An unmanned aerial vehicle aerial micro target detection method based on spatial and frequency domain feature decoupling

Abstract

The application belongs to the technical field of computer vision and image processing, and discloses a UAV aerial micro target detection method based on spatial and frequency domain feature decoupling. The method comprises the following steps: acquiring a UAV aerial high-resolution image and performing block preprocessing; inputting the preprocessed image into a detection network, performing frequency domain feature preserving down-sampling through an ADWT (analog to digital wavelet transform) module, and extracting high-frequency authentic shallow features; utilizing an MCG (multi-scale cross-band guided fusion) module to perform residual modulation enhancement on deep low-frequency semantic features through generating multi-scale high-frequency spatial masks, so as to realize decoupling fusion of targets and background clutter; and combining a joint normalized Wasserstein distance (NWD) to measure the category and bounding box of the micro target. The application alleviates the information loss problem caused by traditional spatial down-sampling, improves the micro target detection precision while taking into account the model lightweight, and is suitable for deployment on UAV edge devices.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and image processing technology, and specifically relates to a method for detecting small targets in aerial images based on deep learning, applicable to edge devices of unmanned aerial vehicles. Specifically, it is a method for detecting small targets in aerial images of unmanned aerial vehicles based on the decoupling of spatial and frequency domain features. Background Technology

[0002] With the widespread adoption of unmanned aerial vehicle (UAV) technology, target detection in UAV aerial images plays an irreplaceable and crucial role in time-sensitive tasks such as maritime search and rescue, post-disaster reconnaissance, and urban traffic inspection. Compared to natural scene images, UAV aerial images typically feature a high field of view and wide coverage, resulting in targets such as people in distress at sea, distant ships, or vehicles occupying very few pixels in the overall image—typically ultra-small targets. Simultaneously, aerial scenes are often accompanied by complex background clutter such as sea surface reflections and dense urban building textures, posing a significant challenge to high-precision target detection.

[0003] Existing deep learning-based object detection solutions typically focus on feature extraction, multi-scale fusion, and loss function optimization to improve detection accuracy. However, in scenarios involving the detection of small targets during drone aerial photography, current technologies still exhibit several physical and engineering limitations.

[0004] First, traditional spatial downsampling leads to the irreversible loss of high-frequency features of small targets. Although existing mainstream methods attempt to alleviate feature loss by introducing self-attention mechanisms or adding high-resolution detection heads, this inevitably introduces intensive floating-point operations, which are difficult to meet the stringent computational resource (SWaP) constraints of UAV edge devices (Zhu X, Lyu S, Wang X, et al. TPH-YOLOv5: Improved YOLOv5 based on transformer predictionhead for object detection on drone-captured scenarios[C] / / Proceedings of the IEEE / CVF International Conference on Computer Vision Workshops. 2021: 2778-2788). To adapt to edge devices, existing lightweight models often use stride convolution operations for dimensionality reduction. However, this destroys and discards the high-frequency contours and edge details of the target. The weak semantic information of small targets is almost completely consumed after several spatial downsampling operations, resulting in irreversible spatial aliasing and feature degradation (Zhang R. Making convolutional networks shift-invariant again[C] / / International Conference on Machine Learning. PMLR, 2019: 7324-7334).

[0005] Secondly, traditional multi-scale feature fusion leads to background clutter aliasing and target feature "hollowing out". Existing cross-scale feature fusion techniques usually achieve this through interpolation-based upsampling, followed by element-wise addition or simple concatenation (Lin TY, Dollár P, Girshick R, et al. Feature pyramid networks for object detection[C] / / Proceedings of the IEEE conference on computer vision and pattern recognition. 2017: 936-944). In aerial images with extremely complex backgrounds, this direct spatial aggregation method introduces significant feature aliasing, easily mixing dense urban building textures and other background noise with small targets. This causes the weak semantic response of sub-pixel targets to be diluted or even "hollowed out" by surrounding clutter, resulting in a large number of false detections (Li H. Rethinking features-fused-pyramid-neck for object detection[C] / / European Conference on Computer Vision. 2024: 74-90).

