A real-time aerial small target detection method based on phase perception and frequency domain enhancement

Abstract

This invention discloses a real-time aerial small target detection method based on phase awareness and frequency domain enhancement. First, a spatial-frequency dual-stream decoupling module simultaneously captures the target's local spatial semantics and high-frequency gradient structure during the feature extraction stage, and a hard-coded structural prior is constructed using the Scharr operator. Then, an absolute texture feature is extracted using a low-level frequency enhancer combined with a Log-Gabor filter. Based on this, a semantically gated phase interaction mechanism is used to strip away the easily disturbed amplitude spectrum in the frequency domain and reconstruct the structural prior using pure phase features. This is combined with a high-level semantic mask to achieve strong noise resistance across scales. Finally, at the detection head output, a frequency domain edge-gated hybrid expert module based on the Laplacian operator is used to achieve pixel-level dynamic feature refinement. This invention effectively revives sub-pixel-level target details and suppresses background false alarms, significantly improving the recall rate of aerial small target detection while maintaining extremely high real-time inference efficiency.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and computer vision technology, specifically to deep learning, UAV remote sensing image processing and digital signal frequency domain analysis technology, and particularly to a real-time aerial small target detection method based on phase perception and frequency domain enhancement. Background Technology

[0002] With the rapid popularization of unmanned aerial vehicles (UAVs) and low-altitude remote sensing technologies, aerial image target detection plays an increasingly important role in fields such as real-time urban traffic monitoring, wilderness search and rescue, precision agricultural plant protection, and power line inspection. However, compared with conventional natural scene images, aerial images are usually taken from a high-altitude overhead perspective, resulting in images with extremely complex background textures and very small and densely distributed targets. Although existing target detection algorithms, especially mainstream detection architectures represented by Transformer (such as RT-DETR), perform well on general datasets, they still reveal the following deep-seated technical shortcomings when directly applied to aerial photography scenarios:

[0003] First, feature extraction backbone networks face severe physical-level spatial structure degradation when processing extremely small targets. Key targets in aerial images (such as distant vehicles and pedestrians) often occupy only a few to tens of pixels. During the aggressive downsampling (Stride convolution or pooling) adopted by existing networks to pursue computational efficiency, the feature responses of these extremely small targets decay rapidly. In particular, in order to control the number of model parameters, existing solutions usually tend to directly discard high-resolution shallow features containing rich geometric details (such as the P2 layer). This approach leads to severe irreversible degradation of the fine-grained high-frequency texture and edge prior information of the target before it enters the deep network, greatly increasing the model's false negative rate for small objects.

[0004] Secondly, traditional attention mechanisms are highly susceptible to signal amplitude interference from abnormally bright areas in complex aerial photography backgrounds. Existing cross-scale feature interaction units heavily rely on the amplitude response of feature maps for similarity measurement and weight allocation. However, aerial photography scenes widely contain highly reflective areas such as reflective rooftops, glass curtain walls, shimmering water surfaces, and metal roofs, which exhibit extreme high brightness amplitudes in the image domain. During Softmax normalization calculations, the spurious high responses generated by these noisy areas forcibly occupy attention weights, misleading the network to concentrate limited computational resources on background noise. This results in the severe obfuscation of real, small targets located in dark areas, shadow areas, or low-contrast backgrounds.

[0005] Finally, fixed receptive fields are ill-suited to the dynamic challenges of aerial photography targeting multi-scale objects. In aerial detection tasks, algorithm performance heavily relies on the quality of specific layers in the feature pyramid (such as the P3 layer), as this layer is crucial for balancing the spatial resolution and semantic depth of small targets. However, general convolutional blocks or Transformer modules typically employ a single, fixed receptive field design, creating an irreconcilable technical contradiction: on the one hand, accurately anchoring extremely small targets requires an extremely narrow receptive field to preserve edge sharpness; on the other hand, capturing the complete geometry of slightly larger targets or eliminating environmental false alarms requires a wider contextual receptive field. This "one-size-fits-all" feature extraction approach leads to frequent classification errors or localization drift when dealing with aerial targets of drastically varying scales, resulting in a large number of background false alarms.

[0006] This invention aims to improve the ability to capture the edge details of small targets and the robustness to complex background noise in embedded computing-limited environments by decoupling and reconstructing the features of aerial images in the spatial and frequency domains. It is particularly suitable for application scenarios with extremely high requirements for real-time performance and small target recall, such as UAV power line inspection, urban security monitoring, search and rescue, and precision agriculture. Summary of the Invention

[0007] To overcome the limitations of existing aerial target detection technologies in preserving small target features, resisting interference from complex backgrounds, and adapting to multi-scale receptive fields, this invention provides a real-time aerial small target detection method and system based on phase perception and frequency domain enhancement, namely the LPH-DETR architecture. The core design idea of ​​this invention lies in explicitly guiding a deep learning network to "revive" sub-pixel-level target details and achieve "physical-level" filtering of bright background noise while maintaining high inference speed by introducing frequency domain priors and phase analysis tools.

[0008] The technical solution of this invention is: a real-time aerial small target detection method based on phase sensing and frequency domain enhancement, comprising the following steps:

[0009] (1) Multi-scale feature extraction: The aerial image to be detected is input into the feature extraction backbone network, which includes a spatial-frequency dual-stream decoupling module; the spatial structure information and high-frequency gradient information of the aerial image are extracted by the spatial-frequency dual-stream decoupling module respectively, and dynamic weighted fusion is performed to output a multi-scale basic feature map, which includes at least a bottom-level feature map, a middle-level feature map and a high-level feature map;

[0010] (2) Lossless restoration of low-level texture: The low-level feature map is input to the low-level frequency enhancer to extract the high-frequency texture features of the low-level feature map, and the spatial neighbor pixels of the high-frequency texture features are rearranged to the channel dimension using the spatial-to-depth transformation technique. Then, it is spliced ​​and fused with the middle-level feature map to obtain the middle-level feature map after detail enhancement.

