Low-altitude unmanned aerial vehicle detection method based on super-resolution and multi-dimensional attention fusion

Abstract

The application relates to the technical field of image processing, and discloses a low-altitude unmanned aerial vehicle detection method based on super-resolution and multi-dimensional attention fusion, which comprises the following steps: a SRFormer model is used to perform super-resolution reconstruction on an original image to obtain a super-resolution image; a YOLO11-DFA model is trained using the super-resolution image, so that the YOLO11-DFA model outputs a prediction result of a target unmanned aerial vehicle; and the YOLO11-DFA model replaces a C3K2 module with an RCCA module on the basis of a YOLO11 model, and introduces a CDA module before a Head module. The purpose of the application is to significantly improve the detection precision of a light and small target unmanned aerial vehicle in a low-altitude complex environment by synergistically optimizing image quality enhancement and target feature focusing in view of the particularity of low-altitude unmanned aerial vehicle detection.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method for detecting low-altitude unmanned aerial vehicles based on super-resolution and multi-dimensional attention fusion. Background Technology

[0002] With the opening up of low-altitude economic policies, the application of drones in logistics, inspection, and other fields has surged. However, the low-altitude environment presents challenges such as variable lighting and complex backgrounds (e.g., urban buildings and vegetation obstruction). The contradiction between the large-scale application of drones and the needs of airspace safety supervision is becoming increasingly prominent. Traditional drone collision avoidance and early warning methods rely on high-precision radar or laser sensors, which have limitations such as high equipment costs and poor adaptability to complex low-altitude environments. Although visual detection technology partially compensates for these shortcomings through cost advantages, in practical applications, the low signal-to-noise ratio characteristics of long-range, small drone targets in complex backgrounds such as dynamic clouds and urban buildings, as well as the motion blur and multi-angle attitude changes caused by high-speed drone maneuvers, exacerbate the prediction errors of traditional visual detection models, seriously restricting the effectiveness of airspace collaborative management. Existing visual target detection methods, such as YOLO, although having fast inference speeds, lack effective solutions for key issues such as feature degradation of long-range aerial drone targets, sudden changes in lighting, and rotor frequency domain feature extraction.

[0003] Therefore, in order to improve the accuracy and precision of target UAV detection in low-altitude backgrounds, technological innovation is needed to enhance the detection accuracy of low-altitude light and small UAVs under complex background interference, target rotation, and scale changes. Summary of the Invention

[0004] The purpose of this invention is to address the unique characteristics of low-altitude UAV detection by proposing a method based on super-resolution and multi-dimensional attention fusion. By collaboratively optimizing image quality enhancement and target feature focusing, this method significantly improves the detection accuracy of small and light UAVs in complex low-altitude environments.

[0005] To achieve the above-mentioned objectives, the embodiments of the present invention provide the following technical solutions:

[0006] A low-altitude UAV detection method based on super-resolution and multi-dimensional attention fusion includes the following steps:

[0007] The SRFormer model is used to perform super-resolution reconstruction on the original image to obtain a super-resolution image.

[0008] The YOLO11-DFA model is trained using super-resolution images, enabling it to output predictions of the target UAV. The YOLO11-DFA model replaces the C3K2 module with the RCCA module and introduces the CDA module before the Head module, based on the YOLO11 model.

[0009] Compared with existing technologies, the advantages of this invention are as follows: This invention combines a cross-dimensional attention module (CDA) and a region-channel collaborative attention module (RCCA). This multi-dimensional attention fusion strategy significantly improves the detection accuracy of low-altitude, small, and lightweight UAVs under conditions of complex background interference, target rotation, and scale changes. By combining with super-resolution technology, this invention can better capture the edge and texture details of small, lightweight target UAVs, thereby improving detection accuracy. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a schematic diagram of the network structure of the SRFormer model according to an embodiment of the present invention;

[0012] Figure 2 This is a schematic diagram of the network structure of the YOLO11-DFA model in an embodiment of the present invention;

[0013] Figure 3 This is a schematic diagram of the network structure of the RCCA module in an embodiment of the present invention;

[0014] Figure 4 This is a schematic diagram of the MRA layer in an embodiment of the present invention;

[0015] Figure 5 This is a schematic diagram of the CDA module in an embodiment of the present invention. Detailed Implementation

[0016] 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. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0017] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, the terms "first," "second," etc., are used only for distinguishing descriptions and should not be construed as indicating or implying relative importance, or suggesting any such actual relationship or order between these entities or operations. Additionally, the terms "connected," "linked," etc., can refer to a direct connection between elements or an indirect connection via other elements.