[0006] Finally, traditional loss metrics based on Intersection over Union (IoU) are extremely sensitive to the positional deviations of tiny targets. In mainstream object detectors, IoU and its variants are widely used as standard metrics for bounding box regression. However, for sub-pixel-level tiny targets, a prediction offset of just one pixel can cause the IoU value to plummet or even drop to zero. This extreme sensitivity greatly disrupts the stability of the loss space, easily leading to gradient vanishing or breaking, severely hindering the convergence of the network during training and weakening its localization robustness (Xu C, Wang J, Yang W, et al. Detecting tinyobjects in aerial images: A normalized Wasserstein distance and a newbenchmark[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2022, 190:79-93). Summary of the Invention

[0007] The purpose of this invention is to provide a method for detecting small targets in UAV aerial photography based on the decoupling of spatial and frequency domain features, in order to solve the technical problems of loss of high-frequency contours in traditional spatial domain downsampling, background clutter mixing in multi-scale fusion, and gradient breakage in regression loss.

[0008] The technical solution adopted by this invention to solve its technical problem includes the following steps:

[0009] A method for detecting small targets in UAV aerial photography based on the decoupling of spatial and frequency domain features includes the following steps:

[0010] S1. Acquire a high-resolution aerial image of the drone to be detected, and perform block preprocessing on the high-resolution aerial image to obtain the network input image;

[0011] S2. Input the network input image into the backbone network of the pre-trained small target detection network, and use the high-frequency discrete wavelet transform module to perform frequency domain feature-preserving downsampling to extract a high-frequency fidelity shallow feature map;

[0012] S3. Input the shallow feature map and the low-frequency semantic feature map from the deep layer of the micro-target detection network into the multi-scale cross-band guided fusion module, generate a high-frequency spatial mask using the shallow feature map, and perform residual modulation enhancement on the low-frequency semantic feature map using the high-frequency spatial mask. Then, concatenate the enhanced low-frequency feature map with the shallow feature map to obtain decoupled fusion features.

[0013] S4. Input the decoupled and fused features into the detection head, combine them with the joint normalized Wasserstein distance metric to perform bounding box regression prediction, and output the category and bounding box of the small target.

[0014] Furthermore, the step of using a high-frequency discrete wavelet transform module to perform frequency domain feature-preserving downsampling to extract a high-frequency fidelity shallow feature map specifically includes: performing global average pooling and linear mapping on the input spatial feature map to generate adaptive frequency domain channel attention weights; multiplying the adaptive frequency domain channel attention weights with the input spatial feature map channel by channel to obtain a pre-cleaned feature map; performing frequency domain decomposition on the pre-cleaned feature map using a two-dimensional discrete wavelet transform to separate it into a low-frequency semantic sub-band and three high-frequency contour sub-bands; concatenating the four sub-bands along the channel dimension, and performing channel dimensionality reduction and feature fusion through convolution to output the shallow feature map.

[0015] Furthermore, the two-dimensional discrete wavelet transform employs discrete wavelet basis functions with orthogonal or bioorthogonal properties, wherein the discrete wavelet basis functions include any one of Haar wavelet, Daubechies wavelet, Symlets wavelet, or Bioorthogonal wavelet.

[0016] Furthermore, the step of generating a high-frequency spatial mask using the shallow feature map specifically includes: obtaining the shallow feature map, inputting it in parallel into a dilated convolution branch containing multiple dilation rates, and extracting multi-scale physical boundary structure features; after channel fusion of the multi-scale physical boundary structure features, processing them through a Sigmoid activation function to generate the high-frequency spatial mask.

[0017] Furthermore, the generation of the high-frequency spatial mask and the residual modulation enhancement of the low-frequency semantic feature map specifically satisfy the following calculation formula:

[0018]

[0019]

[0020] in, This represents the shallow feature map. This represents the low-frequency semantic feature map. and These represent channel max pooling and average pooling, respectively. This indicates a splicing operation. Indicates the expansion rate of Hollow convolution, This represents the Sigmoid activation function; This represents the generated high-frequency spatial mask. This represents the enhanced low-frequency feature map. This indicates element-wise multiplication along spatial dimensions.

[0021] Furthermore, the method of combining the joint normalized Wasserstein distance metric for bounding box regression prediction specifically includes: during the model training phase, modeling both the ground truth bounding box of the small target and the bounding box predicted by the small target detection network as a two-dimensional Gaussian distribution; calculating the normalized Wasserstein distance between the two-dimensional Gaussian distribution corresponding to the ground truth bounding box and the two-dimensional Gaussian distribution corresponding to the predicted bounding box; constructing a joint regression loss function that includes the normalized Wasserstein distance loss and the complete intersection-union ratio loss, and combining the classification loss to perform backpropagation gradient optimization on the small target detection network.