[0011] (3) Semantic gated phase cross-scale interaction: The high-level feature map is upsampled and a semantic mask is generated through the semantic gated phase interaction module; at the same time, Fourier transform is performed on the mid-level feature map after detail enhancement to separate the amplitude spectrum and phase spectrum, and only the phase spectrum is used for inverse transformation to obtain the phase structure feature; the phase structure feature is multiplied with the semantic mask to suppress high-frequency background noise and obtain the purified cross-scale interaction feature;

[0012] (4) Expert feature refinement based on frequency domain edge gating: The multi-scale features after fusion by the feature pyramid network are input into the detection head, and a hybrid expert module is introduced at the mid-level output of the detection head; the local edge intensity of the feature map is extracted using the Laplacian operator as the routing prior, and the gating weight is dynamically calculated for each pixel of the feature map based on the routing prior to activate the fine-grained expert branch, the mid-scale expert branch or the context expert branch, and the final mid-level detection features are output.

[0013] (5) Target Decoding and Output: The final mid-level detection features refined by the hybrid expert module, together with the high-level detection features in the multi-scale basic features, are input into the target detection decoder; interactive decoding is performed through a fixed number of object query vectors in the decoder, and finally the category probability and normalized position bounding box coordinates of the small target to be detected in the aerial image are output.

[0014] As a further improvement of the present invention, in step (1), the spatial-frequency dual-stream decoupling module includes a parallel spatial fidelity stream and a frequency-aware stream; the spatial fidelity stream uses depthwise separable convolution to preserve the spatial structure of the target; the frequency-aware stream uses an isotropic Scharr operator to construct a gradient filter to capture high-frequency components of the image; the output features of the two streams are concatenated and then subjected to pixel-level adaptive weight allocation through an attention unit that includes global channel pooling and spatial filtering.

[0015] As a further improvement of the present invention, in step (2), the low-level frequency enhancer adopts a dual-stream structure when extracting high-frequency texture features: the prior stream introduces a Log-Gabor filter to capture wideband texture information, and the parameters of the Log-Gabor filter are frozen; the adaptive stream uses a learnable depth convolution to extract specific texture patterns driven by data; the outputs of the two are concatenated and the space-to-depth transformation technique is performed to compress and fold the high-resolution features into the low-resolution channels.

[0016] As a further improvement of the present invention, in step (3), the calculation process of the semantic gated phase interaction module includes: discarding the amplitude spectrum information of the mid-level feature map after the detail enhancement, forcing the amplitude of the complex tensor to a constant 1, and retaining only the phase angle information to reconstruct the phase structure feature that is invariant to brightness; multiplying the phase structure feature by the semantic mask and calculating the mean of the spatial dimension to generate a bias vector that is injected into the key matrix of the cross-scale attention mechanism.

[0017] As a further improvement of the present invention, in step (4), the hybrid expert module includes three sets of parallel expert branches: the fine-grained expert branch uses deep convolution to process extremely small targets; the mesoscale expert branch uses dilated convolution to capture complete geometry; the context expert branch uses a bottleneck structure of downsampling and upsampling to help eliminate background false alarms; the gating network concatenates the spatial statistical features of global average pooling with the frequency domain edge statistical features extracted by the Laplacian operator, and generates pixel-level gating weights for the three expert branches through a multilayer perceptron.

[0018] The overall technical solution of this invention is achieved collaboratively through four innovative modules with a logically progressive relationship. First, this invention implements a spatial-frequency dual-stream decoupling strategy in the feature extraction stage of the backbone network. By constructing a spatial-frequency dual-stream decoupling module (GraceSFBlock), the input single-channel feature sequence is dynamically divided into parallel spatial fidelity streams and frequency-aware streams. In the spatial stream, lightweight depthwise separable convolutions are used to extract deep semantic information; while in the frequency-aware stream, this invention innovatively introduces a hard-coded Scharr gradient filter, which uses isotropic operators to pre-capture the high-frequency gradient components carried by extremely small targets in the image. The two feature streams are concatenated at the end and input into a global attention unit, where a nonlinear weighted fusion is performed using a dual attention mechanism of channel and spatial features, thereby establishing sensitivity to edge priors at the bottom layer of the backbone network.

[0019] To address the issue of low-level P2 feature maps being easily lost during aggressive downsampling in aerial images, this invention further constructs a Low-Level Frequency Enhancer (LLFE) to perform lossless folding operations. This module employs a dual-stream hybrid texture extraction architecture, where the prior stream utilizes a parameter-frozen Log-Gabor filter to capture broadband texture edges with multi-directional characteristics, while the adaptive stream supplements data-driven local patterns through learnable depthwise convolutions. To avoid information entropy loss caused by traditional pooling operations, this invention introduces a space-to-depth transformation technique. By periodically rearranging pixels in the spatial neighborhood to the channel dimension, this module can compress the spatial resolution to the P3 scale while losslessly "folding" high-frequency textures that would otherwise be lost into subsequent network layers through channel amplification, achieving physical-level preservation of low-level details of minute targets.

[0020] In the cross-scale feature interaction stage, to completely eliminate false alarm interference from bright backgrounds such as reflective metal and shimmering water surfaces in aerial photography scenes, this invention implements semantically gated cross-scale phase interaction (PhaseAIFI). This mechanism maps mid-level features to the complex frequency domain through a two-dimensional real-number fast Fourier transform and performs amplitude spectrum stripping. By forcibly normalizing the complex amplitude values ​​of the feature map to a constant 1, this invention retains only the phase spectrum information that reflects the geometric contours of the object for inverse transformation reconstruction. Since the phase spectrum has significant invariance to changes in illumination intensity, the reconstructed features exhibit extremely high structural purity. Subsequently, this invention uses a spatial semantic mask generated from high-level P5 features to multiply and modulate the reconstructed phase features, generating a semantically guided structural prior bias vector and injecting it into the key matrix of cross-scale attention, forcing the network to focus on the structural features of the real target in complex energy distributions.

