Unmanned aerial vehicle small target real-time detection method and system for foggy environment

Abstract

The invention provides an unmanned aerial vehicle small target real-time detection method and system oriented to a foggy day environment. The method comprises the following steps: constructing an AERO-YOLO target detection model comprising a backbone feature extraction network, an image adaptive feature enhancement module and a feature fusion module; the detection performance is improved by adopting a channel and space attention fused feature enhancement mechanism; structured pruning and layer fusion are carried out on the basis of a gamma coefficient of a Batch Normalization parameter, so that lightweight of the model is realized; and deploying the optimized model to an unmanned aerial vehicle end side, and outputting a detection result in real time. According to the method, fog noise interference is effectively suppressed, small target edge and texture features are amplified, the method is particularly excellent in low-contrast and small-size target detection tasks, and the adaptability in a foggy environment is higher; the problems of difficult feature extraction, limited model deployment, insufficient generalization ability and the like in unmanned aerial vehicle small target detection in a foggy environment are solved, and the detection precision and the real-time performance are both considered.

Description

Technical Field

[0001] This invention relates to the field of target detection technology, and in particular to a method and system for real-time detection of small targets by unmanned aerial vehicles (UAVs) in foggy environments. Background Technology

[0002] With the widespread application of drones in fields such as intelligent transportation, public safety, emergency rescue, and environmental monitoring, target detection technology based on drone platforms has gradually become an important direction in computer vision research. However, in low-visibility environments such as fog, atmospheric scattering leads to decreased image contrast, blurred edges, and degraded textures. In particular, small targets are easily confused with the background, making it difficult for detection models to accurately extract effective features, resulting in missed detections and false detections.

[0003] Existing fog-based target detection methods mainly fall into two categories: one separates image dehazing and target detection, but the dehazing process often introduces noise and weakens the true target features. The other employs feature enhancement or attention mechanisms, but these models are complex and computationally intensive, making real-time deployment on embedded UAV devices difficult. Traditional detection algorithms such as YOLO, SSD, and Faster-RCNN are mostly trained on clear images, lacking adaptability to fog-induced degradation features, especially struggling to identify small, distant targets. Even with fog adaptation modules, they often rely on high computational resources, which is incompatible with the lightweight requirements of UAV platforms. Furthermore, real-world fog data acquisition is costly, labeling is difficult, samples are limited, and images vary significantly under different fog concentrations, resulting in insufficient model generalization ability. While synthetic fog image methods based on atmospheric scattering models can expand data, without incorporating depth information, the generated fog distribution is not realistic enough to simulate the degradation features from a real UAV perspective.

[0004] Therefore, there is an urgent need for a small target detection method in foggy weather that can enhance the degradation characteristics of foggy weather, is lightweight and efficient, and can be operated in real time on the UAV side to improve detection accuracy and deployment practicality. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a real-time detection method and system for small targets on UAVs in foggy environments. It solves problems such as difficulty in feature extraction, limited model deployment, and insufficient generalization ability in small target detection on UAVs in foggy environments, achieving a balance between detection accuracy and real-time performance, and meeting the real-time detection requirements of UAVs.

[0006] The technical solution of this invention is: a real-time detection method for small targets by unmanned aerial vehicles (UAVs) in foggy environments, comprising the following steps:

[0007] S1) Acquire aerial images at different distances and fog densities and preprocess them;

[0008] S2) Construct an AERO-YOLO target detection model consisting of a backbone feature extraction network, an image adaptive feature enhancement module, and a feature fusion module;

[0009] S3) Train the AERO-YOLO object detection model. During the training process, the image adaptive feature enhancement module and the AERO-YOLO object detection model participate in forward propagation and backward propagation together.

[0010] During the inference phase, the image adaptive feature enhancement module applies a joint weighting of channel attention and spatial attention to the input feature map using fixed parameters obtained from training. The fused feature enhancement mechanism suppresses fog noise interference and amplifies the edge and texture features of small targets.

[0011] S4) The trained AERO-YOLO object detection model is pruned and lightweighted using the model pruning and lightweighting optimization unit;

[0012] S5) Deploy the optimized lightweight AERO-YOLO target detection model to the UAV edge computing platform to detect the target category, location and confidence level in real time in foggy remote sensing environment, and output the detection box to the UAV flight control system for target localization and path correction.