[0022] Furthermore, the preprocessing of the high-resolution aerial image into segments specifically includes: using a sliding window strategy to segment the high-resolution aerial image into sub-images of a preset size; wherein the pixel overlap rate between adjacent sub-images is set to 10% to 30%.

[0023] Furthermore, the micro-target detection network supports architecture scaling by introducing depth scaling factors and width scaling factors to dynamically adjust the number of channels in the convolutional layers and the number of stacked layers in the feature extraction module.

[0024] The present invention also provides an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to execute the above-described method for detecting small targets in UAV aerial photography based on decoupling of spatial and frequency domain features.

[0025] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method for detecting small targets in UAV aerial photography based on the decoupling of spatial and frequency domain features.

[0026] Compared with the prior art, the micro-target detection method of the present invention has the following beneficial technical effects:

[0027] 1. This invention achieves frequency-domain feature-preserving downsampling, which helps alleviate the SWaP limitations and high-frequency feature degradation problems of UAVs. Addressing the issue of sub-pixel edge loss of tiny targets due to pure spatial downsampling in existing technologies, this invention uses an ADWT module to transfer the downsampling process to the frequency domain. Since the Haar wavelet transform is essentially a parameter-free orthogonal linear projection, this invention folds the spatial resolution down to the channel dimension with extremely low computational overhead. This mechanism, without introducing heavy large-kernel convolutions or complex self-attention operations, helps preserve the key contour information of extremely small targets and is more suitable for the SWaP limitations of edge devices such as UAVs.

[0028] 2. An innovative residual modulation fusion architecture achieves physical separation of small targets from complex background clutter. Addressing the problem of background aliasing and target "hollowing" caused by direct feature pyramid stitching in existing methods, the MCG module of this invention differs from traditional pure spatial stitching fusion methods. This invention innovatively utilizes high-frequency prior features to generate a spatial mask as a spatial attention guide, and through a specific residual modulation mechanism, it directionally enhances the low-frequency semantic base of the deep network. The effect of this technique is that while preserving the global contextual semantic structure, it also enhances the local response intensity of small targets with the help of high-frequency guidance, effectively suppressing false alarms caused by strong background noise such as sea surface reflections and urban building textures.

[0029] 3. The gradient space of regression optimization is smoothed, improving detection accuracy while maintaining a lightweight model. Addressing the issue of gradient breakage caused by the extreme sensitivity of traditional IoU loss to tiny targets, this invention introduces a joint normalized Wasserstein distance metric mechanism. This model the target bounding box as a Gaussian distribution, effectively smoothing the drastic error fluctuations caused by minute positional deviations of ultra-small targets, thus enhancing the model's localization robustness. Combining these techniques, this invention maintains extremely low parameter levels and high inference speed, achieving high recall and average precision for densely distributed tiny targets even in complex aerial photography scenarios with extreme scale compression, thus balancing accuracy and efficiency. Attached Figure Description

[0030] Figure 1 This is a flowchart illustrating a method for detecting small targets in UAV aerial photography based on the decoupling of spatial and frequency domain features, provided in an embodiment of the present invention.

[0031] Figure 2 This is a schematic diagram of the overall architecture of the micro-target detection network provided in an embodiment of the present invention.

[0032] Figure 3 This is a schematic diagram of the internal structure of the high-frequency discrete wavelet transform (ADWT) module provided in an embodiment of the present invention.

[0033] Figure 4 This is a schematic diagram of the internal structure of the multi-scale cross-band guided fusion (MCG) module provided in an embodiment of the present invention. Detailed Implementation

[0034] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0035] like Figure 1 Overall algorithm flowchart and Figure 2 As shown in the overall architecture diagram of the micro-target detection network, this embodiment of the invention provides a method for detecting micro-targets in UAV aerial photography based on the decoupling of spatial and frequency domain features. This method alleviates the information loss problem caused by traditional spatial domain downsampling, and its core execution steps are as follows:

[0036] In step S1, a high-resolution aerial image taken by the UAV is acquired and preprocessed by slicing. Due to the large size of the aerial image and the high density of small targets, this embodiment employs a sliding window strategy to slice the high-resolution image into sub-images of a preset size, and sets pixel overlap regions between adjacent sub-images. The processed sub-images are then input into a trained small target detection network. Preferably, the pixel overlap rate between adjacent sub-images is set to a range of 10% to 30% to prevent small targets located at the cutting edges from being truncated.