[0018] Example:

[0019] This invention is achieved through the following technical solution: a low-altitude unmanned aerial vehicle (UAV) detection method based on super-resolution and multi-dimensional attention fusion, comprising the following steps:

[0020] Step 1: Use the SRFormer model to perform super-resolution reconstruction on the original image to obtain a super-resolution image.

[0021] The original images of four environmental backgrounds—low-coverage sky, city, field, and mountain—were obtained from the open-source Det-fly dataset under different lighting conditions and from multiple perspectives. The length and width of the target drone in the original images were less than 100 and 50 pixels, respectively.

[0022] The acquired original images are randomly divided into two groups. The first group of original images is subjected to bicubic downsampling to obtain the low-resolution image corresponding to the original image. This low-resolution image and the original image form an HR-LR image pair, which is then input into the SRFormer model for training. This allows the SRFormer model to perform super-resolution reconstruction on the original image and output an image with twice the super-resolution of the original image.

[0023] Please see Figure 1The SRFormer model can improve the spatial resolution and detail fidelity of images from lightweight drones. The SRFormer model includes multiple permutation self-attention blocks and a convolutional feedforward network. The multiple permutation self-attention blocks are sequentially connected and then connected to the convolutional feedforward network. The HR-LR image input to the SRFormer model first undergoes pixel-level encoding to obtain the input image feature map X. in ; Input image feature map X in The first permutation self-attention block is input, and after processing through multiple permutation self-attention blocks, it is input into a convolutional feedforward network to obtain the first feature map P; finally, the input image feature map X is processed. in The image is then concatenated with the first feature map P to obtain a super-resolution image.

[0024] Among them, the input image feature map Before inputting the permutation self-attention block (H, W, and C represent the height, width, and number of channels of the input image feature map, respectively), it is first divided into N non-overlapping S×S windows to obtain window features. Subsequently, after the input is permuted from the attention block, it passes through three linear transformation layers L. Q L K L V We obtain the query matrix Q, key matrix K, and value matrix V, in the following specific form:

[0025]

[0026] In this case, Q retains the same input channel dimension, while the channel dimensions of K and V are compressed to C / r. 2 r is the token reduction factor to reduce computational complexity.

[0027] To increase the number of tokens involved in the self-attention mechanism's computation while controlling computational resources, the spatial tokens in K and V are replaced with channel dimensions in the channel permutation mechanism, thereby obtaining the permuted features. Using Q and K p V p Self-attention calculation is performed, and relative position encoding B is introduced to maintain spatial awareness. The self-attention calculation formula is as follows:

[0028]

[0029] PSA() represents the operation of mapping spatial information to the channel dimension, which needs to meet the following two key points: First, avoid downsampling from interfering with the original feature expression, and each token participates in attention calculation independently to improve the ability to recover local details; Second, support large window modeling and lower computational overhead. Compared with models such as SwinIR that use 8×8 windows, this solution can use 24×24 windows, which reduces computational complexity while maintaining a larger receptive field.

[0030] Step 2: Train the YOLO11-DFA model using super-resolution images so that the YOLO11-DFA model outputs the prediction results of the target UAV; the YOLO11-DFA model is based on the YOLO11 model, replacing the C3K2 module with the RCCA module and introducing the CDA module before the Head module.

[0031] As a first possible implementation, the super-resolution image output by the SRFormer model is input into the YOLO11-DFA model for training to obtain the first YOLO11-DFA model; the second set of original images from step 1 is input into the YOLO11-DFA model for training to obtain the second YOLO11-DFA model. Real-time images are then input into the first and second YOLO11-DFA models respectively. It is found that the first YOLO11-DFA model provides better prediction results for the target UAV; therefore, the first YOLO11-DFA model is directly used as the YOLO11-DFA model trained in step 2.

[0032] As a second possible implementation, the second set of original images from step 1 are input into the YOLO11-DFA model for training to obtain the first YOLO11-DFA model; then the super-resolution images output by the SRFormer model are input into the first YOLO11-DFA model for training to obtain the second YOLO11-DFA model, thus obtaining the YOLO11-DFA model trained in step 2.

[0033] As a third possible implementation, the second set of original images from step 1 and the super-resolution images output by the SRFormer model are input together into the YOLO11-DFA model for training, resulting in the YOLO11-DFA model trained in step 2.

[0034] As can be seen, all three methods can train a YOLO11-DFA model, enabling the YOLO11-DFA model to output the prediction results of the target UAV. However, considering factors such as overfitting of training samples, effectiveness of super-resolution processing, and model robustness, the first feasible implementation method is preferred in this embodiment.