[0013] Preferably, in step S2), the image adaptive feature enhancement module includes a channel global statistical extraction unit, an attention weight generation unit, and a feature saliency enhancement unit.

[0014] Preferably, in step S2), the channel global statistical extraction unit is used to perform global average pooling on the input feature map to extract the saliency distribution of fog noise, thereby obtaining the sensitivity of different channels to small target information.

[0015] Preferably, in step S2), the attention weight generation unit generates channel attention weights based on channel statistics sensitivity through a two-layer mapping network to distinguish between fog noise regions and effective target regions.

[0016] Preferably, in step S2), the feature saliency enhancement unit is used to broadcast the channel attention weights and apply them to the original feature map, suppress channels dominated by fog interference, and focus on enhancing channels containing small target features.

[0017] Preferably, in step S4), the model pruning and lightweight optimization unit includes a channel importance evaluation unit, a redundant channel screening unit, and a convolution and normalization fusion unit.

[0018] Preferably, in step S4), the channel importance evaluation unit calculates the contribution of each convolutional channel to the target detection task based on the scaling factor of the BatchNormalization layer.

[0019] Preferably, in step S4), the redundant channel removal unit prunes the channels according to a preset pruning ratio based on the channel importance index. Sort the channels and remove low-contribution channels.

[0020] Preferably, in step S4), the convolution and normalization fusion unit structurally fuses the convolutional layer and the BatchNormalization layer to reduce inference computation overhead.

[0021] In a second aspect, the present invention provides a real-time small target detection system for unmanned aerial vehicles (UAVs) in foggy environments, comprising:

[0022] The image acquisition module is used to acquire aerial images at different distances and with different fog densities and to preprocess them;

[0023] The model building module is used to build and train the AERO-YOLO object detection model, which consists of a backbone feature extraction network, an image adaptive feature enhancement module, and a feature fusion module.

[0024] The model training module trains the AERO-YOLO object detection model. During the training process, the image adaptive feature enhancement module participates in forward and backward propagation together with the AERO-YOLO object detection model.

[0025] During the inference phase, the image adaptive feature enhancement module applies a joint weighting of channel attention and spatial attention to the input feature map using fixed parameters obtained from training. The fused feature enhancement mechanism suppresses fog noise interference and amplifies the edge and texture features of small targets.

[0026] The model optimization module prunes and lightweights the trained AERO-YOLO object detection model, and then deploys the optimized and lightweight AERO-YOLO object detection model to the UAV edge computing platform.

[0027] The real-time detection module calls the AERO-YOLO target detection model in foggy remote sensing environments to detect the target's category, location, and confidence level in real time, and outputs the detection box to the UAV flight control system for target localization and path correction.

[0028] The beneficial effects of this invention are as follows:

[0029] 1. This invention effectively suppresses fog noise interference and amplifies the edge and texture features of small targets by combining the image adaptive feature enhancement module with channel attention and spatial attention mechanisms. It performs particularly well in low-contrast, small-sized target detection tasks, with detection accuracy comparable to mainstream lightweight models, and is more adaptable to foggy environments.

[0030] 2. This invention achieves model lightweighting through structured pruning and layer fusion based on the Batch Normalization parameter γ coefficient, reducing the number of parameters by more than 35% and FLOPs by more than 22%. After TensorRT inference acceleration, the inference speed on the Jetson Orin NX embedded platform can reach 19 FPS, which is about 30%-40% higher than YOLOv8n, meeting the real-time detection requirements of UAV edge devices.

[0031] 3. This invention combines monocular depth estimation with atmospheric scattering models to generate synthetic foggy images. The generated images have real atmospheric scattering characteristics and can simulate foggy environments with different fog paths and concentrations. This solves the problems of limited real foggy data samples and difficulty in annotation, and significantly improves the model's generalization ability.