[0021] Finally, in the feature refinement stage of the detection head, this invention deploys a hybrid expert module (P3ExpertBlock) based on frequency domain edge gating. This module dynamically schedules computational resources based on pixel-level edge strength priors and quickly calculates local frequency domain statistics at each location of the feature map using fixed Laplacian convolutional kernels. The gating network adaptively activates parallel fine-grained expert, mesoscale hole expert, and downsampling context expert branches based on routing weights generated from these statistics. This "location-specific" routing mechanism ensures that the network activates high-resolution fine-grained branches when processing sharp-edged target pixels, while focusing on context-aware branches when processing flat background regions. This significantly enhances the feature fitting depth of the detection head for multi-scale targets in aerial scenes while suppressing false alarms in the background.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0023] First, this invention fundamentally overcomes the technical bottleneck of feature vanishing in traditional downsampling paths for small aerial targets by implementing a collaborative mechanism of Low-Level Frequency Enhancer (LLFE) and Spatial-to-Depth (SPD) transformation. Because the LLFE module can accurately capture high-frequency edge information in the underlying P2 features and, in conjunction with the SPD transformation, performs sub-pixel-level spatial neighborhood pixel rearrangement, this invention successfully avoids the irreversible loss of information entropy caused by traditional pooling or stride convolution. This "lossless folding" technique can perfectly "revive" and compress the low-level details of tiny targets into the mid-level channel dimension without increasing the computational burden of high-resolution feature maps, greatly improving the model's recall rate for extremely small targets and solving the long-standing problem of tiny objects being easily submerged by noise in aerial detection.

[0024] Secondly, this invention fully utilizes the illumination invariance of image phase information by introducing a semantically gated phase interaction mechanism. By thoroughly eliminating amplitude spectrum information in the frequency domain that is highly susceptible to ambient light, metallic reflections, and wave interference, and by combining a high-level semantic mask to precisely guide the reconstructed pure phase structure features, this invention cuts off the interference path of bright background noise at the physical and mathematical logic level. This feature interaction mode, driven by structure rather than energy, enables the model to effectively eliminate false alarm signals caused by reflective roofs, reflective glass, etc., which are common in aerial photography scenes. Even in scenes with drastic changes in lighting or low contrast, it can still accurately anchor the geometric contours of the target, significantly reducing the false alarm rate.

[0025] Furthermore, the hybrid expert module (P3ExpertBlock) driven by frequency domain edge priors endows the network with excellent receptive field adaptation capabilities. By calculating the Laplacian edge intensity statistics in real time, this invention can achieve "location-specific" scheduling of computational resources based on the feature attributes of different pixel regions. In sharp-edged, structurally complex small target regions, fine-grained expert branches are automatically activated to preserve edge details, while in flat, homogeneous background regions, the focus is on activating contextual expert branches with a wider receptive field. This pixel-level dynamic routing mechanism perfectly matches the characteristics of drastic fluctuations in target scale and diverse backgrounds in aerial photography scenes, effectively improving the robustness of classification and localization tasks.

[0026] Finally, this invention achieves high accuracy while maintaining extremely high real-time processing performance through hard-coded prior design logic. Because this invention extensively incorporates mathematical prior knowledge such as the Scharr operator, Log-Gabor filter, and Laplace operator—which do not require parameter updates—in GraceSFBlock, LLFE, and expert routing, it not only provides the model with powerful physical inductive biases but also significantly reduces the size of learnable parameters in the network. This lightweight enhancement strategy based on classical signal processing theory enables the LPH-DETR model to significantly enhance frequency domain awareness capabilities while greatly reducing inference latency. This allows the system to be perfectly adapted to embedded edge devices (such as UAV onboard computing platforms), demonstrating significant engineering application value in complex real-time situational awareness tasks in the field. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0028] Figure 1 This is an overall architecture diagram of a target detection method based on phase sensing and frequency domain enhancement (LPH-DETR) according to the present invention.

[0029] Figure 2 This is a schematic diagram of the spatial-frequency dual-stream decoupling (GraceSFBlock) module provided by the present invention.

[0030] Figure 3 This is a schematic diagram of the Low-Level Frequency Enhancer (LLFE) module provided by the present invention.

[0031] Figure 4 This is a schematic diagram of the structure of the semantically gated cross-scale phase interaction (PhaseAIFI) module provided by the present invention.

[0032] Figure 5 This is a schematic diagram of the structure of the P3ExpertBlock hybrid expert module based on frequency domain edge gating provided by the present invention.

[0033] Figure 6 This is a flowchart of the LPH-DETR detection method of the present invention.

[0034] Figure 7 This is a flowchart of the aerial small target detection method of the present invention.

[0035] Figure 8This is a comparison chart of the detection performance of the target detection method based on phase sensing and frequency domain enhancement (LPH-DETR) and the baseline model (RT-DETR) in complex aerial photography scenarios provided in this embodiment of the invention.

[0036] Referring to Figure 1, the real-time aerial small target detection network (LPH-DETR) based on phase sensing and frequency domain enhancement provided by this invention mainly consists of four parts: a backbone network, a neck network, a head network, and a target decoder. The backbone network is used to detect small targets from the input aerial images. Multi-scale features are extracted and divided into four stages (Stage 1 to Stage 4), outputting downsampled feature maps. The neck network employs an improved PANet structure to fuse multi-scale features, resulting in fused features. The data is sent to the detection head, and finally the RT-DETR decoder outputs the target category and normalized bounding box coordinates.

[0037] 2. Specific implementation of the spatial-frequency dual-stream decoupling module (GraceSFBlock)

[0038] Referring to Figure 2, in the feature extraction stage, this invention divides a single feature channel into two: a spatial fidelity stream concatenates depthwise convolution and a pointwise convolution to extract deep semantics; a frequency-aware stream introduces a parameter-frozen Scharr operator as a hard-coded prior to capture high-frequency gradients. After concatenating the two feature streams, they are fed into a global attention unit (GAU), which then sequentially passes them through channel attention (based on global average pooling) and spatial attention (based on...). (Convolution) generates a joint weight mask. Finally, the module achieves adaptive fusion of spatial semantics and high-frequency details through dynamic weighting and residual connections.

