An unmanned aerial vehicle target detection method based on frequency domain enhancement and state space modeling

Abstract

The application provides a UAV target detection method based on frequency domain enhancement and state space modeling, and relates to the technical field of UAV target detection. First, the application acquires a to-be-detected image and inputs the DFS-DETR model constructed based on RT-DETR; the backbone network of C2f-WDMamba is adopted to perform feature extraction, so that backbone features are obtained; an AIFI encoder of WAS-FFN is introduced to perform coding, so that coded features are obtained; a cross-scale fusion structure of CFGN is adopted to perform fusion, so that fused features are obtained; the fused features and the coded features are input into a decoder and a detection head to perform prediction, so that a target detection bounding box and a target classification result are output. The application can improve the small target detection effect in a complex UAV aerial scene.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) target detection technology, and in particular to a UAV target detection method based on frequency domain enhancement and state space modeling. Background Technology

[0002] With the rapid development of drone technology, it has been widely applied in fields such as security patrol, disaster monitoring, and traffic management. Image sensors on drones can detect ground targets from the air, providing new methods for rapid localization and identification. Meanwhile, real-time target detection algorithms, as an important research area in computer vision, have been widely used in drone aerial photography. However, due to limitations in flight altitude, imaging resolution, and shooting angle, small targets (such as pedestrians and vehicle parts) captured on drone platforms often occupy a very small area at the pixel level and are easily interfered with by the background, posing a significant challenge to subsequent detection and tracking. Currently, mainstream target detection algorithms rely on manually designed components, such as non-maximum suppression (NMS) and anchor boxes generated based on professional knowledge. These components require extensive adjustments for different tasks. For example, NMS is sensitive to thresholds and may incorrectly suppress highly overlapping or adjacent objects; anchor box parameters need to be adjusted differently for specific datasets, lacking generalization ability. Therefore, researching efficient, accurate, and real-time drone small target detection algorithms is not only of significant academic value but also crucial for practical applications such as smart cities, security monitoring, and emergency response.

[0003] As an end-to-end detection model, DETR abandons redundant operations like NMS, directly inputting the original image and outputting the object detection result. Recent research has improved the small object detection capability of the DETR model, but due to its high computational cost and poor real-time performance, it is not suitable for real-time scenarios. To address these issues, Zhao et al. proposed RT-DETR, the first end-to-end framework to surpass the YOLO model in real-time conditions. RT-DETR is an object detection model that combines the advantages of the efficient YOLO framework and the DETR architecture, aiming to achieve real-time high-precision object detection. Its advantage is attributed to its innovative hybrid encoder design, which performs feature interactions across multiple scales to improve the algorithm's ability to detect objects of different sizes.

[0004] Although RT-DETR performs well on general benchmarks such as COCO, its architecture is not inherently optimized for aerial image problems. When facing complex real-world scenarios, especially small targets, multiple targets, occlusion, and severe background interference, its performance still suffers from the following bottlenecks: 1) Insufficient feature extraction capability. Although RT-DETR has made trade-offs in speed, its backbone mainly uses ResNet for feature extraction, lacking the ability to extract deeper semantic information; 2) Lack of an effective cross-scale and contextual information integration mechanism. When facing targets with drastic scale changes, the original model does not make full use of multi-scale information for contextual modeling, resulting in weak detection capabilities for distant targets, small objects, or objects in complex backgrounds. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a UAV target detection method based on frequency domain enhancement and state space modeling. By introducing C2f-WDMamba, WAS-FFN and CFGN into RT-DETR, the three stages of backbone feature extraction, AIFI encoding and cross-scale fusion are formed into a collaborative processing chain, thereby improving the ability to preserve small target details and the effect of multi-scale feature integration in complex UAV aerial photography scenarios.

[0006] To achieve the above objectives, the present invention provides the following solution: A UAV target detection method based on frequency domain enhancement and state space modeling includes: Acquire the image to be detected and input it into the DFS-DETR model built based on RT-DETR; The backbone network in the DFS-DETR model is used to extract features from the image to be detected, resulting in backbone features. The backbone network adopts C2f-WDMamba, which processes the input features through dynamic dilatation routing, multi-level wavelet high-frequency enhancement, and state space modeling. The backbone features are input into the AIFI encoder in the DFS-DETR model for encoding to obtain encoded features. The AIFI encoder incorporates WAS-FFN, which uses windowed frequency domain modeling, channel reweighting, and learnable frequency band weights to process the backbone features. The encoded features are input into the cross-scale fusion structure in the DFS-DETR model for fusion to obtain fused features; the cross-scale fusion structure adopts CFGN, which includes C2FM and SPGF. CFGN fuses features of different scales through a central enhancement module and a scale-aware gating mechanism. The fused features and encoded features are input into the decoder and detection head for prediction, and the output is the target detection bounding box and the target classification result.

[0007] Preferably, the backbone network in the DFS-DETR model is used to extract features from the image to be detected, resulting in backbone features, including: The input features corresponding to the image to be detected are segmented into channels to obtain the first branch features and the second branch features; The second branch features are input into the stacked WDMamba Blocks for processing to obtain enhanced features; The first branch features and the enhanced features are concatenated, and the output channels are compressed through convolution to obtain the backbone features.

[0008] Preferably, the second branch features are input into the stacked WDMamba Block for processing to obtain enhanced features, including: Perform depthwise convolution and batch normalization operations on the features of the second branch; A lightweight feedforward network is used to process the features after depthwise convolution and batch normalization operations. The features processed by the lightweight feedforward network are input into the MSFM structure; The output of the MSFM structure is further processed by deep convolution, batch normalization, and feedforward network, and the output features are enhanced through residual connections.

[0009] Preferably, the features processed by the lightweight feedforward network are input into the MSFM structure, including: The input channel is split into a global channel, a local channel, and an identity channel; In the global channel, MBWTConv2d is used to perform dynamic expansion routing and multi-layer wavelet high-frequency enhancement and IWT reconstruction. The features after multi-layer wavelet high-frequency enhancement and IWT reconstruction are modeled in state space by Mamba and modulated with FiLM to obtain the global branch output. The outputs of the global branch, the local channels, and the identity channels are concatenated and then projected through an activation function and PWConv to obtain the output of the MSFM structure.

[0010] Preferably, MBWTConv2d is used in the global channel to perform dynamic dilation routing and multi-layer wavelet high-frequency enhancement and IWT reconstruction, including: Multi-scale adaptive convolution kernel selection is achieved through dynamic dilated convolution. Low-frequency LL and high-frequency HF are obtained by wavelet transform WT decomposition; The highest-level high-frequency HF is subjected to dual enhancement and mixing treatment; The processed high-frequency HF and low-frequency LL are reconstructed by inverse wavelet transform (IWT). A small target mask is constructed based on the gradient magnitude, and a soft mask is generated through adaptive Top-K selection; Soft Mask is used to generate weighted values ​​for high-frequency enhancement.

[0011] Preferably, the backbone features are input into the AIFI encoder in the DFS-DETR model for encoding to obtain encoded features, including: Set residual branches for the main features; The main features are sequentially subjected to dimensionality increase and depthwise convolution. Windowed frequency domain modeling, local fast Fourier transform, and spectral splitting are performed on the features after deep convolution. Learnable gain modulation is applied to the split high-frequency and low-frequency bands; Channel attention and convolution dimensionality reduction are performed on the features after inverse fast Fourier transform and window restoration, and then fused with residual branches to obtain encoded features.

