Frequency Domain Guided Multi-Scale Deformable Alignment UAV Target Detection Method and System

Abstract

This invention discloses a frequency-domain guided multi-scale deformable alignment method and system for UAV target detection. The method first acquires and preprocesses a dataset of UAV aerial images. Next, it improves the YOLOv8 model, performing target detection based on the preprocessed UAV aerial image dataset and outputting the target detection results. Finally, the improved YOLOv8 model is integrated into a pre-configured training environment, using training and validation images from the UAV aerial image dataset for training and validation. This invention significantly enhances the model's ability to perceive small targets, thereby significantly improving the localization and classification accuracy of small targets, reducing detection errors, and improving overall accuracy.

Description

Technical Field

[0001] This invention belongs to the field of computer vision technology and involves technologies such as deep learning, target detection and recognition, feature extraction and fusion. Specifically, it is a frequency domain-guided multi-scale deformable alignment UAV target detection method (FMDANet). Background Technology

[0002] In recent years, with the rapid development of drone technology, drones have been widely used in urban management, disaster emergency response, and agricultural and forestry inspection due to their advantages such as vertical take-off and landing, maneuverability, and low-altitude operation. Meanwhile, the development of deep learning technology has significantly improved the target detection capabilities of aerial images: one-stage models such as YOLO, Faster R-CNN (accelerated region convolutional neural network), and RetinaNet (a one-stage target detection network based on focus loss) achieve an effective balance between detection speed and accuracy; while end-to-end detection models based on Transformer (a deep neural network with self-attention structure), such as DETR (Detection Transformer, an end-to-end target detection framework based on Transformer), Deformable DETR (a deformable detection Transformer model), and RT-DETR (a real-time end-to-end detection model), reduce the dependence on anchor boxes and NMS (non-maximum suppression), further optimizing inference latency. However, when these advanced algorithms are applied to drone aerial photography scenarios, they still face a series of technical challenges that urgently need to be addressed. On the one hand, top-down or oblique-view imaging results in a very small proportion of target pixels, and features are easily submerged after network downsampling. Combined with factors such as occlusion, low illumination, and noise blurring, target visibility is reduced, leading to a high risk of false positives and false negatives. On the other hand, mainstream detection models typically have high computational costs, making it difficult to achieve low-latency real-time inference under the constraints of computing power and power consumption of UAV onboard equipment. Therefore, there is an urgent need to develop efficient detection methods suitable for UAV platforms to balance accuracy and speed under limited resources and effectively improve the detection performance of small targets in complex environments. Summary of the Invention

[0003] To address the problems existing in the prior art, this invention proposes a frequency domain-guided multi-scale deformable alignment method for UAV target detection. The main contents of this invention are: (1) In order to overcome the challenge caused by the small target size, this invention eliminates the P5 scale sampling layer in the backbone network of the model, adopts a P2 / P3 / P4 three-scale feature fusion structure in the neck network of the YOLOv8 model, and introduces a higher resolution P2 detection scale in the detection head, thereby supporting the discrimination with richer fine-grained texture and edge information, significantly improving the accuracy and stability of small target detection, and achieving more accurate target recognition. (2) In order to improve the small target representation ability in the downsampled features and reduce the computation and parameter quantity of the backbone, this invention introduces a lightweight hybrid domain feature extraction module (HDM) on the backbone side. By jointly modeling the spatial-frequency domain information and channel response, it adaptively strengthens the key textures and edge details related to the target, significantly reducing the computation and parameter quantity without significantly reducing the accuracy. (3) To enhance multi-scale perception capabilities, this invention introduces a progressive multi-kernel perception cascade module (MPC) in the Backbone. This module enhances scale-sensitive representation by aggregating multi-kernel features at each level, thereby improving the robustness of target detection from the UAV's perspective. (4) To maintain stable detection performance in complex scenarios such as insufficient lighting and partial target occlusion, this invention introduces a cross-scale alignment and gated weighted feature fusion module (HDAF) on the Neck side. This module filters and strengthens key features through operations such as frequency domain processing, cross-scale alignment, gated weighting, and redundancy suppression, significantly improving the model's detection accuracy for small and occluded targets, especially in complex environments.

[0004] In one aspect, the specific steps of the frequency-domain guided multi-scale deformable alignment UAV target detection method are as follows:

[0005] S1 acquires a dataset of drone aerial images and preprocesses it:

[0006] S2 modifies the original YOLOv8 model structure:

[0007] This invention improves the YOLOv8 model and proposes a frequency-domain guided multi-scale deformable alignment UAV target detection method. Target detection is performed based on a preprocessed UAV aerial image dataset, and the target detection results are output. The specific processing method of this algorithm is as follows:

[0008] (1) The sampling layer at the P5 scale was removed from the backbone network of the YOLOv8 model. In the Neck of the model, a three-scale feature fusion structure of P2 / P3 / P4 was adopted to retain more fine-grained textures. Compared with using only P3 / P4 / P5, it can effectively reduce the risk of missing small target information.