[0039] 3. Specific implementation of the low-level frequency enhancer (LLFE)

[0040] Referring to Figure 3, in order to compensate for the structural degradation of small targets in deep networks, the present invention... and A low-level frequency enhancer is introduced at the junction of layers. The layer feature maps first enter the hybrid texture extraction module, employing a parallel dual-stream architecture: the prior stream uses a parameter-frozen Log-Gabor filter to capture high-frequency absolute texture edges; the adaptive stream uses learnable depthwise convolutions (DConv) to supplement data-driven texture patterns. Subsequently, an innovative spatial-to-depth transformation (SPD) is introduced, which, without losing pixels, halves the spatial dimension and quadruples the channel dimension of the feature maps, achieving a transformation to depth. Layers are seamlessly folded and spliced ​​together.

[0041] 4. Specific implementation of semantically gated cross-scale phase interaction (PhaseAIFI)

[0042] Referring to Figure 4, this module aims to eliminate interference from bright backgrounds in aerial photography. In the frequency domain purification path, for The mid-layer features undergo a two-dimensional real-number Fast Fourier Transform (rFFT2) to decouple the amplitude spectrum from the phase spectrum. This module completely discards the amplitude spectrum, which is susceptible to illumination (forced to be a constant of 1), and reconstructs the spatial structure features using only pure phase angle information. In the semantic mask path, High-level features are upsampled and a spatial probability map (semantic mask) ranging from 0 to 1 is generated. Finally, the pure phase structure features are multiplied element-wise with the semantic mask to generate a structural prior bias, which is then injected into the key matrix of cross-scale attention to achieve robust feature interaction.

[0043] 5. Specific implementation of the hybrid expert module (P3ExpertBlock) based on frequency domain edge gating

[0044] Referring to Figure 5, this module is deployed on a device that carries a small target. Output branch. The network employs parallel configurations of fine-grained experts, mesoscale hole experts, and downsampling context experts. Its core lies in a structure-aware gating network: it utilizes a hard-coded Laplacian operator to extract local edge intensity as frequency domain statistical features, combining this with spatial statistical features obtained through global average pooling, and inputting this into a multilayer perceptron (MLP) to dynamically generate routing weights. This allows for the automatic activation of fine-grained experts in high-frequency edge regions and the activation of context experts in flat, low-frequency regions, thereby refining multi-scale target features according to local conditions and eliminating false alarms.

[0045] P3ExpertBlock Module Structure Diagram

[0046] 6. Aerial Small Target Detection Method and Procedure

[0047] Referring to Figures 6 and 7, the present invention provides a real-time aerial small target detection method based on phase awareness and frequency domain enhancement. The complete detection process includes: First, acquiring the aerial image to be detected and inputting it into a backbone network containing a spatial-frequency dual-stream decoupling module to extract multi-scale features; second, inputting the low-level feature map into a low-level frequency enhancer to extract high-frequency textures and performing lossless folding and fusion using spatial-to-depth transformation; next, through a semantically gated phase interaction module, stripping the amplitude spectrum of the middle-level feature map and multiplying the pure phase features with a high-level semantic mask to complete cross-scale noise reduction interaction; subsequently, introducing a hybrid expert module at the output of the detection head to dynamically allocate routing weights using frequency domain edge strength priors to refine the features; finally, inputting the refined features into a decoder to output the small target's category and bounding box.

[0048] 7. Comparative verification of detection results and manifestation of beneficial effects

[0049] Referring to Figure 8, this invention demonstrates the visual comparison between the method of this embodiment and a benchmark model (such as RT-DETR) in a real complex aerial photography scenario, thus intuitively reflecting the technical advancement of this invention.

[0050] In the baseline model detection results shown on the left side of Figure 8, due to the excessive reliance of the traditional self-attention mechanism on the feature energy amplitude, the network generates obvious false alarm bounding boxes (such as "false alarms in reflective areas" as shown by the dashed circle) when facing strong light reflection on the top of buildings or specular reflection areas on metal surfaces. At the same time, due to the loss of the underlying P2 information caused by the traditional downsampling strategy, the baseline model has a large area of ​​missed detection when dealing with extremely small vehicles located in shadow or edge areas.

[0051] In contrast, as shown on the right side of Figure 8, the LPH-DETR method provided by this invention exhibits significantly improved detection performance in the same scene. Through the phase noise reduction mechanism of the PhaseAIFI module, the system successfully ignores the energy interference of bright backgrounds and eliminates false alarms in reflective areas. Simultaneously, thanks to the lossless reconstruction of the underlying texture by the LLFE module, tiny targets that were previously missed in the baseline model are accurately anchored and correctly classified. This comparative result strongly supports the significant technical contribution of the spatial-frequency dual-stream collaboration and phase-aware interaction described in this invention in improving recall and reducing false alarm rates. Detailed Implementation

[0052] like Figure 1-8 As shown, this invention provides a real-time aerial small target detection method based on phase sensing and frequency domain enhancement, comprising the following steps:

[0053] (1) Multi-scale feature extraction: The aerial image to be detected is input into the feature extraction backbone network, which includes a spatial-frequency dual-stream decoupling module; the spatial structure information and high-frequency gradient information of the aerial image are extracted by the spatial-frequency dual-stream decoupling module respectively, and dynamic weighted fusion is performed to output a multi-scale basic feature map, which includes at least a bottom-level feature map, a middle-level feature map and a high-level feature map;

[0054] (2) Lossless restoration of low-level texture: The low-level feature map is input to the low-level frequency enhancer to extract the high-frequency texture features of the low-level feature map, and the spatial neighbor pixels of the high-frequency texture features are rearranged to the channel dimension using the spatial-to-depth transformation technique. Then, it is spliced ​​and fused with the middle-level feature map to obtain the middle-level feature map after detail enhancement.