[0012] Preferably, learnable gain modulation is performed on the split high-frequency band and low-frequency band, including: Generate a high-frequency mask for the high-frequency band and a low-frequency mask for the low-frequency band; Two sets of learnable gain parameters are used to modulate the high-frequency band and the low-frequency band respectively; Perform inverse fast Fourier transform and window restoration on the modulated spectrum.

[0013] Preferably, the encoded features are input into the cross-scale fusion structure of the DFS-DETR model for fusion to obtain fused features, including: Perform feature mapping to reduce the dimensionality of the encoded features; The reduced-dimensional features are subjected to dual-branch modulation with center enhancement and background suppression to obtain the modulated features. The modulation features are subjected to pyramid pooling, cross-scale similarity calculation and context aggregation, and then refined by a local enhancement module to obtain high-level fusion features; Extract low-level features from encoded features; Gated fusion of high-level and low-level features is performed to obtain fused features.

[0014] Preferably, the dimensionality-reduced features are subjected to dual-branch modulation with center enhancement and background suppression to obtain modulated features, including: The center enhancement module processes the dimensionality-reduced features to obtain the center enhancement mask; The background suppression mask is obtained by processing the dimensionality-reduced features through global average pooling. Modulated features are obtained by modulating the dimensionality-reduced features based on the center enhancement mask and the background suppression mask; Calculate key features, value features, and query features for modulation features; Cross-scale similarity calculation is performed based on key features and query features, and context aggregation is performed based on value features.

[0015] Preferably, gated fusion is performed on the high-level fusion features and the low-level features to obtain fused features, including: Construct low-level spatial structure branches and high-level semantic branches using SPGF; After interpolating the high-level semantic branches, they are spliced ​​with the low-level spatial structure branches. Convolutional fusion is performed on the concatenated features; The convolutionally fused features are subjected to gated fusion modulation to obtain fused features.

[0016] The present invention discloses the following beneficial effects: This invention constructs a DFS-DETR model based on RT-DETR and employs C2f-WDMamba in the backbone network. Input features are processed through dynamic dilated routing, multi-level wavelet high-frequency enhancement, and state-space modeling. Simultaneously, WAS-FFN is introduced into the AIFI encoder, utilizing windowed frequency domain modeling, channel reweighting, and learnable frequency band weights to process backbone features. Based on this technical solution, in UAV aerial images where small targets have a small pixel proportion, weak edge texture, and are easily affected by complex backgrounds, the relevant features of small targets can be supplemented and modulated at both the backbone feature extraction and encoding stages, thereby improving the feature representation of distant small targets, weakly textured targets, and complex background regions. The ablation results in the disclosure document show that after introducing only C2f-WDMamba, AP 50 The percentage increased from 32.1% to 35.8%, AP 50-95 The percentage increased from 18.1% to 20.6%, and the number of parameters decreased from 19.88M to 14.30M, indicating that the above-mentioned backbone feature processing method can improve the feature extraction quality in small target detection scenarios with a lower parameter burden.

[0017] This invention also employs CFGN for cross-scale fusion of encoded features. The CFGN includes C2FM and SPGF, and different scale features are fused through a central enhancement module and a scale-aware gating mechanism. The fused features are then input into the decoder and detector head for prediction. Compared to existing RT-DETR methods, which suffer from insufficient integration of cross-scale and contextual information in UAV scenarios and where deep semantics easily obscure details of shallow, small targets, this technique enables a more stable correspondence between high-level semantic information and low-level spatial texture during the fusion process, and improves the consistency of target localization and classification across different scales. The results documented in the disclosure document show that the model combining the three improvements achieves AP on the VisDrone2019 dataset.50 35.9%, AP 50-95 The accuracy rates were 20.7% for small targets, 12.0% for medium targets, and 30.1% for medium targets. Among these, the accuracy for small targets improved to 10.8% when CFGN was introduced alone, indicating that this cross-scale fusion processing chain has a direct effect on the detection results of small targets in complex UAV scenarios. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart of the method provided in an embodiment of the present invention; Figure 2 This is an overall framework diagram of DFS-DETR provided in an embodiment of the present invention; Figure 3 A C2f-WDMamba framework diagram provided for embodiments of the present invention; Figure 4 A schematic diagram of WAS-FFN-AIFI provided for an embodiment of the present invention; Figure 5 A CFGN framework diagram provided for embodiments of the present invention; Figure 6 A visual heatmap provided for embodiments of the present invention; Figure 7 A visualization of the detection results provided in an embodiment of the present invention; Figure 8 A visualization of the receptive field provided for an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] The purpose of this invention is to provide a UAV target detection method based on frequency domain enhancement and state space modeling. By making targeted improvements to the backbone network, AIFI encoder and cross-scale fusion structure, the adverse effects of complex backgrounds, scale variations and long-distance imaging conditions on the representation of small targets are mitigated, thereby improving the detection stability of small and medium-sized targets in the UAV target detection process.

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0023] Figure 1 The method flowchart provided in the embodiments of the present invention is as follows: Figure 1 As shown, this invention provides a UAV target detection method based on frequency domain enhancement and state space modeling, characterized by comprising: Step 100: Acquire the image to be detected and input the image to be detected into the DFS-DETR model built based on RT-DETR; Step 200: Extract features from the image to be detected using the backbone network in the DFS-DETR model to obtain backbone features; wherein, the backbone network adopts C2f-WDMamba, and C2f-WDMamba processes the input features through dynamic dilatation routing, multi-level wavelet high-frequency enhancement and state space modeling; Step 300: Input the backbone features into the AIFI encoder in the DFS-DETR model for encoding to obtain encoded features; wherein, the AIFI encoder introduces WAS-FFN, and WAS-FFN uses windowed frequency domain modeling, channel reweighting and learnable frequency band weights to process the backbone features; Step 400: Input the encoded features into the cross-scale fusion structure in the DFS-DETR model for fusion to obtain fused features; wherein, the cross-scale fusion structure adopts CFGN, which includes C2FM and SPGF. CFGN performs fusion processing on features of different scales through a central enhancement module and a scale-aware gating mechanism. Step 500: Input the fused features and encoded features into the decoder and detection head for prediction, and output the target detection bounding box and target classification result.