[0032] 4. After being lightweighted, the model of this invention can be deployed on various UAV end-side embedded platforms such as Jetson Orin NX and RK3588, and outputs detection results to the flight control system in real time for target positioning and path correction. It is suitable for multiple practical application scenarios such as intelligent transportation, public safety, and emergency rescue, and has broad application prospects. Attached Figure Description

[0033] Figure 1 This is a flowchart illustrating the method of Embodiment 1 of the present invention;

[0034] Figure 2 This is a schematic diagram of the framework of the AERO-YOLO target detection model in Embodiment 1 of the present invention;

[0035] Figure 3 This is a comparison diagram of Embodiment 1 of the present invention and the baseline model in a foggy scene;

[0036] Figure 4 This is a comparison chart of the detection speed and detection accuracy of Embodiment 1 of the present invention with other models;

[0037] Figure 5 This is a schematic diagram of the system framework of Embodiment 2 of the present invention. Detailed Implementation

[0038] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings:

[0039] Example 1

[0040] like Figure 1 As shown, this embodiment provides a real-time detection method for small targets by UAVs in foggy environments, including the following steps:

[0041] S1) Aerial images taken at different distances and with different fog densities are obtained by using a visible light camera device mounted on the drone and then preprocessed.

[0042] In this embodiment, aerial images under clear weather conditions are first acquired using a visible light camera mounted on a drone. These images are then combined with the publicly available VisDrone2019 dataset and self-acquired small target images from the drone as basic data samples. The training sample set is then expanded by processing the basic data samples accordingly, as follows:

[0043] S11) Extract the pixel-level transmittance distribution of aerial images using a monocular depth estimation model;

[0044] The monocular depth estimation model is an end-to-end depth regression model based on a convolutional neural network; by using fog-free images... The depth is estimated by inputting the data into the monocular depth estimation model, and the corresponding depth map is obtained. Based on the aforementioned depth map Construct pixel-level transmittance distribution ,Right now:

[0045] ; (1)

[0046] In the formula, This is the fog attenuation coefficient; Represents pixel position coordinates;

[0047] S12) Based on the atmospheric scattering model, fog concentration distribution characteristics are constructed, and composite fog images with multiple fog paths and concentrations are generated; specifically as follows:

[0048] S121), given an atmospheric light value Under these conditions, the atmospheric scattering model is based on the transmittance distribution. For fog-free images Degradation is performed to generate the corresponding foggy image. ,Right now:

[0049] ; (2)

[0050] In the formula, This refers to the atmospheric light value. This is the fog attenuation coefficient; Transmittance distribution.

[0051] S122) By randomly sampling different fog attenuation coefficients within a preset range With air light value The system generates synthetic fog images with various fog paths and fog concentration levels, and combines these synthetic fog images with real fog drone images to form a training sample set for fog target detection.

[0052] S2) Construct an AERO-YOLO object detection model consisting of a backbone feature extraction network, an image adaptive feature enhancement module, and a feature fusion module; and train the AERO-YOLO object detection model using training samples.

[0053] like Figure 2 As shown, in this embodiment, the backbone feature extraction network includes:

[0054] Input convolutional layer: The input 640×640 scale UAV foggy image undergoes initial convolution, normalization, and non-linear activation to map the three-channel RGB image to a high-dimensional feature space, obtaining an initial feature map containing low-level information such as edges and textures. This provides a foundation for subsequent multi-scale feature extraction.

[0055] First-stage downsampling and C2f feature extraction module: using the initial feature map As input, the initial feature map is processed by convolution or downsampling operations with a stride of 2. Downsampling is performed, and a set of feature extraction units based on the C2f structure are cascaded together to convert the initial feature map... The spatial resolution is reduced to approximately 1 / 8 the scale of the original input image to obtain a shallow feature map. The C2f module is a core structure in the YOLOv8 backbone. It enhances gradient flow through feature branching and residual aggregation, and can fully extract edge and contour information of the image while keeping the number of parameters small, thus reducing the interference of fog noise on shallow features.

[0056] The second-stage downsampling and C2f deep feature extraction module: using the shallow feature map output from the first stage... As input, continue to use convolutions with a stride of 2 on the feature map. Downsampling is performed, and several C2f modules are stacked to reduce the spatial resolution of the feature map to approximately 1 / 16 of the scale of the original input image, resulting in the intermediate-level feature map. This stage expands the receptive field while preserving the target's shape and structure, which is beneficial for modeling the contextual relationship between small targets and the background in foggy scenes.