[0055] (3) Semantic gated phase cross-scale interaction: The high-level feature map is upsampled and a semantic mask is generated through the semantic gated phase interaction module; at the same time, Fourier transform is performed on the mid-level feature map after detail enhancement to separate the amplitude spectrum and phase spectrum, and only the phase spectrum is used for inverse transformation to obtain the phase structure feature; the phase structure feature is multiplied with the semantic mask to suppress high-frequency background noise and obtain the purified cross-scale interaction feature;

[0056] (4) Expert feature refinement based on frequency domain edge gating: The multi-scale features after fusion by the feature pyramid network are input into the detection head, and a hybrid expert module is introduced at the mid-level output of the detection head; the local edge intensity of the feature map is extracted using the Laplacian operator as the routing prior, and the gating weight is dynamically calculated for each pixel of the feature map based on the routing prior to activate the fine-grained expert branch, the mid-scale expert branch or the context expert branch, and the final mid-level detection features are output.

[0057] (5) Target Decoding and Output: The final mid-level detection features refined by the hybrid expert module, together with the high-level detection features in the multi-scale basic features, are input into the target detection decoder; interactive decoding is performed through a fixed number of object query vectors in the decoder, and finally the category probability and normalized position bounding box coordinates of the small target to be detected in the aerial image are output.

[0058] In step (1), the spatial-frequency dual-stream decoupling module includes a parallel spatial fidelity stream and a frequency-aware stream; the spatial fidelity stream uses depthwise separable convolution to preserve the spatial structure of the target; the frequency-aware stream uses an isotropic Scharr operator to construct a gradient filter to capture high-frequency components of the image; the output features of the two streams are concatenated and then subjected to pixel-level adaptive weight allocation through an attention unit that includes global channel pooling and spatial filtering.

[0059] In step (2), the low-level frequency enhancer uses a dual-stream structure when extracting high-frequency texture features: the prior stream introduces a Log-Gabor filter to capture wideband texture information, and the parameters of the Log-Gabor filter are frozen; the adaptive stream uses a learnable depth convolution to extract specific texture patterns driven by data; the outputs of the two are concatenated and the spatial-to-depth transformation technique is performed to compress and fold the high-resolution features into the low-resolution channels.

[0060] In step (3), the calculation process of the semantic gated phase interaction module includes: discarding the amplitude spectrum information of the mid-level feature map after the detail enhancement, forcing the amplitude of the complex tensor to a constant 1, and retaining only the phase angle information to reconstruct the phase structure feature that is invariant to brightness; multiplying the phase structure feature by the semantic mask and calculating the mean of the spatial dimension to generate a bias vector that is injected into the key matrix of the cross-scale attention mechanism.

[0061] In step (4), the hybrid expert module includes three sets of parallel expert branches: the fine-grained expert branch uses deep convolution to process extremely small targets; the mesoscale expert branch uses dilated convolution to capture complete geometry; the context expert branch uses a bottleneck structure of downsampling and upsampling to help eliminate background false alarms; the gating network concatenates the spatial statistical features of global average pooling with the frequency domain edge statistical features extracted by the Laplacian operator, and generates pixel-level gating weights for the three expert branches through a multilayer perceptron.

[0062] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the core mathematical formulas and operator definitions in the embodiments of the present invention.

[0063] 1. Specific implementation of the spatial-frequency dual-stream decoupling module (GraceSFBlock)

[0064] Referring to [fig:gracesf_module], the spatial-frequency dual-stream decoupling module provided by this invention has its core logic in synchronously extracting deep semantics and underlying physical edges through parallel heterogeneous branches. In a specific embodiment of this invention, the processing flow of this module is as follows:

[0065] Feature Dimension Preprocessing and Decoupling Segmentation First, given the input feature map... Its dimension is represented as ,in For batch size, For the number of channels, and These are the height and width of the feature map, respectively. Preferably, it is first passed through a... Pointwise convolution is used to increase the dimensionality, expanding the number of channels to the original value. Times (in this embodiment, Subsequently, the expanded features are uniformly divided into two sub-features along the channel dimension: spatial flow features. and frequency flow characteristics This explicit channel splitting ensures that in subsequent processing, some neurons focus on learning spatial geometric semantics, while others focus on capturing frequency domain gradient responses.

[0066] Basic Dimensions and Backbone Network Symbol X: The input feature map of the module, whose dimensions are represented as B × C × H × W.

[0067] B: Batch Size; C: Number of Channels in the Feature Map; H, W: Spatial Height and Width of the Feature Map; k: Channel Augmentation Coefficient; P2, P3, P5: Feature Layers in the Backbone Network Corresponding to Different Downsampling Rates.

[0068] 1. Semantic extraction of spatially faithful streams for spatial stream features This invention employs a depthwise separable convolutional architecture to balance computational efficiency and performance. Specifically, it first utilizes a kernel size of... Depthwise Convolution This captures local spatial context information without changing the number of channels. Then, through a... Pointwise convolution ( This enables cross-channel information interaction and ultimately extracts structural features with strong semantic expressive power. Its mathematical expression is:

[0069]

[0070] High-frequency edge capture of frequency-aware stream for frequency stream features This invention employs a hard-coded Scharr gradient filter. In this embodiment, the Scharr operator, as an isotropic differential operator, has its parameters fixed during training (i.e., it does not participate in gradient updates). Through the level ( ) and vertical ( Performing convolution operations in both directions can accurately extract edge transition information of tiny targets in an image, generating high-frequency edge components. The introduction of this physical prior effectively compensates for the lack of perception of subtle contours in deep networks after multiple downsampling.

[0071] Adaptive feature fusion based on Global Attention Unit (GAU) yields spatial features. and frequency characteristics Then, the two features are concatenated along the channel dimension to obtain the fused candidate features. To achieve deep complementarity between the two types of heterogeneous information, this invention employs a Global Attention Unit (GAU) for dynamic weighting:

[0072] Channel attention mechanism: First, for Global average pooling (GAP) is performed to extract global spatial statistics. Subsequently, two fully connected layers (with weight matrices respectively) are used. and ) and SiLU activation function ( Generate channel-dependent weights :

[0073]

[0074] Spatial attention mechanism: After being applied to the feature map, using a Large-kernel convolutional layers extract the relationships between local spatial features, and then pass them through a sigmoid function ( Generate spatial attention mask :

[0075]

[0076] Finally, the module combines the residual output with the feature fusion, which has undergone double attention weighting, and the original input features. Residual connections are performed to ensure that the original semantic flow of the backbone network is not disrupted while introducing frequency domain enhancement information. The final output features... Represented as:

[0077]

[0078] in, This indicates element-wise multiplication. This module significantly improves the model's accuracy in extracting key structures of small targets in aerial photography scenes by decoupling and reconstructing spatial and frequency data.