[0009] (2) In the Backbone of the YOLOv8 model, a Hybrid Domain Feature Extraction (HDM) module is introduced at the P2, P3, and P4 scales. The HDM divides the channel into frequency domain spatial processing branch and channel processing branch according to the proportion, and fuses the outputs of the two branches to obtain enhanced features. This significantly reduces the amount of computation and parameters without significantly reducing the accuracy, and is used for subsequent detection.

[0010] The specific implementation process of the HDM module is as follows:

[0011] Let the input feature map be X. First, input... Dividing by channel dimension yields... and Two parts; secondly, and Features are obtained by passing the data through two convolutional networks. and Then After depthwise convolution, global average pooling is performed along the channel dimension, followed by a sigmoid operation to obtain the output. ;at last After compressing the number of channels to 1 using a 1×1 convolution, the output features are calculated again using batch normalization, a sigmoid operation, and a frequency-domain-based spatial-frequency hybrid transform operator. ; feature map and Perform element-wise dot product to transform the feature map and Perform element-wise dot products, and finally add and fuse the results to obtain the output feature. .

[0012] (3) Introducing the progressive multi-core perception cascade module MPC into the backbone of the YOLOv8 model, the MPC module enhances the feature representation by progressively aggregating multi-scale receptive fields and combining them with lightweight residual structures. Under the premise of controlling computational overhead, it significantly enhances the robustness of small target detection from the perspective of UAV.

[0013] (4) The cross-scale alignment and gated weighted feature fusion module HDAF is introduced in the Neck of the YOLOv8 model. The HDAF module uses the frequency domain to process the bias between learning scales, uses the bias to guide deformable convolution to embed the shallow semantic information in the network model into the deep location information map, and finally performs gated weighting to complete adaptive fusion, and finally outputs a feature map with high semantic feature information with accurate localization.

[0014] S3 training network model:

[0015] First, the improved YOLOv8 model is integrated into the pre-configured training environment, and the parameter file in the network model is modified as required. Then, the model is trained and validated using pre-defined training and validation sets of images from the UAV aerial image dataset.

[0016] S4 Analysis of Experimental Results:

[0017] The improved YOLOv8 model automatically generates a corresponding weight file after training. This pre-trained weight file is then imported into the testing program to validate the input test set images. After execution, the program outputs the object detection results from the input images. Finally, these outputs are carefully evaluated to verify whether their recognition performance meets the experimental expectations.

[0018] In another aspect, the present invention also provides a frequency-domain guided multi-scale deformable alignment UAV target detection system, comprising the following modules:

[0019] The data acquisition module is used to acquire and preprocess drone aerial image datasets.

[0020] The target detection module is used to improve the YOLOv8 model, perform target detection based on the preprocessed drone aerial image dataset, and output the target detection results.

[0021] The training and validation module is used to integrate the improved YOLOv8 model into a pre-configured training environment, using training and validation images from the drone aerial image dataset to train and validate the model.

[0022] The beneficial effects of this invention are:

[0023] This invention discloses a UAV aerial target detection method based on a frequency domain-guided multi-scale deformable alignment detection network: (1) In order to enhance the model's ability to detect small targets, this invention eliminates the P5-scale sampling layer in the Backbone and adopts a P2 / P3 / P4 three-scale feature fusion structure in the Neck, so that high-resolution P2 features directly participate in the three-scale detection, which significantly enhances the model's ability to perceive small targets. While preserving deep semantic information, it makes full use of shallow fine-grained textures, reducing the probability of losing small target information, thereby significantly improving the localization accuracy and classification accuracy of small targets. At the same time, eliminating the detection of ultra-low resolution P5 feature layers avoids redundant features that contribute limited to small targets from participating in prediction, thus balancing parameter scale and computational efficiency while ensuring detection performance. (2) In order to reduce the computation and parameter quantity of the backbone network while capturing local features in the image, this invention inserts a lightweight hybrid domain feature extraction module (HDM) into the Backbone. This module constructs two heterogeneous convolutional branches by proportionally dividing the input channels, and jointly introduces frequency domain spatial attention and channel attention mechanisms to achieve feature enhancement with extremely low additional parameters. This module can enhance the expression of key textures and edge details related to the target by adaptively weighting and fusing the branch outputs, so that the network can focus more on important regions related to the target, providing more stable and discriminative basic features for subsequent detection, while taking into account both lightweight and feature representation capabilities. (3) In order to better balance scale robustness and computational overhead, this invention introduces a progressive multi-core perception cascade module MPC in the Backbone. This module uses the progressive aggregation features of different receptive fields of 3×3, 5×5, and 7×7, and strengthens the integration of contextual information through a lightweight residual structure, thereby enhancing the adaptability to multi-scale targets and the robustness of small target detection from the perspective of UAVs. (4) In order to cope with the detection challenges in complex environments such as low light, target occlusion, and dense scenes, this invention introduces a scale feature fusion module HDAF in the Neck. This module efficiently fuses shallow localization information from the backbone or neck with deep semantic information from the neck. By extracting the discrepancies between features at different scales, it guides deformable convolutions to spatially align the features, and finally achieves adaptive fusion through gated weighting. The HDAF module effectively enhances the model's visibility and discrimination ability for key details in low-light, occluded, and dense scenes, thereby significantly reducing detection errors and improving overall accuracy. Attached Figure Description