[0037] In step S2, as Figure 2 The backbone network and Figure 3 As shown, the preprocessed image undergoes feature extraction in the backbone network. To address the problem of lost high-frequency spatial details of small targets due to traditional stride convolution, this invention employs a high-frequency discrete wavelet transform (ADWT) module for frequency-domain feature-preserving downsampling. Specifically, this includes three sub-steps: channel pre-cleaning, frequency-domain downsampling, and feature reconstruction. First, global average pooling is performed on the input spatial feature map. Adaptive frequency-domain channel attention weights are generated using a multilayer perceptron and a sigmoid function, and these weights are multiplied channel-by-channel with the input feature map to obtain a pre-cleaned feature map. Then, a two-dimensional Haar discrete wavelet transform is used to perform frequency-domain decomposition on the pre-cleaned feature map, separating a low-frequency semantic subband (LL) and three high-frequency contour subbands (LH, HL, HH) while minimizing information loss. Finally, the four subbands are concatenated along the channel dimension, achieving halved spatial resolution while quadrupling the number of channels with aliasing-free downsampling. Subsequently, 1×1 convolution is used for channel dimensionality reduction and feature fusion, outputting a high-frequency fidelity shallow feature map.

[0038] In step S2, the specific execution process of the high-frequency discrete wavelet transform (ADWT) module is as follows:

[0039] First, to prevent noise from spreading across frequency bands, the input feature map is... An adaptive frequency-domain aware channel gating method is introduced for pre-cleaning. Global average pooling (GAP) and multilayer perceptron (MLP) are used to evaluate the semantic purity of each channel within the global receptive field, generating an adaptive frequency-domain aware attention mask. :

[0040]

[0041] in, This represents the Sigmoid function. The mask... Compared with the original input features Adaptive dynamic filtering is achieved by multiplying each channel sequentially.

[0042] Subsequently, orthogonal projection downsampling is performed using two-dimensional Haar wavelet basis functions. Let the one-dimensional low-pass filter be... The high-pass filter is The four two-dimensional orthogonal filters are defined by their Kronecker products:

[0043]

[0044]

[0045]

[0046]

[0047] in, It represents the Kronecker product.

[0048] Given an intermediate feature map, the above filter is applied with a sliding window of stride 2 to decompose it into a low-frequency semantic base. and three high-frequency profile components , Its downsampling operation is formulated as follows:

[0049]

[0050]

[0051] in, Finally, through Convolution merges all subbands along the channel dimension.

[0052] In step S3, as Figure 2 The neck network and Figure 4As shown, to address the background clutter problem during feature pyramid fusion, this invention designs a multi-scale cross-band guided fusion (MCG) module for decoupling and brightening deep and shallow features. First, a high-frequency spatial mask is generated: the shallow feature map output from step S2 is obtained, and the physical boundary structure is extracted using multi-scale parallel dilated convolution. After channel fusion and Sigmoid activation, a high-frequency spatial mask for more accurate localization is generated. Then, residual soft attention modulation is performed: the low-frequency semantic feature map obtained from deep network upsampling is obtained, and the high-frequency spatial mask is used for residual modulation enhancement. The residual modulation process follows the physical modulation logic described above, i.e., by accumulating the residuals of the product terms of the low-frequency semantic feature map and the high-frequency spatial mask, explicit enhancement of the target features is achieved. This operation preserves global low-frequency semantic features while explicitly enhancing the local response features of small target regions through the high-frequency mask. Finally, the enhanced low-frequency feature map and the original high-frequency shallow feature map are concatenated along the channel dimension, and the output fused tensor after feature decoupling is sent to the detection head.

[0053] In step S3, the specific execution process of the multi-scale cross-band guided fusion (MCG) module is as follows:

[0054] set up This represents the high-level semantic features after upsampling. This represents the corresponding low-level features from the frequency-domain sensing backbone network. First, from... Mining high-frequency spatial attention masks :

[0055]

[0056] in, and These represent channel max pooling and average pooling, respectively. This indicates a splicing operation. Indicates the expansion rate of Hollow convolution, This represents the Sigmoid function.

[0057] Subsequently, a residual soft attention mechanism is introduced, using the generated high-frequency mask. To modulate high-level semantic features :

[0058] in, This represents element-wise multiplication along a spatial dimension. The residual formula... It acts as a soft gating mechanism, preserving the integrity of the low-frequency semantic base while amplifying boundary activation.