[0079] Tfid and Tfreq: respectively, spatial fidelity flow carrier characteristics and frequency-sensing flow carrier characteristics after segmentation;

[0080] FS: Structural features with strong semantic expressive power extracted from spatially faithful streams;

[0081] ConvDW: Depthwise Convolution.

[0082] ConvPW: Pointwise Convolution.

[0083] GAP: Global Average Pooling.

[0084] W1, W2: Learnable weight matrices of the fully connected layer;

[0085] δ and σ: are the SiLU activation function and the Sigmoid activation function, respectively;

[0086] ⊗: Element-wise multiplication.

[0087] 2. Specific Implementation of Low-Level Frequency Enhancer (LLFE)

[0088] Referring to Figure 3, in aerial image detection tasks, due to the extremely small size of the target (usually smaller than...), The key texture features of a pixel are mainly concentrated in the low-level feature maps of the backbone network (such as the P2 layer). However, traditional downsampling operations (such as convolution or pooling with a stride of 2) cause severe signal aliasing and loss of detail. To address this, this invention implements a Low-Level Frequency Enhancer (LLFE), which achieves deep mining and preservation of low-level features through physical prior guidance and sub-pixel lossless folding technology. The specific implementation steps are as follows:

[0089] In this embodiment, the LLFE module first captures the high-frequency responses of the input P2 layer features in multiple directions. Unlike traditional self-learning convolution, this invention introduces a Log-Gabor filter with fixed parameters as the prior flow.

[0090] Specifically, the Log-Gabor filter possesses a Gaussian transfer function on the logarithmic frequency scale, overcoming the low-frequency over-response problem caused by the DC component in traditional Gabor filters. Its transfer function in the frequency domain... Represented as radial component Angular component The product of . In the spatial domain, its impulse response function. Defined as:

[0091]

[0092] in, Indicates the center wavelength of the filter. and These determine the spatial bandwidth of the filter in the horizontal and vertical directions, respectively. The direction angle of the filter. This is the phase shift. (Rotated coordinates): Represents the original pixel coordinates After the direction angle The mathematical definition of the rotated coordinates is: . It is a Gaussian window that determines the degree of spatial localization of the filter, that is, how much of the image the filter "sees". (Oscillatory component): Represents the frequency characteristics of the filter, responsible for generating a response in a specific direction to extract edges.

[0093] By setting multiple scales (e.g., 3) and multiple directions (e.g., 4) of Log-Gabor kernels, this module can accurately capture the absolute edges of tiny targets in aerial images. In a preferred embodiment, the Log-Gabor branches do not participate in backpropagation updates, thereby ensuring that the extracted features have stable physical meaning.

[0094] To compensate for the limitations of fixed operators in data adaptability, LLFE incorporates a parallel adaptive feature stream and hybrid feature fusion. This branch utilizes learnable features... Depthwise convolution extracts local statistical features. Subsequently, high-frequency edge features extracted from the prior flow are... Dynamic features extracted with adaptive flow Perform channel splicing:

[0095]

[0096] in The SiLU activation function is used. This design achieves a complementary advantage between "physical rules" and "data-driven" approaches, resulting in a hybrid output feature. It possesses both clear edge orientation and rich semantic expressiveness. Fprior and Fadapt: ​​These are edge prior features and adaptive features extracted based on physical operators, respectively. BN refers to Batch Normalization, which aims to force a set of feature data to transform it into a set with a mean of 1 / 2. variance is The standard distribution of .

[0097] Lossless Feature Folding Based on Spatial-to-Depth Transformation (SPD)

[0098] When fusing the enhanced features of layer P2 to layer P3, this invention abandons the pooling operation that may cause feature vanishing and introduces a space-to-depth transformation.

[0099] The specific operation logic is as follows: For an input size of... Feature map SPD conversion according to period (Corresponding to a downsampling factor of 2) Pixel rearrangement. For example, rearranging spatially adjacent pixels... The four pixels in the pixel block are determined by their relative coordinates within the block. They were assigned to different channels.

[0100] The rearranged feature map dimension becomes This physical location permutation operation reduces spatial resolution while achieving "lossless folding" of the original low-level features. Finally, this folded feature is concatenated with the original P3 layer features of the backbone network, and then... Convolution performs channel compression and interaction to generate the final fused features. :

[0101]

[0102] SPD: Spatial to Depth Transform operation, used to achieve lossless downsampling. Through the above implementation method, the present invention successfully "revives" subpixel-level texture information that originally disappeared in the deep network at the P3 scale.

[0103] 3. Specific Implementation of Semantically Gated Cross-Scale Phase Interaction (PhaseAIFI)

[0104] Referring to Figure 4, aerial images in complex environments (such as direct sunlight, water reflection, and metal roof reflection) generate extreme high-brightness noise. This noise manifests as extremely high response amplitudes in the spatial domain feature map, thus misleading traditional self-attention mechanisms. To address this issue, this invention implements a cross-scale interaction mechanism based on frequency domain phase awareness (Semantic-Guided PhaseAIFI), utilizing the illumination invariance of phase information to purify the target's geometric structure. The specific implementation steps are as follows:

[0105] Frequency domain component extraction based on Fast Fourier Transform: First, this module receives the mid-layer feature map from the backbone network. In order to analyze the structural features of the target in the frequency domain, this embodiment... Each channel performs a two-dimensional real fast Fourier transform (rFFT2) independently:

[0106]