[0024] Figure 1 This is a diagram of the overall model structure.

[0025] Figure 2 This is a diagram of the HDM feature fusion structure.

[0026] Figure 3MPC structure diagram;

[0027] Figure 4 HDAF structure diagram;

[0028] Figure 5 For comparison of detection results Figure 1 ;

[0029] Figure 6 For comparison of detection results Figure 2 ;

[0030] Figure 7 For comparison of detection results Figure 3 . Detailed Implementation

[0031] The present invention will be further described below with reference to the accompanying drawings and specific examples. However, it should be noted that the present invention is not limited to the following embodiments. Those skilled in the art should recognize that various substitutions and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the scope of protection of the present invention should not be limited to the content disclosed in the embodiments.

[0032] In one aspect, this invention provides a frequency-domain guided multi-scale deformable alignment method for UAV target detection, comprising the following steps:

[0033] 1. Data Acquisition:

[0034] This invention uses the VisDrone2019 dataset, publicly released by the Machine Learning and Data Mining Laboratory of Tianjin University, as the source of experimental data. This dataset contains 10,209 images, covering various scene types in 14 cities across my country, and involving different weather conditions and lighting conditions. The dataset is divided according to a predetermined ratio, with 6,471 images used as the training set, 548 as the validation set, and 1,610 as the test set, providing a complete and standardized data foundation for the algorithm training and performance evaluation of this invention.

[0035] The VisDrone2019 dataset contains annotations for ten target categories, including pedestrians, bicycles, cars, vans, trucks, tricycles, tuk-tuks with sunshades, buses, and motorcycles, totaling approximately 2.6 million annotations. This dataset exhibits high complexity and diversity in terms of target categories, scene backgrounds, and imaging conditions, effectively testing the robustness and generalization ability of target detection algorithms in complex real-world environments. It serves as an ideal benchmark dataset for validating and optimizing the target detection method of this invention.

[0036] After preprocessing the dataset, a corresponding dataset.yaml configuration file was first constructed. The paths to the training and validation sets, as well as the number of classes in the VisDrone2019 dataset, were written into this file to ensure the training framework could load the data correctly. Subsequently, the default.yaml file in the YOLOv8 source code was configured according to experimental requirements. The input image size was set to 640×640, the number of training epochs was set to 300, the batch size was set to 8, and the number of target classes was configured to 10, thus meeting the training requirements of the target detection task of this invention.

[0037] The environment is set up according to the environment required by the model and the basic hardware conditions. The environment of this invention is: CUDA 11.8, deep learning framework PyTorch 1.13.1, Intel Core i9-13900ks CPU, 128G memory, and NVIDIA GeForce RTX 4090 GPU with 24G video memory.

[0038] 2. Improved the original YOLOv8 model structure, combined with the attached... Figure 1 The main improvements are as follows:

[0039] (1) Compared with the traditional YOLOv8 network structure, this invention reduces the maximum downsampling factor of its backbone from 1 / 32 to 1 / 16, retaining only the P4 feature layer and no longer generating the P5 layer. In the Neck part, feature fusion is no longer performed on the three scales of P3, P4, and P5. Instead, the entire scale is shifted down by one scale, and multi-scale feature fusion is carried out around P2, P3, and P4 in a top-down and bottom-up manner. Target detection is then performed on the features after multi-scale feature fusion. In dense small target scenes, the P5 layer features have low information density due to their low resolution, resulting in significant loss of details and limited contribution to the detection task. They may even introduce redundant computation. By removing the P5 layer and enhancing the participation of high-resolution features such as P2 and P3, this structure reduces the dependence on low-information-density large-scale features, reduces feature redundancy and computational overhead, and improves the utilization rate of shallow fine-grained features, thus being more conducive to the accurate detection of small and medium-scale targets.