[0024] Specifically, this embodiment proposes a DFS-DETR model based on RT-DETR. First, due to the weak global context modeling capability of RT-DETR, its ability to extract small target features is insufficient. To address this issue, this embodiment designs a C2f-WDMamba structure to replace the traditional convolutional backbone. It achieves long-range dependency capture of spatiotemporal features through state space modeling, improving the representation capability of small target features. Second, to solve the problem of weakened small target features caused by the lack of spatial frequency information modeling in AIFI in RT-DETR, this embodiment introduces a WAS-FFN structure into the AIFI module of RT-DETR. By introducing a frequency enhancement mechanism and learnable frequency band weights in the feedforward layer, it achieves spatial-frequency-channel triple feature fusion, improving the efficiency and stability of model feature interaction. Finally, to compensate for the insufficient combination of high-level and low-level semantics caused by the insufficient utilization of frequency information in RT-DETR, this embodiment designs a CFGN structure. It achieves lightweight frequency domain selectivity and cross-scale semantic alignment capabilities by employing dilated convolution and spectral channel gating mechanisms at high levels. The specific technical solution of this embodiment is as follows: Real-time object detection on unmanned aerial vehicles (UAVs) not only needs to cope with the high-density distribution of small targets and complex occlusion, but also needs to overcome computational bottlenecks to meet real-time requirements. Scene classification in images is also a significant challenge. The YOLO series of models is a real-time, single-stage object detection framework that transforms the object detection task into a regression problem. Its main features are direct prediction of bounding boxes and probabilities from images, high speed, strong generalization ability, and strong adaptability. Due to its balance of performance and real-time capabilities, many UAV-OD models are based on the YOLO series. However, this object detection framework generates a large number of redundant boxes during object detection. Therefore, most frameworks use non-maximum suppression (NMS) to handle these redundant boxes. However, using NMS increases computational complexity and also increases the difficulty of model lightweighting and performance optimization.

[0025] DETR employs end-to-end direct detection, abandoning redundant operations like NMS. DETR first uses a CNN to extract features, then processes these features through a Transformer encoder to extract global self-attention. Next, a decoder module, combined with randomly initialized object queries, generates the object detection result. This process eliminates the need for complex post-processing, but its enormous computational cost prevents its use in real-time applications. RT-DETR (Real-Time Detection Transformer) is an object detection model that combines the advantages of the efficient YOLO framework and the DETR architecture, aiming to achieve real-time, high-precision object detection. The RT-DETR image to be detected first passes through a backbone network to initially extract features and reduce the size of the feature map. Then, an efficient encoder performs feature interactions across multiple scales to improve the algorithm's ability to detect objects of different sizes. The efficient encoder consists of Intra-scale Feature Interaction (AIFI) and Cross-Scale Feature Fusion Module (CCFM). Finally, the feature maps from multiple scales and the output of the encoder are integrated as input to the decoder for decoding and the final prediction of the detection head. However, its performance still faces bottlenecks when dealing with complex scenes of drone images, especially small targets, multiple targets, occlusion, and severe background interference.

[0026] Accurate target identification in aerial images with complex backgrounds places high demands on the feature extraction capabilities of the model's backbone network, directly impacting the model's detection accuracy and inference speed. Lightweight vision models can be categorized into two types: CNN-based and ViT (Vision Transformer)-based. CNN-based networks include MobileNet and GhostNet. The MobileNet series (MobileNetV1-V3) reduces the number of parameters and computational complexity by introducing methods such as depthwise separable convolution, inverted residuals with linear bottlenecks, and Neural Architecture Search (NAS). GhostNet, on the other hand, further reduces feature mapping computation by designing Ghost Modules and utilizing a series of linear transformations to obtain feature maps. While these CNN-based methods achieve high efficiency on mobile devices, they are still limited by local receptive fields, resulting in limited global modeling capabilities. ViT-based networks, however, possess a global receptive field, enabling them to establish long-range dependencies within the feature space. MobileViT combines the local modeling of convolution with the global awareness mechanism of VisionTransformer, enabling the model to learn feature representations through simple training. SHViT introduces a single-head attention module to prevent head redundancy and improves accuracy by combining global and local information in parallel. However, these ViT-based model structures are essentially based on attention mechanisms and still suffer from problems such as computational redundancy and insufficient local detail in feature representation.

[0027] In recent years, State Space Modeling (SSM) has provided a new approach for global modeling. The Mamba series of models utilizes a linear recurrence mechanism, where computational complexity is linearly related to the length of the input sequence, enabling efficient feature propagation in both temporal and spatial dimensions. This makes them highly efficient in long sequence modeling and has been widely applied in the field of computer vision. The MobileMamba architecture achieves global perception and dynamic fusion of local details through the coupling of separable state updates and lightweight convolutions.

[0028] Building upon this foundation, this embodiment references the ideas of MobileMamba in improving the RT-DETR backbone network and makes corresponding improvements for UAV detection tasks. Specifically, this embodiment introduces wavelet multi-level decomposition and reconstruction mechanism separation in the band convolution module (MBWTConv2d), and integrates Dynamic Dilated Routing and Gradient Prior-Based Masking (SOAM) within the module to achieve a balance between global modeling and local details.

[0029] In drone aerial photography, significant differences in target scale often occur due to variations in altitude and viewing angle. This can cause shallow texture details to be easily obscured by deep semantics during cross-scale feature fusion, making multi-scale feature fusion a common solution. For example, Feature Pyramid Network (FPN) achieves multi-scale information fusion through bottom-up feature extraction and fusion paths. PANet enhances low-level features and proposes adaptive feature layers to improve the model's feature extraction capabilities. BiFPN proposes a weighted bidirectional feature network and a custom composite scaling method to improve the model's feature representation capabilities. However, these structures that perform convolutions in the spatial domain can lead to an overemphasis on high-level semantic information and a loss of low-level detail information.

[0030] To further improve the multi-scale feature fusion capability of networks, researchers have introduced frequency domain analysis into feature modeling. Frequency domain enhancement methods such as Discrete Wavelet Transform (DWT), Fast Fourier Transform (FFT), and Discrete Cosine Transform (DCT) are used to extract features from different frequency bands and supplement spatial convolutional features. FFCNet combines convolutional networks with frequency learning to compensate for the loss of phase information caused by convolution operations; however, this method is mostly used for single-scale features and is limited in multi-scale tasks. The CSFCN (Context and Spatial Feature Calibration Network) module proposed by Li et al. achieves cross-scale frequency domain modeling by establishing cross-frequency domain dependencies between high-level and low-level features to achieve full-frequency domain fusion.

[0031] Therefore, this embodiment improves the RT-DETR feature fusion stage by referencing the idea of ​​the CSFCN module and making targeted adjustments. This embodiment improves the high-level semantic branch into a center-focused filter modulation with selective suppression (C2FM-SS) structure, and designs a scale-aware parallel fusion structure and a modality gate (SAPF-MG) mechanism in the low-level detail branch to achieve model foreground enhancement, background suppression and cross-scale semantic alignment.

[0032] Feed-Forward Networks (FFNs) are the core components of Transformer encoders and decoders, primarily responsible for nonlinear transformations of features and semantic reconstruction. Traditional FFNs typically use fully connected structures, resulting in weak capabilities in extracting hidden correlations between spatial location and frequency distribution, thus limiting their ability to model detailed textures. To address this, some researchers have incorporated frequency domain modeling into the Transformer structure. FNet uses Fourier transforms to replace the sub-attention layers in the Transformer encoder, achieving globally efficient feature mixing. The SEFF (Spectral Enhanced Feed-Forward) module proposed by Sun et al. enhances the frequency domain pathway based on traditional feed-forward networks and maps spatial features to the frequency domain through a two-dimensional Fast Fourier Transform, achieving global enhancement of frequency components.

[0033] Based on this, this embodiment makes targeted adjustments to the FFN in the AIFI coding layer of RT-DETR. This embodiment designs a new FFN module based on the SEFFN module. Specifically, this embodiment first introduces a local window, uses two-dimensional fast Fourier transform within the local window, and uses learnable dual-frequency gain for modulation. Finally, it combines channel attention to achieve reweighting of different frequencies, so that the model can obtain higher texture accuracy and boundary recognition ability.

