates the coordinates of the bottom right corner of the prediction box. This represents the coordinates of the bottom right corner of the actual bounding box; The formula for calculating the bounding box loss function Inner-WIoU is as follows: ； ； ； ； in, Indicates the current The mean, This represents the intersection-union ratio (IUU) between the predicted bounding boxes and the ground truth bounding boxes. Used to measure the center distance between the predicted bounding box and the ground truth bounding box; This represents the natural exponential function, that is, e raised to the power of n; This represents the width of the smallest bounding rectangle between the predicted bounding box and the ground truth bounding box. This represents the height of the smallest bounding rectangle between the predicted bounding box and the ground truth bounding box; The superscript * in the text indicates that and Separate from the computation graph; This represents the bounding box loss function value. β represents the monotonic focus factor; β represents the outlier of the anchor box, and the higher the outlier, the worse the quality of the predicted box. α represents the non-monotonic frequency modulation factor, controlled by outliers β and α; α and δ represent model parameters, with α taking the value of 1.9 and δ taking the value of 3.

Images