[0040] (2) This invention introduces the lightweight hybrid domain feature extraction module (HDM) into Backbone, as shown in the attached figure. Figure 2 As shown, HDM modules are inserted after Conv modules at different scales (P2 and P3), and two HDM modules are inserted after the Conv module at scale (P4) to enhance the details related to small targets during the downsampling stage. Taking a general HDM module as an example, its internal processing flow is shown in equations (1) to (6). Let the input feature map be... The number of channels is C, and the spatial dimensions are H×W:

[0041]

[0042]

[0043]

[0044]

[0045]

[0046] In formula (1) Indicates that the input will be entered. The channels are divided in a 1:3 ratio. and These respectively represent the input features The two parts are obtained by dividing the channel dimension in a 1:3 ratio.

[0047] In formula (2) This represents a 5×5 depthwise convolution kernel. This represents a 3×3 ordinary convolution kernel. This represents a 1×1 ordinary convolution kernel.

[0048] In formula (3) This represents a 1×1 ordinary convolution kernel.

[0049] In formula (4) This represents a 3×3 depthwise convolution kernel. This represents global average pooling along the channel dimension. This represents the sigmoid operation.

[0050] In formula (5) This indicates a mapping operation that uses a 1×1 convolutional kernel to compress a feature map with a channel dimension of dim into a single channel. This indicates a batch normalization layer, which is used to standardize features along a mini-batch dimension, improving the stability of network training. This represents the sigmoid operation. This describes a frequency-domain hybrid spatial-frequency transformation operator. Its implementation is as follows: First, the input feature map with channel dimension *dim* is compressed into a single-channel feature map using a 1×1 convolution. Next, batch normalization (BN) and sigmoid activation are applied to this feature map, resulting in a standardized and nonlinearly transformed feature map. Then, a two-dimensional complex Fast Fourier Transform (FFT) is performed to obtain the real and imaginary features. Next, the real and imaginary features are concatenated along the channel dimension, and a linear transformation is performed in the frequency domain using a 1×1 grouped convolution. Finally, the transformed real and imaginary features are recombine into complex features, mapped back to the spatial domain using a two-dimensional inverse Fast Fourier Transform (IFFT), and the real part is taken as the output feature. This operator can perform learnable filtering and reconstruction of the spatial saliency map in the frequency domain, thereby achieving frequency domain enhancement of spatial information.

[0051] Finally, as attached Figure 2 As shown, the feature map and Perform element-wise dot product to transform the feature map and Perform element-wise dot products, and finally add the two results together. The fusion process is shown in equation (6):

[0052]

[0053] In formula (6) This indicates an element-wise multiplication (dot product) operation.

[0054] The HDM module enables a lightweight interaction between semantic and spatial-frequency domain features, outputting feature maps with both strong semantics and high spatial-frequency resolution. While significantly reducing computational load, this module not only enhances the network's ability to discriminate subtle features but also comprehensively improves the model's stability and reliability under complex conditions.

[0055] (3) This invention introduces the Progressive Multi-Core Perception Cascade Module (MPC) into Backbone, as shown in the appendix. Figure 3 As shown, an MPC module is inserted between the two HDM modules at the P4 scale. By aggregating multi-scale features step by step, the scale awareness capability is enhanced, improving the robustness of small target detection in environments with large scale variations. Taking a general MPC module as an example, its internal processing flow is as follows. Let the input feature map be... The number of channels is C, and the spatial dimensions are H×W:

[0056]

[0057] In formula (7) , , This represents depthwise convolution kernels of sizes 3×3, 5×5, and 7×7. This represents a 1×1 ordinary convolution kernel.

[0058] The MPC module extracts and fuses features from targets at different scales through the above operations, improving the ability to perceive multi-scale targets. Ultimately, the output feature map combines multi-layered information, effectively enhancing the robustness of small target detection while keeping computational costs under control.

[0059] (4) This invention introduces the HDAF module, a cross-scale alignment and gated weighted feature fusion module, into Neck. (See attached...) Figure 1 , 4 As shown, the Neck contains a first HDAF module, a second HDAF module, a third HDAF module, and a fourth HDAF module. Each HDAF module receives two inputs: a large-scale detail feature map (FL) from a shallower layer of the network, and a semantic feature map (FS) from a deeper layer of the network, which has been upsampled or downsampled to have the same resolution as FL. The detail feature map (FL) contains more fine-grained spatial information representing the edge contours, texture variations, and local structures of the target; the semantic feature map (FS) provides stronger semantic expressive power. To meet the channel consistency requirements of the fusion computation within the HDAF module, the FL and / or FS can be aligned in terms of channel number using a convolutional module (e.g., a 1×1 convolution) before being input into the corresponding HDAF module.