[0034] The RT-DETR framework faces challenges in object detection, including loss of low-level details, insufficient cross-scale feature fusion capabilities, and weak long-distance dependency modeling in small target detection scenarios captured by UAVs. To address these issues, this embodiment proposes a novel detection model, DFS-DETR, based on RT-DETR-r18. This embodiment mainly includes three points: 1) In this embodiment, a C2f-WDMamba (Wave and Dynamic Mamba) module is introduced into the backbone network. To address its insufficient local representation capabilities, a high-frequency enhancement and saliency screening structure based on multi-level wavelet transform is designed to enhance the details of small targets in the shallow and middle layers. 2) In this embodiment, a WAS-FFN (Window-based Adaptive Spectral FFN) module is introduced into the AIFI encoder. FFT representation, frequency band separation, and enhancement structures are specifically designed to improve the recognition ability of small targets. 3) In the cross-scale fusion stage, this embodiment introduces a CFGN (Center-Focused and Gated Network) module. CFGN includes two parts: C2FM (Center-Coupled Filtering Module) and SPGF (Scale-Parallel Gating Fusion). A scale-aware gating strategy combining center enhancement and background mechanisms is designed to achieve stable small target detail compensation and semantic alignment. The overall framework diagram of the model is shown below. Figure 2 As shown.

[0035] The original RT-DETR framework uses a ResNet convolutional residual structure for the backbone, which performs well in conventional object detection tasks, but still falls short in small target detection for drones. Because ResNet's convolutional stride is large and continuous downsampling easily leads to the loss of shallow spatial texture, small targets in images are difficult to detect, resulting in insufficient model representation of small targets. Furthermore, small targets in drone scenarios are often distinguished by contextual semantics (such as vehicle groups and roads), while ResNet's local convolutional kernels have an upper limit and lack the ability to model long-range dependencies. In drone aerial images, small targets have sharp edges and fine textures; ResNet only performs convolutions in the spatial domain, underutilizing frequency domain features, which easily leads to the loss of small targets.

[0036] To address the above issues, this embodiment proposes a backbone network suitable for small target detection tasks using unmanned aerial vehicles (UAVs). Specifically, this embodiment designs a C2f-WDMamba module in the backbone network to replace the ResNet Block, such as... Figure 3 As shown.

[0037] In designing the C2f-WDMamba module, this embodiment utilizes YOLOv8's cross-stage feature reuse and efficient channel flow design principles to enable the fusion of local convolutional features and global spatial modeling at multiple scales. The MobileMamba Block module, based on a state-space model (SSM) global modeling structure, possesses excellent long-range dependency capture capabilities, but still suffers from weak local modeling capabilities and a lack of mechanisms for enhancing the high-frequency structure of small targets in UAV aerial photography missions. Therefore, this embodiment systematically improves the MobileMambaBlock module by constructing WDMambaBlock: 1) Designing a dynamic dilation route to expand the effective receptive field of local structures through a learnable dilation selection mechanism; 2) Proposing MBWTConv2d for UAV small target detection tasks, combining multi-level wavelet high-frequency enhancement with SOAM masking to enhance detail and texture; 3) Constructing a steady-state modulation mechanism to stably fuse the high-frequency enhancement branch and the state-space backbone branch. Specific implementation details are as follows: In the proposed architecture, the C2f structure can provide lightweight multi-branch feature fusion capabilities, input features C2f first performs channel splitting:

[0038] x2 then enters the stacked WDMamba Block to enhance long-range dependencies and high-frequency responses:

[0039] Finally, all branches are merged into the output, and the output channel is compressed using convolution:

[0040] The WDMamba Block mainly consists of four stages. First, the input features are processed with DWConv and BN operations to supplement local spatial context information. Then, lightweight FFN is used for channel recalibration and nonlinear enhancement. Afterward, the input enters the MSFM (Multi-branch Spectral-Spatial Fusion Mamba Module). The output of the MSFM is again processed with DWConv, BN, and FFN operations. Finally, residual connections are used to form an information backflow, outputting the final output of the WDMamba Block. The MSFM structure is the core component of the WDMamba Block, splitting the input channels into three parts: global, local, and identity. Through this three-channel division, this embodiment can obtain a mixture of global dependencies, local structures, and the original representation. In the global channel of MSFM, this embodiment designs MBWTConv2d specifically for UAV aerial missions. 1) To address the issue of small target edges being easily obscured, this embodiment designs a dynamic dilation route to enhance the local receptive field. This embodiment achieves multi-scale adaptive convolution kernel selection by introducing dynamic dilation convolution:

[0041]

[0042] in, It is a 3×3 depthwise convolution; For large kernel depthwise convolutions (such as 7×7); This is a dilated convolution with an inflation rate of d=2; To smooth the temperature coefficient.

[0043] 2) This embodiment designs a multi-layered Wavelet high-frequency enhancement and IWT reconstruction structure. First, the low-frequency LL and high-frequency HF are obtained through WT decomposition:

[0044] Then, the highest-level high-frequency HF undergoes dual enhancement and mixing treatment:

[0045]

[0046] Finally, the high-frequency and low-frequency reconstruction weights of the output are improved to enhance stability.

[0047] 3) To make the network sensitive to small target regions, this embodiment also designs a small target mask to enhance detail and texture. This embodiment constructs a small target mask based on gradient magnitude:

[0048] in, and For Sobel gradient kernel, It is a Laplacian-of-Gaussian edge-enhanced kernel.

[0049] Then, a Soft Mask is generated through adaptive Top-K selection:

[0050] Finally, weighted values ​​are generated for high-frequency enhancement:

[0051] in It is the fine-tuning coefficient. This indicates that the gradient backpropagation operation has been stopped. The scaling factor (taken in this article) ).

[0052] After the features are enhanced by high frequency and masked by small targets, this embodiment uses Mamba's state space structure to model long sequences, and finally performs lightweight FiLM modulation on the output:

[0053] in, Adaptive prediction based on input features. After obtaining the modulation coefficients, a global branch output is generated:

[0054] For local channels, this embodiment retains DWConv while improving BN initialization and stability of deeper layers. For identity channels, this embodiment maintains gradient passthrough, improving training stability. Finally, the three channels are fused together for output:

[0055] Where x1 is a local branch, x i This is the identity mapping branch.

[0056] Finally, after passing through the ReLU activation function and PWConv projection, the final output of MSFM is obtained:

[0057] The C2f-WDMamba module designed in this embodiment improves the feature representation of small targets in UAV images through the synergistic effect of dynamic routing design, frequency domain high-frequency enhancement, and state space global modeling, effectively improving the accuracy and robustness of the model in the UAV small target detection task.

[0058] The AIFI architecture in RT-DETR, based on a Transformer architecture with a multi-head attention mechanism (MHSA) and a feedforward network (FNN), has limitations when applied to small target scenarios using UAVs. The feedforward network (FFN) in AIFI is a typical MLP structure, and its form is... It only performs linear mapping in the channel dimension, which cannot clearly distinguish high-frequency details. The original FFN processes each token one by one, with no connection between regions and lacking a spatial receptive field. Therefore, it is not capable of expressing the features of small targets with weak edge contrast and local structures that are difficult to capture in drone scenarios.