[0035] like Figure 2As shown, the YOLO11-DFA model includes a backbone network, a neck network, and a head network.

[0036] The backbone network comprises, in sequence, a first CBS module, a second CBS module, a first C3K2 module, a third CBS module, a second C3K2 module, a fourth CBS module, an RCCA module, a fifth CBS module, a third C3K2 module, an SPPF module, and a C2PSA module. It should be noted that "first," "second," etc., are used only for ease of describing the connection relationships, and the network structures of "first CBS" and "second CBS" are identical; therefore, when describing their specific functions or internal structures, they are collectively referred to as "CBS," and the same applies to others. Each CBS module includes a Conv layer, a BN layer, and a SiLU activation function, used to progressively extract features from the reconstructed super-resolution image. In the backbone network, the C3K2 module further enhances the feature representation output by the CBS modules.

[0037] The improvement to the backbone network lies in replacing the third C3K2 module in the traditional YOLOv11 model with an RCCA module (Region-Channel Collaborative Attention). This module uses multiple local region divisions (MRA) to implement a multi-head attention mechanism to extract local salient features, and introduces a task-ware weighted mechanism to enhance channel performance. Figure 3 As shown, the RCCA module incorporates an MRA (Multi-Region Attention) layer to focus on key regions and highlight important features, providing more accurate information for subsequent detection. Following the RCCA module, the CBS module is used again to further mine features. The SPPF (Spatial Pyramid Pooling) module integrates information at different scales, enhancing the backbone network's adaptability to target size. Finally, the C2PSA (Channel and Spatial Information Extraction) module further synthesizes features, laying the foundation for multi-scale target detection.

[0038] like Figure 4 As shown, the MRA layer inside the RCCA module divides different regions of the image using Local Self-Attention, focusing on the feature relationships within each region and enhancing the expression of local features. Simultaneously, weights are allocated along the channel dimension, rationally distributing channel resources according to the importance of different regions, highlighting feature channels in key regions and suppressing channels in unimportant regions. This enhances the expression of both local and global features, improving the network's ability to detect occluded targets and ensuring accurate detection even when low-altitude targets, such as drones, are obstructed.

[0039] The calculation formula for the RCCA module is as follows:

[0040]

[0041] Where F is the feature map of the input RCCA module; Reshape() is the reshaping operation; F area The reshaped features are shown below. This process reshapes the feature map F into multiple regions of size A×A (default A=4), which are used to divide the feature map F into regions, preparing for subsequent multi-region attention calculations.

[0042]

[0043] Where Linear() is the linear transformation operation; split(3) means dividing the linearly transformed features into three matrices: query Q, key K, and value V; N h Let be the dimension of the key vector. This process uses a linear transformation to obtain the query Q, key K, and value V matrix, which is then used to compute multi-region attention.

[0044]

[0045] Here, Att(Q,K,V) represents the multi-region attention feature; Softmax() is the Softmax activation function. This process calculates the attention score by multiplying the query and key by the dot product, then normalizes it using the Softmax function, and finally multiplies it by the value matrix to obtain the multi-region attention feature, thus enhancing the expression of local features.

[0046]

[0047] Among them, F attn This represents the concatenated multi-region attention features; `concat` is the concatenation operation. This process concatenates the multi-region attention features back to the original feature map's dimensions C×H×W for easier subsequent operations.

[0048]

[0049] Among them, F pos This is the feature map after incorporating location information; DepthwiseConv() is a depthwise separable convolution operation, implemented in each MAR layer. This process adds the multi-region attention features to the value matrix after depthwise separable convolution, incorporating location information and enhancing the spatial representation of the features.

[0050]

[0051] Among them, F out The feature map output by the RCCA module; Conv 1×1 This is a 1×1 convolution operation. This process performs channel compression and feature fusion on the feature map using 1×1 convolution to obtain the final output feature map F.out This enables the enhancement of both local and global features.

[0052] Please continue reading Figure 2 In the neck network, the features output by the C2PSA module are upsampled and concatenated with the features output by the RCCA module to obtain feature map P1. Feature map P1 is then processed by the fourth C3K2 module to obtain feature map P2. Feature map P2 is upsampled and concatenated with the features output by the second C3K2 module to obtain feature map P3. Next, feature map P3 is processed by the fifth C3K2 module to obtain feature map P4. Feature map P4 is processed by the sixth CBS module and concatenated with feature map P2 to obtain feature map P5. Finally, feature map P5 is processed by the sixth C3K2 module to obtain feature map P6. Through multiple upsampling, concatenation, and C3K2 module processing in the neck network, a multi-scale feature fusion path is constructed, enabling the neck network to capture feature information of the target at different scales. This effectively improves the ability to detect both small and large targets simultaneously, achieving effective extraction of multi-scale features and preparing for subsequent detection.