[0060] Furthermore, detailed feature maps of each HDAF module. Specifically, this includes: the first detailed feature map. For the backbone network Backbone The output features in the scale branch are obtained after channel alignment by the convolution module; the second detail feature map. For the backbone network in The output features in the scale branch are obtained after channel alignment by the convolution module; the third detail feature map The output features of the first C2f module in Neck; the fourth detail feature map. Backbone network The output features of the SPPF module in the scale branch are obtained after channel alignment by the convolution module. The fourth detail feature map FL4 is simultaneously fed into the first layer of the Neck network and the fourth HDAF module.

[0061] Semantic feature maps of each HDAF module Specifically, this includes: the first semantic feature map The first upsampling layer in Neck provides the output feature map; the second semantic feature map is also shown. The output feature map of the second upsampling layer in Neck; the third semantic feature map. The first downsampling layer in Neck is the output feature map; the fourth semantic feature map is the output feature map. This is the output feature map of the second downsampling layer in Neck.

[0062] Based on this, the HDAF module first performs... Frequency domain enhancement is performed, and its key features are extracted and represented as follows: Subsequently, the two sides joined forces. and Learning scale bias, used for guidance Spatial alignment is then achieved. Finally, adaptive fusion is realized through a gated weighting mechanism, outputting an enhanced feature map that simultaneously contains rich location and semantic information. Let the input feature map be... and The number of channels are respectively and Given a spatial dimension of H×W, the processing flow of the HDAF module is as follows:

[0063]

[0064]

[0065]

[0066]

[0067]

[0068]

[0069]

[0070] In formula (8) This represents a space-frequency domain hybrid transform operator based on the frequency domain. After passing through this operator, we get .

[0071] In formula (9) and This represents feature maps from shallow and deep layers that have been upsampled to the same resolution. This indicates a frequency domain enhancement operation, which uses Fast Fourier Transform (FFT), 1×1 convolutional layers, and Inverse Fast Fourier Transform (IFFT) to process features in the frequency domain to enhance key information and suppress redundant components. This indicates a channel splicing operation. This indicates that a 1×1 convolution kernel is used to reduce the number of channels from... Mapped to The process.

[0072] In formula (10) This indicates the bias and mask prediction layers implemented using 3×3 convolutional layers, used to process the fused feature maps. Perform mapping to obtain the intermediate output tensor. The aforementioned The number of output channels is 3 * k * k * G (where k = 3, G = 1), and its spatial size is... Maintain consistency.

[0073] In formula (11) This indicates a channel partitioning operation, used to divide... Dividing it into three parts along the channel dimension, we obtain... , , Each of these copies has k * k * G channels.

[0074] Formula (12) will show , The spatial offset is obtained by stitching along the channel dimension. Its number of channels is 2 * k * k * G; The modulation mask is obtained by sigmoid operation. To constrain the mask values ​​to The interval has k * k * G channels. Finally, the spatial offset parameters required for deformable alignment are obtained. With modulation weights .

[0075] In formula (13) Represents deformable convolution, which uses Input feature map and explicitly introduce and As an alignment parameter; where, Used for adaptive displacement adjustment of convolution sampling positions. Used to weight and modulate the response contribution of the corresponding sampling point, so as to achieve the Spatial alignment and detail calibration are performed to obtain the calibrated feature map. ).

[0076] In formula (14) This indicates a channel splicing operation. The gating mechanism is implemented using a 1×1 convolutional layer, which controls the spliced ​​calibration features ( ) and shallow features ( Weighted learning. This represents the sigmoid operation, which constrains the gating coefficient W to the interval [0, 1] to adaptively characterize the importance of different positions and channels.

[0077] Finally, as attached Figure 4 As shown, the fusion process uses weighting coefficients. Adaptive fusion of shallow and deep features yields the output feature. The fusion process is shown in equation ().

[0078]

[0079] In formula (15) This represents the gating weighting coefficient. This represents the deep calibration features after deformable convolution alignment. This represents a shallow, large-scale feature map. This represents a 1×1 ordinary convolution kernel.

[0080] The HDAF module effectively mitigates the cross-scale mismatch problem in spatial location and semantic representation between shallow large-scale feature maps and deep upsampled feature maps at the same resolution through a step-by-step processing flow of "frequency domain enhancement—bias prediction—deformable alignment—gated weighted fusion," significantly improving the accuracy and robustness of multi-scale feature fusion. In low-light, partially occluded, and densely packed small target scenes, this module can highlight key detail responses, suppress background and noise interference, and provide more stable and discriminative feature representations for subsequent heads, thereby improving detection performance in small targets and complex scenes.

[0081] In another aspect, the present invention also provides a frequency-domain guided multi-scale deformable alignment UAV target detection system, comprising the following modules:

[0082] The data acquisition module is used to acquire and preprocess drone aerial image datasets.