[0059] Finally, the modulated semantic features With low-level features splicing, through convolution Integrating cross-frequency band information:

[0060]

[0061] In step S4, regression prediction is performed using the joint normalized Wasserstein distance. In the detection head, both the ground truth bounding box and the network-predicted bounding box of the small target are modeled as a two-dimensional Gaussian distribution, and the normalized Wasserstein distance between them is calculated. A joint regression loss function, comprising the normalized Wasserstein distance loss and the complete intersection-union ratio (CIU) loss, is constructed, and the entire detection network is optimized using backpropagation gradients in conjunction with the classification loss. During the inference phase, the trained network directly outputs the class label and relatively accurate predicted bounding box coordinates of the small target in the aerial image to be detected.

[0062] First, the prediction box and real frame Modeled as a two-dimensional Gaussian distribution and Calculate the second-order Wasserstein distance between these two distributions. :

[0063]

[0064] in, Indicates the coordinates of the bounding box center. and These represent the width and height, respectively.

[0065] To integrate this distance metric into the loss function of the neural network, it is mapped to a nonlinear exponential transformation. Within the range, the normalized Wasserstein distance (NWD) is obtained:

[0066]

[0067] in, It is an empirical constant that is dynamically related to the average scale of the target.

[0068] Furthermore, to combine the advantages of precise fitting of large targets with robust localization of small targets, a joint metric is constructed:

[0069]

[0070] in, Here are the hyperparameters used for balancing. The corresponding joint bounding box regression loss is defined as:

[0071]

[0072] In step S5, a model family is constructed and adaptively deployed based on scaling factors. To adapt to UAV edge devices with different computing power levels, this invention introduces depth and width scaling factors to dynamically adjust network parameters, constructing a multi-level detection model family. Specifically, this includes a Nano version of an ultra-lightweight model suitable for extremely low-power devices, and a Small version of a lightweight model. The lightweight detection network can be deployed on several UAV-borne edge computing platforms. The edge computing platform includes an embedded graphics processor, an edge neural network computing chip, a dedicated NPU acceleration engine, and a field-programmable gate array (FPGA) and a system-on-a-chip (SoC) based on a heterogeneous computing architecture. This allows for hardware-level parallel acceleration to meet the low-latency inference requirements of the airborne edge in a limited environment without cloud network connectivity.

[0073] In addition, to verify the effectiveness of this method, this embodiment conducted multi-dimensional comparative experiments with mainstream object detection networks on two large public aerial photography datasets, VisDrone and DOTA-V1.5. Experiments based on the VisDrone dataset show that, compared to classic lightweight models, the proposed lightweight model significantly reduces the number of parameters and computational cost, which helps alleviate the computational limitations of edge devices such as micro-drones. Dual-resolution experiments based on the DOTA-V1.5 dataset show that this invention still achieves excellent high-frequency feature fidelity under extremely compressed resolution; while in high-resolution tests, it achieves superior average accuracy with very low computational cost, indicating that the multi-scale cross-band guided fusion module reduces background aliasing and redundant computation, achieving feature response enhancement and computational efficiency optimization. The comparative experimental results are shown in Tables 1 and 2, respectively.

[0074] Table 1 shows the comparison experiment using the VisDrone dataset.

[0075]

[0076] Table 2 shows the comparison experiment of the DOTA-V1.5 dataset.

[0077]

[0078] This invention also provides an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to execute the above-described method for detecting small targets in UAV aerial photography based on decoupling of spatial and frequency domain features.

[0079] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for detecting small targets in UAV aerial photography based on the decoupling of spatial and frequency domain features.

[0080] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for detecting small targets in UAV aerial photography based on the decoupling of spatial and frequency domain features, characterized in that, Includes the following steps: S1. Acquire a high-resolution aerial image of the drone to be detected, and perform block preprocessing on the high-resolution aerial image to obtain the network input image; S2. Input the network input image into the backbone network of the pre-trained small target detection network, and use the high-frequency discrete wavelet transform module to perform frequency domain feature-preserving downsampling to extract a high-frequency fidelity shallow feature map; S3. Input the shallow feature map and the low-frequency semantic feature map from the deep layer of the micro-target detection network into the multi-scale cross-band guided fusion module, generate a high-frequency spatial mask using the shallow feature map, and perform residual modulation enhancement on the low-frequency semantic feature map using the high-frequency spatial mask. Then, concatenate the enhanced low-frequency feature map with the shallow feature map to obtain decoupled fusion features. S4. Input the decoupled and fused features into the detection head, combine them with the joint normalized Wasserstein distance metric to perform bounding box regression prediction, and output the category and bounding box of the small target.