[0059] Therefore, this embodiment makes targeted adjustments to the FFN structure in AIFI, referencing the SEFFN module. SEFFN, through frequency domain transformation and modulation, can distinguish between high-frequency and low-frequency information, providing better representation of fine textures in small targets. However, SEFFN is designed for image inpainting and low-frequency enhancement tasks, and its direct application in RT-DETR has limitations. If a two-dimensional FFT is performed globally, its complexity is... However, high-resolution feature maps are insufficient to meet the real-time requirements of UAV deployment. Furthermore, global spectral modulation disperses enhancement across the entire image, preventing high-frequency enhancement from accurately targeting small targets. Therefore, this embodiment designs a WAS-FFN model based on SEFFN, specifically: 1) Windowed frequency domain modeling: Compared to global FFT, windowed modeling is more efficient in high-resolution scenarios and suitable for UAV small target detection; 2) This embodiment constructs high- and low-frequency sub-bands on each window spectrum and introduces two sets of learnable gain parameters for frequency band modulation, improving sensitivity to the detailed textures of small UAV targets; 3) This embodiment introduces a channel attention mechanism after returning to the spatial domain via inverse FFT to enhance the extraction of small target channels. Then, convolutional operations are used for dimensionality reduction and a residual mapping is formed with the input, achieving frequency domain enhancement that is friendly to small targets. The architecture of WAS-FFN-AIFI is as follows: Figure 4 As shown, the specific implementation details are as follows: Input feature map The residual branch is then performed, followed by the introduction of a spatial local receptive field through dimensionality increase and depthwise convolution operations:

[0060] This enables the model to have local structure awareness before entering the frequency domain. Then, windowed frequency domain modeling is performed:

[0061] High-frequency enhancement is restricted to a local window, specifically the vicinity of the target region. This embodiment reduces the computational load by dividing the window. Reduced to This improves the model's real-time performance. After windowing, local FFT and spectral splitting are performed:

[0062]

[0063] After splitting the spectrum into high-frequency and low-frequency bands, in order to separate semantic information and detail information, this embodiment generates high-frequency masks and low-frequency masks, and then performs learnable gain modulation:

[0064] Enhance high-frequency textures and edges; small targets are the most important. Suppress low-frequency background clutter and avoid interference from large objects. After spectral enhancement, inverse FFT and window restoration are performed:

[0065] After regressing to the spatial domain, due to high-frequency and low-frequency masking and modulation, the details of small targets are enhanced while low-frequency backgrounds are suppressed. Then, the output features are combined with a channel attention mechanism to enhance channels sensitive to the target and suppress interfering background channels.

[0066] Finally, channel reduction is performed, and the result is fused with the residual to output the final output:

[0067] The WAS-FFN module proposed in this embodiment enhances the high-frequency texture details of small targets and suppresses low-frequency background noise through windowed frequency domain modeling, high-frequency and low-frequency learnable modulation, and channel attention synergy, effectively improving the model's feature representation ability in UAV small target detection tasks.

[0068] In the original RT-DETR, the Neck uses a combination of FPN and PAN to achieve multi-scale feature fusion through lateral connections and bottom-up path enhancement. However, this type of structure has shortcomings in UAV small target detection tasks: 1) After high-level features are concatenated with low-level features, the scale and semantics of the high-level features weaken the texture of small targets in the low-level features during the fusion process, resulting in a decrease in small target feature extraction; 2) PAN upsampling usually uses nearest neighbor or bilinear interpolation, which is insufficient for distinguishing structural differences between multiple scales, making it difficult to form cross-scale geometric consistency in spatial features; 3) In UAV small target detection tasks, the target proportion is often very small and easily submerged by complex background noise. The RT-DETR Neck structure lacks feature modulation structures such as spatial attention and center enhancement mechanisms, which can easily lead to false detections and false negatives. To address these issues, this embodiment references the idea of ​​CSFCN (Context and Spatial Feature Calibration Network). The CSFCN module is divided into two structures: CFC (Context Feature Calibration) and SFC (Spatial Feature Calibration). The PSP (Pyramid Spatial Pooling) mechanism in the CSFCN module pools features at different scales, so that the key and value contain multi-scale information during construction, which can supplement the missing semantic information of small targets. CSFCN achieves cross-scale semantic alignment through the query-key similarity matrix, which can realize spatial alignment and selectively focus on feature regions to reduce semantic pollution.

[0069] However, the CSFCN module still has shortcomings in small target detection tasks for UAVs. Its Query structure uses global linear projection, which fails to distinguish between foreground and background, resulting in a lack of prominence in the center of small targets. Furthermore, its introduced large-scale pooling (PSP) over-enhances background features, damaging small targets in the image and causing instability in cross-scale fusion. Therefore, this embodiment designs CFGN based on the CSFCN module, comprising two structures: C2FM and SPGF. The CFGN module is as follows: 1) This embodiment designs an explicit center enhancement mechanism to distinguish between foreground and background, effectively solving the problem of small targets being submerged by background noise; 2) This embodiment introduces a background suppression mechanism to suppress large areas of background such as sky and roads commonly found in UAV aerial images; 3) This embodiment designs a cross-scale fusion structure with scale awareness and modal gating mechanisms, which can preserve the detailed texture of shallow features and suppress background noise in deep semantics, achieving more stable cross-scale feature fusion. The structure is as follows: Figure 5 As shown, the specific implementation details are as follows: The CFGN module consists of two structures: C2FM and SPGF, which operate on the large-scale (P5) and small-scale (P3, P4) aspects of RT-DETR, respectively, forming a cross-layer semantic enhancement path. After feature input, feature mapping and dimensionality reduction are performed first.

[0070] in, The convolution is 3×3. Then, center enhancement and background suppression are performed. After dimensionality reduction, the features are modulated using a dual-branch mechanism, including center enhancement and background suppression. The center enhancement mask strengthens the candidate region response.

[0071] in, Perform depthwise separable convolution, BatchNorm, ReLU activation, and 1×1 convolution in sequence. Center enhancement can improve the response to small targets, enhance the extraction of feature details and textures, and reduce the likelihood of small targets being obscured by the background. Background suppression, on the other hand, utilizes global average pooling to suppress interference from non-target regions and improve the contrast between the detected target and the background.

[0072] GAP stands for Global Pooling.

[0073] The final output after modulation is:

[0074] The modulated output enters pyramid spatial pooling, where different features are pooled. First, the key and value are calculated:

[0075] Then calculate the query:

[0076] Calculate cross-scale similarity:

[0077] Context aggregation:

[0078] The final output is obtained after refinement using the LocalAtten local enhancement module:

[0079]

[0080] SPGF integrates features from P3, P4, and P5, and addresses structural misalignment and loss issues during cross-scale fusion through scale consistency and gated fusion. SPGF is divided into low-level spatial structure branches. and high-level semantic branches The two branches are concatenated and then fused using convolution:

[0081] In the final output, this embodiment introduces gated fusion modulation, adaptively selecting the weights of structural and semantic branches to prevent semantic features from causing excessive interference to low-level textures:

[0082] The CFGN module proposed in this embodiment achieves small target detail preservation, noise suppression, and cross-scale fusion dynamic adjustment through center enhancement, background suppression, scale alignment, and modality selection, thereby improving the model's semantic expression ability and localization accuracy in UAV small target detection scenarios.