[0057] Third-stage downsampling and spatial pyramid pooling fast module SPPF: using the mid-layer feature map output from the second stage. As input, the middle layer feature map is processed by a convolution with a stride of 2. Downsampling was performed to obtain a deep feature map at a scale of approximately 1 / 32 of the original input image. Furthermore, a fast spatial pyramid pooling module is introduced to perform multi-scale pooling and aggregation of features under different receptive fields. This module integrates a wide range of global contextual information without significantly increasing computational cost, enabling the network to maintain its overall perception capability for distant small targets even in complex foggy scenes.

[0058] Multi-scale feature output layer: The backbone feature extraction network ultimately outputs three feature maps of different resolutions from the above three stages, namely shallow feature maps. Mid-layer feature map and deep feature map These correspond to approximately 1 / 8, 1 / 16, and 1 / 32 scales of the original image, respectively. These three feature layers simultaneously preserve both shallow edge details and deep semantic information. They are fed as input into the image adaptive feature enhancement module and subsequent feature fusion module for multi-scale detection of small targets in foggy weather.

[0059] This embodiment embeds the image adaptive feature enhancement module between the backbone feature extraction network and the feature fusion module to suppress fog interference and enhance the feature representation of small targets. Through the image adaptive feature enhancement module, the features of small targets and high-contrast regions are amplified, while the noise response of fog regions is suppressed, improving the detection performance of the model in foggy environments. The image adaptive feature enhancement module uses a feature enhancement mechanism based on the fusion of channel attention and spatial attention to suppress fog noise interference and amplify the edge and texture features of small targets, thus improving the detection performance of small targets in foggy conditions. The fusion module is used to realize multi-scale feature interaction and detection.

[0060] The image adaptive feature enhancement module includes a channel global statistical extraction unit, an attention weight generation unit, and a feature saliency enhancement unit; wherein:

[0061] The aforementioned channel global statistical extraction unit is used to perform global average pooling on the input multi-scale feature map to extract the saliency distribution of fog noise, thereby obtaining the sensitivity of different channels to small target information.

[0062] ; (3)

[0063] In the formula, The multi-scale feature map output by the backbone feature extraction network; This indicates a global average pooling operation; This is a channel statistics vector;

[0064] The attention weight generation unit generates channel attention weights based on channel statistical vectors using a two-layer mapping network to distinguish between fog noise regions and effective target regions, thereby achieving feature selective enhancement. Its calculation formula is as follows:

[0065] ; (4)

[0066] in, These are the trainable weights. It is the ReLU activation function. It is the Sigmoid activation function. The generated channel attention weights.

[0067] The feature saliency enhancement unit is used to broadcast the channel attention weights and apply them to the multi-scale feature map to obtain the channel-enhanced feature map. Channels dominated by fog interference are suppressed, while channels containing small target features are enhanced to improve the feature representation capability of small target detection. The calculation formula is as follows:

[0068] ; (5)

[0069] Where ⊙ represents a channel-by-channel multiplication operation;

[0070] In channel enhancement feature map Above, average pooling and max pooling are performed along the channel dimension respectively to obtain two single-channel spatial feature maps. and The data is then stitched together along the channel dimension and then passed through a convolutional layer. Spatial attention map generated with Sigmoid activation function ,Right now:

[0071] ; (6)

[0072] The spatial attention map assigns lower weights to large, low-information fog background regions and higher weights to salient regions containing small targets.

[0073] Spatial attention map Broadcast along the channel dimension and then applied to the channel-enhanced feature map. The final enhanced features are obtained by adding them to the multi-scale feature maps using a residual method. ,Right now:

[0074] ; (7)

[0075] Where ⊙ represents element-wise multiplication;

[0076] The aforementioned channel and spatial attention joint weighting operations are applied to shallow features. Mid-layer features and deep features The corresponding final enhanced features are obtained. , , ; These are shallow enhancement features. Mid-layer enhancement features Deep enhancement features The final enhancement feature mentioned above. While suppressing fog noise interference, it significantly improves the contrast and distinguishability of small targets in foggy weather on the feature map, and sends them as input to the feature fusion module for detection.