[0083] The target detection module is used to improve the YOLOv8 model, perform target detection based on the preprocessed drone aerial image dataset, and output the target detection results.

[0084] The training and validation module is used to integrate the improved YOLOv8 model into a pre-configured training environment, using training and validation images from the drone aerial image dataset to train and validate the model.

[0085] Example:

[0086] This invention first preprocesses the VisDrone2019 dataset, strictly dividing it into 6471 training samples and 548 validation samples. Based on this, the model training rounds are set to 300, with a batch size of 8 images per iteration. Subsequently, the improved network structure is used to train the dataset. During training, the Wandb platform is used to monitor training logs such as loss curves and accuracy metrics in real time, and the model weights and related experimental results are fully saved after training. Based on this, this invention systematically analyzes the actual contribution of each module and the overall performance improvement by combining ablation experiments and visualizations of object detection in typical scenarios. To quantitatively evaluate the impact of different modules on model performance improvement, this invention designs and conducts a series of ablation experiments, and the comparison results are shown in Table 1.

[0087] Table 1. Detection results after adopting different improvement strategies

[0088]

[0089] Note: √ indicates that the module was used.

[0090] mAP in the table 50 The average accuracy of the model is expressed as (%) when IoU = 0.5; mAP50-95 represents the average accuracy of the model when IoU = 0.5 to 0.95; Param represents the number of model parameters; FLOPs represents the number of floating-point operations required for one forward inference; GFlops represents the number of floating-point operations required for one forward inference.

[0091] The ablation experiment results show that Model B, which employs a down-shifted three-scale feature fusion structure, has a higher mAP compared to the baseline model A (YOLOv8s). 50 The percentage increased from 39.4% to 44.2%, a rise of 4.8 percentage points. (mAP) 50-95 The accuracy improved from 23.4% to 26.8%, an increase of 3.4 percentage points. Simultaneously, the number of parameters decreased significantly from 11.1M to 3.7M, resulting in a substantial reduction in model storage overhead. FLOPs only increased from 28.5G to 32G, with a computational increase of approximately 3.5G, making the overall cost increase relatively limited. Model C was obtained by introducing an HDM hybrid domain feature extraction module based on Model B. Compared to Model B, Model C showed improved mAP. 50 Only a 0.5% decrease, mAP 50-95The decrease was only 0.5 percentage points. Meanwhile, the number of parameters continued to be compressed from 3.7M to 3.3M, and FLOPs decreased from 32G to 26.2G, reducing computation by approximately 5.8G. This achieved further reduction in model complexity with almost no sacrifice in accuracy. Subsequently, an MPC progressive multi-core perceptron cascade module was added to model C to form model D, mAP 50 The percentage increased from 43.7% to 44.1%, mAP 50-95 The accuracy slightly increased to 26.4%. Simultaneously, the number of parameters decreased from 3.3M to 2.4M, and FLOPs decreased from 26.2G to 25.2G, indicating that MPC helps improve feature representation capabilities with lower complexity. Finally, based on model D, an HDAF cross-scale feature fusion module was introduced to construct the improved model FMDA-YOLO of this invention. 50 The percentage increased from 44.1% to 48.8%, a rise of 4.7 percentage points, mAP 50-95 The percentage increased from 26.4% to 29.9%, a rise of 3.5 percentage points. Compared to the baseline model A, mAP... 50 and mAP 50-95 The accuracy and performance were improved by 9.4 and 6.5 percentage points respectively, and the number of parameters decreased from 11.1M to 3.68M, a reduction of approximately 7.4M overall, which is only about one-third of the original model. Although the number of FLOPs increased to 52.7G, the detection accuracy was significantly improved while maintaining a smaller model size, demonstrating a superior accuracy-complexity trade-off.

[0092] To visually demonstrate the effects of this invention, in conjunction with the appendix... Figure 5 , Figure 6 , Figure 7 The explanation compares the detection results of the improved model and the original model in the same scene. In each set of images, the top image shows the detection results of the original model, and the bottom image shows the detection results of the improved model. The small yellow dashed box marks the region of interest, the large yellow dashed box is a magnified view of the region, and the yellow connecting lines are used to indicate the magnification correspondence. The category text next to the detection box and the decimal number after it indicate the prediction confidence of the target.

[0093] As attached Figure 5 As shown, a comparison of the detection performance of the original YOLOv8(s) model and the improved model of this invention in the scenario of distant small target vehicles is presented. The small yellow dashed box marks the area of ​​distant small target buses, and the right side is a magnified view. As can be seen from the magnified area, the original model only detected one bus (confidence level of 0.61), and missed detections of buses further away; the improved model of this invention can additionally detect distant bus targets at the same location (confidence level of 0.33), and maintain stable detection of nearby buses (confidence level of 0.62), thereby reducing the missed detection rate of distant small target vehicles and improving the detection capability of small targets.