[0083] Furthermore, the dataset and experimental setup for this embodiment are as follows: VisDrone is a widely used dataset in the field of drone aerial image target detection. This dataset includes 6471 training images, 548 validation images, and 1610 test images, each captured by a drone from different angles and in different scenes. The dataset includes ten different object categories: pedestrians, people, cars, vans, buses, trucks, motorcycles, bicycles, tricycles with canopies, and tricycles. All experiments were conducted on two NVIDIA Tesla V10 (32GB) compute servers running Ubuntu 18.04, using PyTorch 2.3.1 deep learning framework, and CUDA version 12.1. During training, the input image size is 640×640, and the batch size is set to 32. Because the DEIM and DETR models and the SwingTransformer structure are more complex and consume more GPU memory, this embodiment sets the batch size to 16 and the number of training epochs to 350 when training the DEIM and DETR models and the SwingTransformer structure. Validation is performed after each training epoch. The optimizer used is AdamW, and the initial learning rate is set to 1×10⁻⁶. -4 Momentum parameters The weight decays to 1×10 -4Other training configurations remain consistent. For Ultralytics models, Early Stopping is enabled by default (patience=100) to avoid overfitting. The optimizer uses the default auto setting. To ensure fairness, all models in this embodiment are compared using the best weights selected during training.

[0084] This embodiment of DFS-DETR is implemented based on the RT-DETR framework, and the experimental code is built based on the Ultralytics open-source library. To comprehensively evaluate model performance, this embodiment uses standard COCO dataset evaluation metrics, including the average precision (AP) calculated using uniform sampling with an IoU threshold ranging from 0.5 to 0.95 and a step size of 0.05. 50-95 (IoU = 0.50 : 0.95) and AP 50 (IoU = 0.50). Furthermore, this embodiment also provides accuracy metrics at different target scales: AP S AP M This is used to evaluate the detection performance of targets at different scales. To further evaluate the model's detection capabilities, this embodiment uses parameters and GFLOPs to evaluate the model's complexity, and frames per second (FPS) to evaluate the model's inference speed on the NVIDIA Tesla V100 GPU.

[0085] To verify the effectiveness of the model proposed in this embodiment, a systematic comparison was conducted with current mainstream real-time object detectors (YOLO series) and end-to-end detectors (DETR series) on the VisDrone2019 dataset. The results are shown in Table 1. The model proposed in this embodiment shows a significant advantage in detection accuracy, especially in small object detection. Regarding small object detection, the model in this embodiment achieves AP... SThe accuracy reached 12.0%, significantly higher than the best YOLO series models YOLOv12m (9.6%) and YOLOv9m (9.2%), and also showed a clear advantage compared to RT-DETR-r18 (10.4%) and DEIM-D-FINE-M (10.4%) in the DETR series. These results indicate that the C2f-WDMamba proposed in this embodiment effectively solves the problem of small target detail loss during downsampling by introducing multi-level wavelet high-frequency enhancement and dynamic dilation routing. The WAS-FFN module designed in this embodiment utilizes windowed modeling and learnable frequency band modulation to improve the high-frequency components of small target features, enabling the model to extract features of small targets at long distances. Furthermore, the CFGN model designed in this embodiment further improves the response of small targets in multi-scale features through center enhancement and background suppression operations in the cross-scale fusion stage, reducing the obscuring of shallow details. In summary, the model in this embodiment has stronger feature representation and target perception capabilities in UAV small target scenarios. 50 and AP 50-95 In this embodiment, the model achieves accuracy rates of 35.9% and 20.7% respectively, demonstrating a significant advantage over other comparative models. This indicates that the model proposed in this embodiment not only significantly improves the detection accuracy for small targets but also brings a comprehensive improvement to the overall accuracy. This is mainly due to the synergistic effect of the backbone network and cross-scale fusion structure designed in this embodiment. The state-space modeling introduced by the C2f-WDMamba model enhances the model's long-range dependency capture capability. Combined with center enhancement and background suppression of CFGN, the model can effectively distinguish between background and target regions, thereby improving the model's detection accuracy. Meanwhile, the model in this embodiment also achieves high accuracy in AP... M The accuracy reached 30.1%, which is comparable to the best results in the YOLO series, indicating that the model in this embodiment maintains stability in detecting medium-sized targets while improving the accuracy of small target detection.

[0086] In terms of parameter count, the model in this embodiment has only 16.54M parameters, falling between YOLOv8s (11.13M) and YOLOv8m (25.85M), a significant reduction compared to the DETR series. This is mainly due to the WDMamba module designed in this embodiment, which efficiently replaces the ResNet residual blocks, reducing redundant parameters. Regarding computational complexity, despite the introduction of multi-scale enhancement and global modeling mechanisms, the model in this embodiment has a GFLOPs of only 73.8, essentially on the same order of magnitude as YOLOv9m's 76.5, maintaining manageable computational costs. This indicates that the model in this embodiment achieves significantly higher detection accuracy than the comparison models while incurring only a small increase in computational overhead. In summary, the model proposed in this embodiment achieves significantly better detection accuracy and stability than existing models while maintaining manageable computational overhead, offering a significant advantage in application scenarios with extremely high detection accuracy requirements but relatively relaxed real-time requirements.

[0087] Table 1 Comparison of YOLO and DETR series experiments

[0088] To verify the effectiveness of the backbone network proposed in this embodiment, various backbone networks were compared on the VisDrone2019 dataset, including lightweight convolutional networks (ConvNeXtV2, MobileNetV4), a lightweight visual Transformer (EfficientViT), and a hierarchical Transformer (Swin Transformer). In the experiments, except for the replacement of the backbone network, the other network structures remained consistent to ensure that performance differences originated solely from the backbone network. The experimental results are shown in Table 2. The backbone network proposed in this embodiment achieved high performance on AP... 50 and AP 50-95 The metrics reached 35.8% and 20.6%, significantly higher than backbone networks of similar scale such as ConvNeXtV2 (28.6%), EfficientViT (28.2%), and MobileNetV4 (30.0%). Compared to complex backbone networks like Swin Transformer, the AP in this embodiment... 50 and AP 50-95 The fact that the accuracy remains largely unchanged demonstrates that the backbone proposed in this embodiment can achieve a significant improvement in accuracy without substantially increasing computational costs. For small target detection in UAV target detection tasks, the backbone network AP in this embodiment... SAchieving a success rate of 11.9%, the best result among all backbone networks, this advantage primarily stems from the multi-level wavelet decomposition and high-frequency enhancement mechanism designed in this embodiment, which preserves and enhances the detailed information of small targets during the feature extraction stage, as well as the long-range dependency modeling capability provided by state-space modeling. This enables the backbone network to effectively reduce missed and false detections of small targets in complex UAV image backgrounds. Regarding real-time performance, and in AP... 50 Compared to the Swin Transformer, which achieves the best results in terms of metrics, the backbone network in this embodiment maintains comparable accuracy while significantly reducing the number of parameters and computational cost. Furthermore, compared to the slightly lighter MobileNetV4, this embodiment achieves a significant improvement in accuracy. This is because this embodiment achieves more efficient information fusion through linear complexity state-space modeling and lightweight frequency domain enhancement. In summary, the backbone network proposed in this embodiment achieves better detection accuracy while keeping computational overhead under control, realizing a balance between detection accuracy and real-time performance, and is more suitable for target detection tasks in UAV scenarios.