2. The method according to claim 1, characterized in that, The step of using a high-frequency discrete wavelet transform module to perform frequency domain feature-preserving downsampling to extract a high-frequency fidelity shallow feature map specifically includes: Global average pooling and linear mapping are performed on the input spatial feature map to generate adaptive frequency domain channel attention weights; The adaptive frequency domain channel attention weights are multiplied channel by channel with the input spatial feature map to obtain a pre-cleaned feature map; The pre-cleaned feature map is decomposed in the frequency domain using two-dimensional discrete wavelet transform, separating it into a low-frequency semantic sub-band and three high-frequency contour sub-bands. The four sub-bands are concatenated along the channel dimension, and channel dimensionality reduction and feature fusion are performed through convolution to output the shallow feature map.

3. The method according to claim 2, characterized in that, The two-dimensional discrete wavelet transform employs discrete wavelet basis functions with orthogonal or bioorthogonal properties, including any one of Haar wavelet, Daubechies wavelet, Symlets wavelet, or Bioorthogonal wavelet.

4. The method according to claim 3, characterized in that, When the discrete wavelet basis function adopts the Haar wavelet, its low-pass filter With high-pass filter Four two-dimensional orthogonal filters are generated using the Kronecker product. Furthermore, the downsampling operation of frequency domain decomposition satisfies the following formula: in, This represents the pre-cleaned feature map. This represents the index coordinates of the feature map in the channel and spatial dimensions. This represents the low-frequency semantic subband obtained from the decomposition. This represents the high-frequency profile subband obtained from the decomposition, and .

5. The method according to claim 1, characterized in that, The process of generating a high-frequency spatial mask using the shallow feature map specifically includes: obtaining the shallow feature map and inputting it in parallel into a dilated convolution branch containing multiple dilation rates to extract multi-scale physical boundary structure features. After channel fusion of the multi-scale physical boundary structure features, the high-frequency spatial mask is generated by processing with the Sigmoid activation function.

6. The method according to claim 5, characterized in that, The generation of the high-frequency spatial mask and the residual modulation enhancement of the low-frequency semantic feature map specifically satisfy the following calculation formula: in, This represents the shallow feature map. This represents the low-frequency semantic feature map. and These represent channel max pooling and average pooling, respectively. This indicates a splicing operation. Indicates the expansion rate of Hollow convolution, This represents the Sigmoid activation function; This represents the generated high-frequency spatial mask. This represents the enhanced low-frequency feature map. This indicates element-wise multiplication along spatial dimensions.

7. The method according to claim 1, characterized in that, The method of combining joint normalized Wasserstein distance metric for bounding box regression prediction specifically includes: During the model training phase, both the true bounding boxes of the small targets and the bounding boxes predicted by the small target detection network are modeled as two-dimensional Gaussian distributions. Calculate the normalized Wasserstein distance between the two-dimensional Gaussian distribution corresponding to the true bounding box and the two-dimensional Gaussian distribution corresponding to the predicted bounding box; A joint regression loss function is constructed, which includes the normalized Wasserstein distance loss and the complete intersection-union ratio loss. The backpropagation gradient is then used to optimize the small target detection network by combining the classification loss.

8. The method according to claim 7, characterized in that, The normalized Wasserstein distance and the joint regression loss function The calculation satisfies the following formula: in, Represents the predicted bounding box. Represents the true bounding box. This represents the second-order Wasserstein distance between two-dimensional Gaussian distributions. and These represent the center coordinates of the predicted bounding box and the ground truth bounding box, respectively. and These represent the width and height, respectively. It is a smoothing constant. These are hyperparameters used to balance different metrics.

9. The method according to claim 1, characterized in that, The preprocessing of the high-resolution aerial image into blocks specifically includes: The aerial high-resolution image is segmented into sub-images of a preset size using a sliding window strategy; The pixel overlap rate between adjacent sub-images is set to 10% to 30%.

10. The method according to any one of claims 1 to 9, characterized in that, The micro-target detection network supports architecture scaling by introducing depth scaling factors and width scaling factors to dynamically adjust the number of channels in the convolutional layers and the number of stacked layers in the feature extraction module.

11. An electronic device, characterized in that, include: At least one processor; A memory that is communicatively connected to the at least one processor; The memory stores a computer program that can be executed by the at least one processor, which is then executed by the at least one processor to enable the at least one processor to perform the UAV aerial photography small target detection method based on spatial and frequency domain feature decoupling as described in any one of claims 1 to 10.

Images