[0053] In the head network, feature map P6 is input into the first CDA module and then passed through the first Head module to obtain the first prediction result. Feature map P4 is input into the second CDA module and then passed through the second Head module to obtain the second prediction result. The resolution of the first prediction result is twice that of the second prediction result. Both prediction results contain information about the target location, category, and confidence level.

[0054] The improvement in the head network lies in the introduction of two CDA modules (Cross-Dimensional Attention modules), such as... Figure 5 The CDA module enhances small target discrimination capabilities through a combination of depthwise separable convolution (DWConv), channel attention, and spatial attention. Channel attention recalibrates features along the channel dimension, weighting them according to their importance to enhance key feature channels and suppress less important ones. Spatial attention focuses on key regions of the target in the spatial dimension, highlighting features at the target's location while ignoring irrelevant background information. By fusing channel and spatial attention information, feature representation is further optimized, making features more discriminative and providing more accurate feature support for subsequent target classification and localization, thus improving detection accuracy.

[0055] The calculation formula for the CDA module is:

[0056]

[0057] Where G is the feature map input to the CDA module; DWConv() is the depthwise separable convolution operation. Feature map G is processed by depthwise separable convolution to obtain feature map F. dw This process reduces computation and parameter count while extracting spatial features.

[0058]

[0059] Where W1 and W2 are learnable weight matrices; GELU() is the Gaussian error linear unit activation function; AvgPool() is the global average pooling operation; The Sigmoid function is used. This process extracts global information in the channel dimension through global average pooling, introduces nonlinearity using the Gaussian error linear unit activation function, and finally obtains the channel attention weights C through a linear transformation of the Sigmoid function. A (F dw ), used to calibrate feature channels.

[0060]

[0061] Among them, Conv k×k For k×k convolution operations; concat[] is the concatenation operation; Avg() is the average pooling operation; Max() is the max pooling operation. This process concatenates the features after average pooling and max pooling, uses k×k convolution for channel compression and feature fusion, and finally obtains the spatial attention weight S through the Sigmoid function. A (F dw (), used to highlight the spatial characteristics of the target area.

[0062]

[0063] Among them, F CDA For fusion of feature maps; C A (F dw ) represents the channel attention weight, and Channel represents the channel dimension; S A (F dw ) represents the spatial attention weight, and Spatial represents the spatial dimension; This is an element-wise multiplication. The process applies channel attention and spatial attention to the input feature map F and the depthwise separable convolutional feature map F, respectively. dw Then, the feature map F is fused in the channel dimension and the spatial dimension to obtain the feature map F that integrates channel and spatial attention information. CDA This enhances the discriminative power of features.

[0064] The improved YOLO11-DFA model identifies features of small targets and sky-covering areas in low-altitude UAV images. The network structure is optimized by removing redundant large target detection branches, reducing computational complexity and avoiding unnecessary resource waste. Network resources are focused on the extraction and detection of small target features, enabling the network to more accurately capture the features of low-altitude small UAV targets, improving detection accuracy, and achieving effective detection of low-altitude small targets to meet the practical needs of low-altitude surveillance.

[0065] After feature extraction and processing are completed by the first and second CDA modules, the first and second Head modules perform final analysis and processing of the features, outputting the target's location information (such as the coordinates of the bounding box), category information (determining whether it is a drone), and confidence information (the reliability of the detection result). This information provides accurate basis for low-altitude airspace supervision, enabling timely detection and tracking of low-altitude drone targets, achieving safe monitoring and real-time early warning of low-altitude airspace, effectively ensuring low-altitude flight safety, and providing strong support for decision-making by relevant regulatory departments.