[0089] Table 2 Comparison Experiments with Different Backbone Networks

[0090] To verify the effectiveness of each module proposed in this embodiment, an ablation experiment was conducted on the VisDrone2019 dataset, and the results are shown in Table 3. The experimental results show that after introducing the C2f-WDMamba module, AP... 50 The percentage increased from 32.1% to 35.8%, a rise of 3.7%. 50-95 The performance improved from 18.1% to 20.6%, a 2.5% increase, while the number of parameters decreased from 19.88M to 14.30M. This indicates that the C2f-WDMamba module achieves long-range dependency capture through state-space modeling, and frequency-domain high-frequency enhancement and dynamic routing address the issue of lost details, effectively improving model performance while achieving the lightweight design goal. As a cross-scale fusion structure, the CFGN module, after introducing RT-DETR, improves AP... 50 Increased to 33.7%, AP S The improvement to 10.8% indicates that center enhancement and background suppression prevent deep semantics from interfering with the details of small targets. Combined with scale-aware gating, this effectively preserves shallow spatial textures, enhancing the model's ability to locate small-scale targets. Adding the WAS-FFN module alone provides limited improvement, but its windowed FFT and learnable frequency band modulation can suppress background noise and amplify target textures, providing auxiliary enhancements for other modules. When all modules are combined, the model achieves optimal performance, with AP... 50 AP 50-95 AP S AP MThe efficiency scores were 35.9%, 20.7%, 12.0%, and 30.1% respectively, all outperforming the single-module configuration, demonstrating that the state-space modeling, cross-scale semantic fusion, and frequency-domain feedforward enhancement mechanism of the three modules can effectively complement each other. Although the model fused with all modules outperforms the single-module fused C2f-WDMamba model in terms of AP... 50 While the improvement is limited, the model in this embodiment achieves enhanced multi-scale robustness, specifically in the moderate target accuracy AP after fusion of all modules in this embodiment. M Compared to the baseline model, this represents a 3.8% improvement, and compared to the single-fusion C2f-WDMamba module, a 0.6% improvement. This demonstrates that through the complementary effects of CFGN and WAS-FFN, the model in this embodiment improves spatial detail capture while maintaining a lightweight design, making it more suitable for UAV aerial image target detection tasks. Regarding model complexity, the final model in this embodiment has only 16.54M parameters, a 3.34% reduction compared to the baseline model. In terms of inference speed, due to Mamba's serial scanning mechanism, the FPS decreased from 111.62 to 16.50, but overall AP... 50 The accuracy is improved by nearly 4 percentage points, and the number of model parameters is reduced. Sacrificing some inference speed to meet the requirements of high accuracy and low storage is reasonable in practical applications. Overall, the model in this embodiment can achieve the best balance in terms of accuracy, robustness and structural efficiency.

[0091] Table 3 Ablation Experiment

[0092] To further verify the effectiveness of the model in this embodiment for small target detection in UAV images, this embodiment performs visualization analysis on RT-DETR-r18 and DFS-DETR on the VisDrone2019 dataset. This embodiment also performs heatmap visualization analysis on the baseline and improved models, as shown below. Figure 6 As shown in the figure, compared with the baseline model, the model in this embodiment exhibits higher thermal values ​​for small target areas such as vehicles and pedestrians, indicating that the backbone network and multi-scale fusion module proposed in this embodiment have more effective extraction capabilities. Furthermore, the visualization results show that the model in this embodiment can produce reasonable responses in small target areas while effectively suppressing irrelevant background areas, demonstrating that the WAS-FFN structure proposed in this embodiment effectively suppresses background noise through windowed modeling and learnable frequency band modulation. In summary, the thermal map visualization results further verify the feature representation capability of the model in this embodiment for small targets in UAV aerial images.

[0093] To visually demonstrate the effectiveness of the model in this embodiment, the detection results under different scenarios are visualized and analyzed. The results are as follows: Figure 7As shown in the figure, the model in this embodiment effectively compensates for the missed detection and false detection problems of the baseline model. In road traffic scenarios, the baseline model has difficulty accurately detecting small-sized or occluded targets in weakly textured areas such as tree shade, while the model in this embodiment can stably detect these small-scale targets, indicating that the C2f-WDMamba backbone network proposed in this embodiment, through state space modeling and multi-level wavelet enhancement mechanism, can make the model more robust in complex backgrounds. In parking lot scenarios, the model in this embodiment has more accurate detection results and clearer segmentation effect for small target areas in the distance, indicating that the CFGN structure proposed in this embodiment effectively improves cross-scale feature consistency and expressive power. In street scenarios, the model in this embodiment can accurately identify small targets at a distance and reduce erroneous responses in non-target areas, indicating that the WAS-FFN structure of this embodiment can effectively remove high-frequency noise, making the model have a more accurate recognition effect on target features.

[0094] To verify the improvement in feature focusing capability of the component proposed in this embodiment at different stages, this embodiment performed an effective receptive field visualization analysis on the benchmark model RT-DETR-r18 and the DFS-DETR of this embodiment. The results are shown in Table 4. Figure 8As shown. In the Backbone P3 layer, the response region of the baseline model is larger than that of the model in this embodiment. This indicates that the C2f-WDMamba backbone proposed in this embodiment can enhance the shallow layer's ability to gather detailed textures, making the model more friendly to small targets. In the Backbone P5 layer, the response region of the baseline model shows a diffusion trend, indicating that the model is easily disturbed by the background. However, the response region of the model in this embodiment still shows a centralization trend, maintaining a compact effective receptive field. At the same time, the area ratio of the model in this embodiment at the t=20% threshold decreases from 0.97 to 0.49 compared to the baseline model, indicating that the model in this embodiment also maintains stronger spatial constraint ability in deep semantics. In summary, the backbone network proposed in this embodiment can accurately preserve the features of the target in the feature extraction stage and effectively suppress irrelevant background context information. In the AIFI architecture, the effective receptive field of the baseline model exhibits a wide diffusion. However, the model in this embodiment, after introducing WAS-FFN, has a more concentrated effective receptive field compared to the baseline model. The area ratio at the t=20% threshold decreases from 1.45 to 0.85 compared to the baseline model, indicating that AIFI-WAS-FFN effectively improves the modeling ability of local detailed features and reduces the model's response to irrelevant regions through frequency domain enhancement and weighted modulation. In the feature fusion stage, the model in this embodiment still exhibits a more compact response region compared to the baseline model. The area ratio at the t=20% threshold in layer P5 decreases from 2.49 to 1.85 compared to the baseline model, indicating that the CFGN proposed in this embodiment improves the representation of shallow small targets through background suppression, center enhancement, and cross-scale gating, while also suppressing the diffused background response in deep features. In summary, the receptive field visualization analysis results show that the DFS-DETR model in this embodiment exhibits a more concentrated effective receptive field compared to the baseline model. By suppressing the response of non-target regions and improving the response of target features, the model enhances its ability to locate small targets, making it more suitable for target detection tasks in complex scenarios such as UAV aerial images.