[0094] As attached Figure 6 As shown, a comparison of the detection performance of the original YOLOv8(s) model and the improved model of this invention in low-light and dark environments is presented. The small yellow dashed boxes mark the dark areas of the road, as can be seen in the magnified view on the right. The original model failed to effectively detect vehicle targets in the dark areas; the improved model of this invention can detect a "car" target in the same dark area (confidence level of 0.51) and provides correct localization and identification results for multiple small targets in the dark areas, thus significantly improving the target detection capability in low-light and dark environments.

[0095] As attached Figure 7 As shown, a comparison of the detection performance of the original YOLOv8(s) model and the improved model of this invention in occluded scenarios is presented. The small yellow dashed box marks the pedestrian area partially obscured by trees on the roadside, with a magnified view on the right. As can be seen from the magnified area, the original model misses pedestrians in this occluded area; the improved model of this invention can detect occluded pedestrian targets at the same location (e.g., providing the detection result and corresponding confidence score for "pedestrian / person"), achieving more accurate localization and recognition, and improving the robustness of detection in complex occluded scenarios. (Comprehensive comparison attached) Figures 5-7 It can be seen that, under challenges such as small targets, low illumination, and target occlusion, the detection performance of the improved model proposed in this invention is superior to that of the original model.

Claims

1. A frequency-domain guided multi-scale deformable alignment method for UAV target detection, characterized in that, Includes the following steps: S1. Obtain the drone aerial image dataset and preprocess it; S2. Improve the YOLOv8 model and perform target detection based on the preprocessed UAV aerial image dataset, and output the target detection results. The improvement of the YOLOv8 model includes the following operations: In the backbone network of the YOLOv8 model, the sampling layer at the P5 scale is removed, and in the Neck of the YOLOv8 model, a feature fusion structure of P2, P3 and P4 at three scales is adopted. In the backbone of the YOLOv8 model, a lightweight HDM feature extraction module is introduced at the P2, P3, and P4 scales. HDM divides the channel into frequency domain spatial processing branch and channel processing branch according to the proportion, and fuses the outputs of the two branches to obtain enhanced features. In the backbone of the YOLOv8 model, the MPC feature perception module is introduced. The MPC feature perception module enhances feature representation by progressively aggregating multi-scale receptive fields and combining them with lightweight residual structures. In the Neck of the YOLOv8 model, the HDAF cross-scale feature fusion module is introduced. The HDAF cross-scale feature fusion module uses the frequency domain to process the bias between learning scales, uses the bias to guide deformable convolution to embed shallow semantic information in the network model into deep location information map, and finally performs gated weighting to complete adaptive fusion, and finally outputs a feature map with high semantic feature information with accurate localization. S3. Integrate the improved YOLOv8 model into the pre-configured training environment, and use the training and validation sets of images from the drone aerial image dataset to train and validate the model.

2. The frequency-domain guided multi-scale deformable alignment UAV target detection method according to claim 1, characterized in that, The P5-scale sampling layer is removed from the backbone network of the YOLOv8 model. Instead, a three-scale feature fusion structure (P2, P3, P4) is adopted in the Neck of the YOLOv8 model. The specific implementation is as follows: In the traditional YOLOv8 network structure, the Backbone part is retained only up to the P4 feature layer, and the P5 layer is no longer generated. In the Neck part, multi-scale feature fusion is carried out around P2, P3, and P4 in a top-down and bottom-up manner, and target detection is completed on the features after multi-scale feature fusion.

3. The frequency-domain guided multi-scale deformable alignment UAV target detection method according to claim 2, characterized in that, The HDM lightweight feature extraction module is introduced into the backbone of the YOLOv8 model at the P2, P3, and P4 scales. HDM divides the channels proportionally into frequency domain spatial processing branches and channel processing branches, and fuses the outputs of the two branches to obtain enhanced features. The specific implementation is as follows: The lightweight hybrid domain feature extraction module HDM is introduced into Backbone. HDM modules are inserted after the Conv modules at different scales of P2 and P3, and two HDM modules are inserted after the Conv module at the P4 scale. The specific implementation process of the HDM module is as follows: Let the input feature map be X. First, input... Dividing by channel dimension yields... and Two parts; secondly, and Features are obtained by passing the data through two convolutional networks. and Then After depthwise convolution, global average pooling is performed along the channel dimension, followed by a sigmoid operation to obtain the output. ;at last After compressing the number of channels to 1 using a 1×1 convolution, the output features are calculated again using batch normalization, a sigmoid operation, and a frequency-domain-based spatial-frequency hybrid transform operator. ; feature map and Perform element-wise dot product to transform the feature map and Perform element-wise dot products, and finally add and fuse the results to obtain the output feature. .