[0107] in, : Refers to the mid-level feature maps from the backbone network, typically represented by the following dimension. . : Represents the discrete coordinates of the feature map in the spatial domain The specific pixel value at that location. and : Represent feature maps respectively The space's height and width. and : Discrete coordinate index of the spatial domain, used to traverse every pixel of the image. and : Coordinate indices in the frequency domain, corresponding to the frequency components in the horizontal and vertical directions, respectively. The frequency domain complex tensor obtained through transformation calculation contains all the frequency information of the image. The imaginary unit, in its mathematical definition, satisfies... . : Natural constant, used as the base of complex exponential functions in calculations. Pi (π) is used to define the period of a complex exponential waveform. According to Euler's formula, this complex tensor can be further decomposed into an amplitude spectrum. and phase spectrum :

[0108]

[0109] It is the amplitude spectrum, which carries the energy intensity of image features in the frequency domain. Physically, it directly corresponds to the light intensity of each pixel in the image. The specific calculation formula is as follows: This involves calculating the modulus of a complex number, which reflects the energy contribution of that frequency component. The phase spectrum calculation formula is as follows: It is the phase spectrum, which carries the core structural information of objects in the image, including geometric shape, edge details, and the relative positions of each target. : Represents the real part of the complex tensor in the frequency domain. : Represents the imaginary part of a complex tensor in the frequency domain. Modulo operation: used to calculate the length of a complex vector. The arctangent function is used to determine the corresponding phase angle based on the ratio of the real to the imaginary parts. To eliminate "energy interference" from the highlighted background, this invention employs a radical frequency domain purification strategy: explicitly stripping the amplitude spectrum and retaining only the phase spectrum.

[0110] The specific operation is as follows: Construct a unit amplitude matrix consisting entirely of 1s. and compared it with the original phase spectrum By combining these, a set of pure phase complex tensors is reconstructed. Subsequently, the inverse fast Fourier transform (iFFT2) is used to map it back to the spatial domain to obtain pure structural features. :

[0111]

[0112] : Represents the two-dimensional real-valued inverse fast Fourier transform (iFFT2). This operation is responsible for converting complex data in the frequency domain back to pixel features in the spatial domain. This is a pure phase complex tensor constructed before the inverse transformation. : Represents the result after amplitude spectrum stripping. Because The reconstruction process discards the original energy distribution information. It has natural robustness (i.e., illumination invariance) to light fluctuations, shadow occlusion and reflection interference in aerial photography scenes. It can clearly outline the geometric contours of small targets, thereby suppressing bright spots in the background.

[0113] Cross-scale semantic gating and bias vector injection

[0114] While phase features offer good structural representation, they lack explicit categorical semantics. Therefore, this invention introduces high-level features. As a gating signal.

[0115] In specific implementation, firstly Upsampled to the same level using bilinear interpolation. Same spatial resolution, and utilizing Convolution generates spatial semantic masks Subsequently, and Element-wise multiplication is performed to filter out residual components belonging to the background in the phase features. Finally, a multi-channel structure bias vector is generated using global average pooling (GAP). :

[0116]

[0117] : Structural prior bias vector. The structural feature map, reconstructed from pure phase information, carries the contour and edge information of objects in the image. Spatial semantic mask, composed of high-level features (such as...) This is generated to filter out residual components in the phase features that belong to the background. : Element-wise multiplication operation, used to perform gated modulation of phase features using semantic masks. Global average pooling is used to compress statistical features in spatial dimensions to generate multi-channel statistics. : Convolutional layers are used to achieve cross-channel interaction and linear mapping of features. : Activation function, which maps the output value to The interval serves as a dynamic weight bias. During the cross-scale attention interaction phase, this invention injects [the bias] into the key matrix. During the generation process, the corrected attention calculation formula is:

[0118]

[0119] : These represent the query, key, and value matrices in the cross-scale self-attention mechanism, respectively. : Structure enhancement operator, by Injected into the key matrix The similarity weight allocation is corrected during the generation process. : scaling factor, where The dimension of the key / query vector is used to prevent gradient vanishing due to excessively large dot product results. T is the transpose of the matrix. : Normalization function, which transforms attention scores into a probability distribution.

[0120] This injection mechanism allows the Query vector (target query item) to be explicitly guided to the target area with a real geometric structure when matching the Key vector, thus fundamentally avoiding false alarm interference caused by the highlighted background.

[0121] 4. Specific Implementation of the Hybrid Expert Module (P3ExpertBlock) Based on Frequency Domain Edge Gating

[0122] Referring to Figure 5, in the feature output stage of the detection head, due to the significant scale uncertainty in the distribution of aerial targets in the image (from vehicles of a few pixels to building outlines of hundreds of pixels), it is difficult to simultaneously achieve fine-grained edge extraction and global context suppression using a fixed convolutional receptive field. Therefore, this invention implements a frequency-gated mixture of experts (P3ExpertBlock) based on frequency domain edge gating, which achieves dynamic refinement of multi-scale features through explicit edge strength priors. The specific implementation steps are as follows:

[0123] To quantify the structural complexity of each region in the feature map, this module first introduces a frequency domain statistic extraction method based on the Laplacian operator. Laplacian convolution kernel The Laplacian operator, as a second-order differential operator, is highly sensitive to abrupt changes in grayscale in an image (i.e., edges and small targets).

[0124] Specifically, for input features Perform a depthwise convolution operation to obtain a high-frequency gradient feature map. Subsequently, global high-frequency statistics for this region are extracted using global average pooling (GAP). :

[0125]

[0126] Frequency domain edge statistical characteristics obtained based on Laplace response calculation;

[0127] At the same time, this module simultaneously extracts the conventional spatial domain global average pooling statistics. By splicing and A binary prior feature vector with both frequency domain intensity and spatial mean is constructed.

[0128] In this embodiment, the hybrid expert system is configured with three groups of expert branches with different receptive field characteristics in parallel:

[0129] Fine-grained expert ( ):use Dense convolutions are used to focus on capturing the sharp edges of tiny objects.

[0130] Mesoscale experts ( ):use Furthermore, dilated convolution with a dilation rate of 2 aims to balance local details with the geometric outline of medium-sized targets.

[0131] Contextual expert ( ): A bottleneck structure that first downsamples and then upsamples is used to obtain broad contextual information to identify large landmarks or suppress large areas of false alarms in the background.