[0066] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A low-altitude UAV detection method based on super-resolution and multi-dimensional attention fusion, characterized in that, Includes the following steps: The SRFormer model is used to perform super-resolution reconstruction on the original image to obtain a super-resolution image. The YOLO11-DFA model is trained using super-resolution images, enabling the YOLO11-DFA model to output the prediction results of the target UAV. The YOLO11-DFA model is based on the YOLO11 model, with the C3K2 module replaced by the RCCA module and the CDA module introduced before the Head module. The YOLO11-DFA model includes a backbone network, a neck network, and a head network. The backbone network includes a first CBS module, a second CBS module, a first C3K2 module, a third CBS module, a second C3K2 module, a fourth CBS module, an RCCA module, a fifth CBS module, a third C3K2 module, an SPPF module, and a C2PSA module connected in sequence. The neck network includes a fourth C3K2 module, a fifth C3K2 module, a sixth CBS module, and a sixth C3K2 module. Features output from the C2PSA module are upsampled and concatenated with features output from the RCCA module to obtain feature map P1. Feature map P1 is then processed by the fourth C3K2 module to obtain feature map P2. Feature map P2 is upsampled and concatenated with features output from the second C3K2 module to obtain feature map P3. Next, feature map P3 is processed by the fifth C3K2 module to obtain feature map P4. Feature map P4 is processed by the sixth CBS module and concatenated with feature map P2 to obtain feature map P5. Finally, feature map P5 is processed by the sixth C3K2 module to obtain feature map P6. The head network includes a first CDA module, a second CDA module, a first Head module, and a second Head module; feature map P6 is input into the first CDA module and then passes through the first Head module to obtain a first prediction result, and feature map P4 is input into the second CDA module and then passes through the second Head module to obtain a second prediction result.

2. The low-altitude UAV detection method based on super-resolution and multi-dimensional attention fusion according to claim 1, characterized in that, The SRFormer model includes multiple permutation self-attention blocks and a convolutional feedforward network. The multiple permutation self-attention blocks are connected sequentially and then connected to the convolutional feedforward network. The steps to obtain a super-resolution image specifically include: The original image is subjected to bicubic downsampling to obtain a low-resolution image corresponding to the original image. This low-resolution image and the original image form an HR-LR image pair. The HR-LR image is first input into the SRFormer model and then subjected to pixel-level encoding to obtain the input image feature map X. in ; Input image feature map X in The first permutation self-attention block is input, and after processing through multiple permutation self-attention blocks, it is input into a convolutional feedforward network to obtain the first feature map P; finally, the input image feature map X is processed. in The image is then concatenated with the first feature map P to obtain a super-resolution image.

3. The low-altitude UAV detection method based on super-resolution and multi-dimensional attention fusion according to claim 2, characterized in that, Input image feature map X in The processing procedure for the first permutation self-attention block is as follows: Input image feature map Before inputting the permutation self-attention block, it is first divided into N non-overlapping S×S windows to obtain window features. H, W, and C represent the height, width, and number of channels of the input image feature map, respectively; after the input is permuted from the attention block, it passes through three linear transformation layers L. Q L K L V We obtain the query matrix Q, key matrix K, and value matrix V, in the following specific form: In this case, Q retains the same input channel dimension, while the channel dimensions of K and V are compressed to C / r. 2 r is the token reduction factor; In the channel permutation mechanism, the spatial tokens in K and V are replaced with channel dimensions to obtain the permuted features. Using Q and K p V p Self-attention calculation is performed, and relative position encoding B is introduced to maintain spatial awareness. The self-attention calculation formula is as follows: Where PSA() represents the operation of mapping spatial information to the channel dimension; T represents matrix transpose; d k Represents the vector dimension.

4. The low-altitude UAV detection method based on super-resolution and multi-dimensional attention fusion according to claim 1, characterized in that, The RCCA module includes multiple MRA layers and a task-ware mechanism; the calculation process of the RCCA module is as follows: Where F is the feature map of the input RCCA module; Reshape() is the reshaping operation; F area A represents the reshaped feature; A represents the reshaped dimension. Where Linear() is the linear transformation operation; split(3) means dividing the linearly transformed features into three matrices: query Q, key K, and value V; N h The dimension of the key vector; Where Att(Q,K,V) represents the multi-region attention feature; Softmax() is the Softmax activation function; Among them, F attn This represents the multi-region attention features after concatenation; concat is the concatenation operation. Among them, F pos This is the feature map after incorporating location information; DepthwiseConv() is a depthwise separable convolution operation. Among them, F out The feature map output by the RCCA module; Conv 1×1 This is a 1×1 convolution operation.

5. The low-altitude UAV detection method based on super-resolution and multi-dimensional attention fusion according to claim 1, characterized in that, The calculation formula for the CDA module is as follows: Among them, F dw G is the feature map obtained after depthwise separable convolution; G is the feature map input to the CDA module; DWConv() is the depthwise separable convolution operation; Among them, C A (F dw W1 and W2 are the channel attention weights; GELU() is the Gaussian error linear unit activation function; AvgPool() is the global average pooling operation. For the Sigmoid function; Among them, S A (F dw ) represents the spatial attention weights; Conv k×k `k×k` is the k×k convolution operation; `concat[]` is the concatenation operation; `Avg()` is the average pooling operation; `Max()` is the max pooling operation. Among them, F CDA To fuse feature maps; This is for element-wise multiplication.

Images