4. The frequency-domain guided multi-scale deformable alignment UAV target detection method according to claim 3, characterized in that, The specific calculation process of the frequency-domain-based spatial-frequency hybrid transform operator is as follows: First, a two-dimensional complex fast Fourier transform is performed to obtain the real and imaginary features. Next, the real and imaginary features are concatenated along the channel dimension and a linear transformation is performed in the frequency domain using 1×1 grouped convolution. Finally, the linearly transformed real and imaginary features are recombine into complex features, which are then mapped back to the spatial domain using a two-dimensional inverse fast Fourier transform, and their real parts are taken as the output features.

5. The frequency-domain guided multi-scale deformable alignment UAV target detection method according to claim 4, characterized in that, The introduction of the MPC feature perception module into the backbone of the YOLOv8 model enhances feature representation by progressively aggregating multi-scale receptive fields and combining them with lightweight residual structures. The specific implementation is as follows: The Progressive Multi-Core Perception Cascade Module (MPC) is introduced into the Backbone. The MPC module is inserted between two HDM modules at the P4 scale to enhance scale perception capabilities by aggregating multi-scale features step by step. The specific implementation process of MPC is as follows; Let the input feature map be X1. After passing through the convolutional network H, it is added element-wise to the input X1 through residual connections to obtain the output feature map. ; The convolutional kernels of the convolutional network H are progressively larger, and a 1×1 convolution operation is added between every two convolutional layers.

6. The frequency-domain guided multi-scale deformable alignment UAV target detection method according to claim 5, characterized in that, The HDAF cross-scale feature fusion module is introduced into the Neck of the YOLOv8 model. This module utilizes the frequency domain to handle the bias between learning scales, and uses bias-guided deformable convolutions to embed shallow semantic information from the network model into a deep location information map. Finally, gated weighting is performed to complete adaptive fusion, outputting a feature map with accurate localization and high semantic feature information. The specific implementation is as follows: Neck introduces a cross-scale alignment and gated weighted feature fusion module, HDAF, where each HDAF module receives two inputs: detail feature maps. The semantic feature map in Neck with the same resolution after all upsampling or downsampling. ; The detailed feature map Specifically, the Neck consists of four HDAF modules: a first HDAF module, a second HDAF module, a third HDAF module, and a fourth HDAF module. Correspondingly, the detail feature maps (FLs) for each HDAF module are as follows: the first detail feature map (FL1) is the output feature of the backbone network at the P3 scale branch, obtained after channel alignment via convolution; the second detail feature map (FL2) is the output feature of the backbone network at the P2 scale branch, obtained after channel alignment via convolution; the third detail feature map (FL3) is the output feature of the first C2f module in the Neck; and the fourth detail feature map (FL4) is the output feature of the SPPF module in the P4 scale branch of the backbone network, obtained after channel alignment via convolution. The fourth detail feature map (FL4) is simultaneously input to both the first layer of the Neck network and the fourth HDAF module.

7. The frequency-domain guided multi-scale deformable alignment UAV target detection method according to claim 6, characterized in that, The specific implementation process of the cross-scale alignment and gated weighted feature fusion module HDAF is as follows: Let the input feature map be and First, input Frequency domain enhancement is achieved through a frequency-domain-based spatial-frequency hybrid transform operator. Secondly, and After concatenation along the channel dimension, a 1×1 convolution is applied to compress the number of channels to a minimum. With the same number of channels, we get Then The input is a bias and mask prediction layer implemented using convolution to obtain the output. and in order of passage Divided into three parts , , ;in , The spatial offset is obtained by splicing along the channel dimension. ,Will The modulation mask is obtained after a sigmoid operation. ; then with For input, and The modulated deformable convolution operator is used as an alignment parameter input. Deformation alignment calculations are performed to obtain the calibrated feature map. Then and After channel splicing Convolution compression to root The same number of channels are used to obtain the gating weights by combining them with a sigmoid operation. Finally, the gating weights are used to... and Perform element-wise weighted fusion to obtain the output features. .

8. A frequency-domain guided multi-scale deformable alignment UAV target detection system, used to implement the UAV target detection method according to any one of claims 1 to 7, characterized in that, Includes the following modules: The data acquisition module is used to acquire and preprocess drone aerial image datasets. The target detection module is used to improve the YOLOv8 model, perform target detection based on the preprocessed drone aerial image dataset, and output the target detection results. The training and validation module is used to integrate the improved YOLOv8 model into a pre-configured training environment, using training and validation images from the drone aerial image dataset to train and validate the model.

Images