[0132] Dynamic Gated Routing and Feature Adaptive Fusion: This invention introduces a lightweight gating network, which consists of a multilayer perceptron (MLP) with two fully connected layers. Based on the input prior feature vector, the gating network uses the Softmax function to calculate the soft voting weights of the three expert branches. ,in Wi: The dynamic routing weight coefficient assigned to the i-th expert branch;

[0133] Final feature output A residual learning framework is adopted, and a learnable scaling factor is introduced. This is used to adjust the contribution of expert enhancement terms. Its mathematical expression is:

[0134]

[0135] : The final refined feature map output by the hybrid expert module. SiLU activation function. : Batch normalization layer. : The original feature map input to this module. The sigmoid activation function maps the scaling factor to... Interval. : A learnable residual scaling factor used to adaptively adjust the contribution of expert augmentations to the final features during training. : Convolutional layer. : Sum the results of the three heterogeneous expert branches. : Assigned to the The dynamic routing weight coefficients of each expert branch are generated in real time by the gated network based on frequency domain edge priors. : No. Each expert branch is paired with the input. The processing results. Among them... As a fine-grained expert, A mesoscale expert. As a context expert, this "site-specific" allocation of computing resources significantly enhances LPH-DETR's fitting ability when processing multi-scale aerial targets.

[0136] This invention is not limited to the above embodiments. Based on the technical solutions disclosed in this invention, those skilled in the art can make some substitutions and modifications to some of the technical features without creative effort, and all such substitutions and modifications are within the protection scope of this invention.

Claims

1. A real-time aerial small target detection method based on phase sensing and frequency domain enhancement, characterized in that, Includes the following steps: (1) Multi-scale feature extraction: The aerial image to be detected is input into the feature extraction backbone network, which includes a spatial-frequency dual-stream decoupling module; the spatial structure information and high-frequency gradient information of the aerial image are extracted by the spatial-frequency dual-stream decoupling module respectively, and dynamic weighted fusion is performed to output a multi-scale basic feature map, which includes at least a bottom-level feature map, a middle-level feature map and a high-level feature map; (2) Lossless restoration of low-level texture: The low-level feature map is input to the low-level frequency enhancer to extract the high-frequency texture features of the low-level feature map, and the spatial neighbor pixels of the high-frequency texture features are rearranged to the channel dimension using the spatial-to-depth transformation technique. Then, it is spliced ​​and fused with the middle-level feature map to obtain the middle-level feature map after detail enhancement. (3) Semantic gated phase cross-scale interaction: The high-level feature map is upsampled and a semantic mask is generated through the semantic gated phase interaction module; at the same time, Fourier transform is performed on the mid-level feature map after the detail enhancement to separate the amplitude spectrum and the phase spectrum, and the phase structure features are obtained by inverse transformation using only the phase spectrum. The phase structure features are multiplied by the semantic mask to suppress high-frequency background noise and obtain purified cross-scale interaction features. (4) Expert feature refinement based on frequency domain edge gating: The multi-scale features after fusion by the feature pyramid network are input into the detection head, and a hybrid expert module is introduced at the mid-level output of the detection head; the local edge intensity of the feature map is extracted using the Laplacian operator as the routing prior, and the gating weight is dynamically calculated for each pixel of the feature map based on the routing prior to activate the fine-grained expert branch, the mid-scale expert branch or the context expert branch, and the final mid-level detection features are output. (5) Target Decoding and Output: The final mid-level detection features refined by the hybrid expert module, together with the high-level detection features in the multi-scale basic features, are input into the target detection decoder; interactive decoding is performed through a fixed number of object query vectors in the decoder, and finally the category probability and normalized position bounding box coordinates of the small target to be detected in the aerial image are output.

2. The real-time aerial small target detection method based on phase sensing and frequency domain enhancement according to claim 1, characterized in that, In step (1), the spatial-frequency dual-stream decoupling module includes a parallel spatial fidelity stream and a frequency-aware stream; the spatial fidelity stream uses depthwise separable convolution to preserve the spatial structure of the target; the frequency-aware stream uses an isotropic Scharr operator to construct a gradient filter to capture high-frequency components of the image; the output features of the two streams are concatenated and then subjected to pixel-level adaptive weight allocation through an attention unit that includes global channel pooling and spatial filtering.

3. The real-time aerial small target detection method based on phase sensing and frequency domain enhancement according to claim 1, characterized in that, In step (2), the low-level frequency enhancer uses a dual-stream structure when extracting high-frequency texture features: the prior stream introduces a Log-Gabor filter to capture wideband texture information, and the parameters of the Log-Gabor filter are frozen; the adaptive stream uses a learnable depth convolution to extract specific texture patterns driven by data; the outputs of the two are concatenated and the spatial-to-depth transformation technique is performed to compress and fold the high-resolution features into the low-resolution channels.

4. The real-time aerial small target detection method based on phase sensing and frequency domain enhancement according to claim 1, characterized in that, In step (3), the calculation process of the semantic gated phase interaction module includes: discarding the amplitude spectrum information of the mid-level feature map after the detail enhancement, forcing the amplitude of the complex tensor to a constant 1, and retaining only the phase angle information to reconstruct the phase structure feature that is invariant to brightness; multiplying the phase structure feature by the semantic mask and calculating the mean of the spatial dimension to generate a bias vector that is injected into the key matrix of the cross-scale attention mechanism.

5. The real-time aerial small target detection method based on phase sensing and frequency domain enhancement according to claim 1, characterized in that, In step (4), the hybrid expert module includes three sets of parallel expert branches: the fine-grained expert branch uses deep convolution to process extremely small targets; the mesoscale expert branch uses dilated convolution to capture complete geometry; the context expert branch uses a bottleneck structure of downsampling and upsampling to help eliminate background false alarms; the gating network concatenates the spatial statistical features of global average pooling with the frequency domain edge statistical features extracted by the Laplacian operator, and generates pixel-level gating weights for the three expert branches through a multilayer perceptron.

Images