[0095] Table 4 Comparison of Effective Receptive Field Area Ratio of Different Components

[0096] Specifically, the DFS-DETR framework proposed in this embodiment systematically improves RT-DETR in three aspects: global modeling, frequency domain enhancement, and cross-scale fusion. The C2f-WDMamba backbone designed in this embodiment, combined with multi-level wavelet high-frequency enhancement and state-space global modeling, effectively improves the ability to represent small target details. The WAS-FFN module achieves selective target enhancement through windowed frequency domain decomposition and learnable frequency band modulation. Finally, the CFGN module designed in this embodiment enhances cross-scale semantic alignment capabilities and improves the model's robustness in complex backgrounds through center enhancement and background suppression operations combined with scale gating mechanisms. Experiments on the VisDrone2019 dataset show that the model in this embodiment has higher accuracy compared to mainstream models with comparable computational overhead.

[0097] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0098] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A UAV target detection method based on frequency domain enhancement and state space modeling, characterized in that, include: Acquire the image to be detected and input the image to be detected into the DFS-DETR model constructed based on RT-DETR; The backbone network in the DFS-DETR model is used to extract features from the image to be detected, thereby obtaining backbone features; wherein, the backbone network adopts C2f-WDMamba, and C2f-WDMamba processes the input features through dynamic dilatation routing, multi-level wavelet high-frequency enhancement, and state space modeling; The backbone features are input into the AIFI encoder in the DFS-DETR model for encoding to obtain encoded features; wherein, the AIFI encoder introduces WAS-FFN, and the WAS-FFN uses windowed frequency domain modeling, channel reweighting and learnable frequency band weights to process the backbone features; The encoded features are input into the cross-scale fusion structure in the DFS-DETR model for fusion to obtain fused features; wherein, the cross-scale fusion structure adopts CFGN, which includes C2FM and SPGF, and the CFGN performs fusion processing on features of different scales through a central enhancement module and a scale-aware gating mechanism; The fused features and the encoded features are input into the decoder and the detection head for prediction, and the target detection bounding box and target classification result are output.

2. The UAV target detection method based on frequency domain enhancement and state space modeling according to claim 1, characterized in that, The backbone network in the DFS-DETR model is used to extract features from the image to be detected, resulting in backbone features, including: The input features corresponding to the image to be detected are segmented into channels to obtain the first branch features and the second branch features; The second branch features are input into the stacked WDMamba Block for processing to obtain enhanced features; The first branch feature and the enhanced feature are concatenated, and the output channel is compressed through convolution to obtain the backbone feature.

3. The UAV target detection method based on frequency domain enhancement and state space modeling according to claim 2, characterized in that, The second branch features are input into the stacked WDMamba Blocks for processing to obtain enhanced features, including: Perform depthwise convolution and batch normalization operations on the features of the second branch; The features processed by the depthwise convolution and batch normalization operations are then processed by a lightweight feedforward network. The features processed by the lightweight feedforward network are input into the MSFM structure; The output of the MSFM structure is further processed by depthwise convolution, batch normalization, and feedforward network, and the enhanced features are output through residual connections.

4. The UAV target detection method based on frequency domain enhancement and state space modeling according to claim 3, characterized in that, The features processed by the lightweight feedforward network are input into the MSFM structure, including: The input channel is split into a global channel, a local channel, and an identity channel; In the global channel, MBWTConv2d is used to perform dynamic dilatation routing and multi-layer wavelet high-frequency enhancement and IWT reconstruction; The features after the multi-layered wavelet high-frequency enhancement and IWT reconstruction are modeled in the state space structure of Mamba and modulated with FiLM to obtain the global branch output. The output of the global branch, the output of the local channel, and the output of the identity channel are concatenated and then subjected to an activation function and PWConv projection to obtain the output of the MSFM structure.

5. The UAV target detection method based on frequency domain enhancement and state space modeling according to claim 4, characterized in that, In the global channel, MBWTConv2d is used to perform dynamic dilation routing and multi-layer wavelet high-frequency enhancement and IWT reconstruction, including: Multi-scale adaptive convolution kernel selection is achieved through dynamic dilated convolution. Low-frequency LL and high-frequency HF are obtained by wavelet transform WT decomposition; The highest-level high-frequency HF is subjected to dual enhancement and mixing treatment; The processed high-frequency HF and low-frequency LL are reconstructed by inverse wavelet transform (IWT). A small target mask is constructed based on the gradient magnitude, and a soft mask is generated through adaptive Top-K selection; The soft mask is used to generate weighted values ​​for high-frequency enhancement.

6. The UAV target detection method based on frequency domain enhancement and state space modeling according to claim 1, characterized in that, The backbone features are input into the AIFI encoder in the DFS-DETR model for encoding to obtain encoded features, including: Set residual branches for the main features; The core features are sequentially subjected to dimensionality upscaling and depthwise convolution. Windowed frequency domain modeling, local fast Fourier transform, and spectral splitting are performed on the features after the depth convolution processing. Learnable gain modulation is applied to the split high-frequency and low-frequency bands; Channel attention and convolution dimensionality reduction are performed on the features after inverse fast Fourier transform and window restoration, and then fused with the residual branch to obtain the encoded features.

7. The UAV target detection method based on frequency domain enhancement and state space modeling according to claim 6, characterized in that, Learnable gain modulation is performed on the split high-frequency and low-frequency bands, including: Generate a high-frequency mask for the high-frequency band and a low-frequency mask for the low-frequency band; The high-frequency band and the low-frequency band are modulated using two sets of learnable gain parameters, respectively. Perform inverse fast Fourier transform and window restoration on the modulated spectrum.

8. The UAV target detection method based on frequency domain enhancement and state space modeling according to claim 1, characterized in that, The encoded features are input into the cross-scale fusion structure of the DFS-DETR model for fusion to obtain fused features, including: Perform feature mapping to reduce the dimensionality of the encoded features; The reduced-dimensional features are subjected to dual-branch modulation with center enhancement and background suppression to obtain the modulated features. The modulation features are subjected to pyramid pooling, cross-scale similarity calculation and context aggregation, and then refined by a local enhancement module to obtain high-level fusion features; Extract low-level features from the encoded features; The high-level fusion feature and the low-level feature are subjected to gated fusion to obtain the fused feature.

9. The UAV target detection method based on frequency domain enhancement and state space modeling according to claim 8, characterized in that, The reduced features are subjected to dual-branch modulation with center enhancement and background suppression to obtain the modulation features, including: The reduced-dimensional features are processed by the center enhancement module to obtain the center enhancement mask; The reduced-dimensional features are processed by global average pooling to obtain a background suppression mask; The dimensionality-reduced features are modulated based on the center enhancement mask and the background suppression mask to obtain the modulated features; Calculate the key features, value features, and query features for the modulation features; Cross-scale similarity calculation is performed based on the key features and the query features, and context aggregation is performed based on the value features.

10. The UAV target detection method based on frequency domain enhancement and state space modeling according to claim 8, characterized in that, Gated fusion is performed on the high-level fusion features and the low-level features to obtain the fusion features, including: The SPGF is used to construct low-level spatial structure branches and high-level semantic branches; After interpolating the high-level semantic branch, it is then spliced ​​with the low-level spatial structure branch. Convolutional fusion is performed on the concatenated features; The convolutionally fused features are subjected to gated fusion modulation to obtain the fused features.

Images