[0077] In this embodiment, the feature fusion module performs multi-scale feature interaction and detection as follows:

[0078] 1) Top-down feature propagation: The input is the final enhanced feature output by the feature saliency enhancement unit, i.e., the shallow enhanced feature. Mid-layer enhancement features Deep enhancement features ;

[0079] Deep enhancement features After 1×1 convolution compression, features are enhanced by 2x upsampling and mid-layer enhancement. The feature maps are then fused and refined by the C2f module to obtain a top-down fused mid-layer feature map. ;

[0080] Then the mid-layer fusion feature map Upsampling and shallow enhancement features The fusion process is refined to obtain a shallow fusion feature map obtained from top to bottom fusion. This approach injects high-level semantic information into the shallow layer, enhancing the context awareness of small targets.

[0081] 2) Bottom-up detail aggregation: To enhance the bidirectional flow of features, shallow feature maps rich in detail are fused together. Feature map after downsampling and fusion with middle layer After fusion and restructuring by the C2f module, enhanced mid-layer fusion features are obtained. Then enhance the mid-layer fusion features. Downsampling and deep enhancement features The fusion process, followed by reorganization via the C2f module, yields enhanced deep fusion features. Simultaneously, the shallow fusion feature map obtained from top-down fusion is... Directly used as an enhanced shallow fusion feature This bottom-up approach feeds shallow details back to deep features, enabling integrated modeling of multi-scale targets.

[0082] 3) Multi-scale detection output: The final output consists of three layers of fused features, namely: a shallow fused feature that emphasizes small targets at close range. Enhanced mid-level integration features that take into account both small and medium-sized objectives Enhanced deep fusion features with a focus on large, distant targets These features combine detailed textures with deep semantics, effectively suppressing fog noise, improving the response to small targets, and providing more discriminative input for subsequent detection heads.

[0083] S3) The AERO-YOLO target detection model is trained using the training set. During the training process, the image adaptive feature enhancement module participates in forward propagation and backward propagation together with the AERO-YOLO target detection model. In the inference stage, the image adaptive feature enhancement module uses the fixed parameters obtained from training to apply a joint weighting of the input feature map based on channel attention and spatial attention. The fused feature enhancement mechanism suppresses fog noise interference and amplifies the edge and texture features of small targets.

[0084] S4) The trained AERO-YOLO object detection model is pruned and lightweighted using the model pruning and lightweighting optimization unit;

[0085] In this embodiment, the model pruning and lightweight optimization unit includes a channel importance evaluation unit, a redundant channel removal unit, and a convolution and normalization fusion unit; wherein:

[0086] The channel importance evaluation unit calculates the contribution of each convolutional channel to the object detection task using a scaling factor based on the Batch Normalization layer, i.e.:

[0087] ; (8)

[0088] In the formula, For the first Key metrics for each channel; This refers to the scaling parameters for the corresponding Batch Normalization layer.

[0089] The redundant channel removal unit prunes channels according to a preset pruning ratio based on channel importance indicators. The channels are sorted and low-contribution channels are removed, with the following pruning conditions:

[0090] ; (9)

[0091] in, The threshold is dynamically determined based on the pruning ratio; below the threshold... The channels are marked as redundant channels.

[0092] The convolution and normalization fusion unit fuses the convolutional layer and the batch normalization layer to reduce inference computation overhead. The fusion computation method is as follows:

[0093] ; (10)

[0094] ; (11)

[0095] in, These represent the weights and biases of the fused convolutional layer. These are the original convolution weights. These are the mean and variance, respectively. These are the scaling and translation parameters for Batch Normalization; It is the stability constant.

[0096] S5) Deploy the optimized lightweight AERO-YOLO target detection model to the UAV edge computing platform to detect the target category, location and confidence level in real time in foggy remote sensing environment, and output the detection box to the UAV flight control system for target localization and path correction.

[0097] This embodiment deploys a lightweight AERO-YOLO object detection model on the Jetson Orin NX platform, converts it to the TensorRT engine format for inference acceleration, and compares it with the YOLOv8n model;

[0098] As shown in Table 1, this embodiment reduces the number of model parameters by about 35% and FLOPs by about 22%, but the detection accuracy decreases by no more than 1%, achieving a balance between accuracy and real-time performance.

[0099] Table 1 Performance comparison between the model in this embodiment and existing models

[0100]

[0101] To verify the effectiveness of the small target UAV detection method in foggy weather presented in this embodiment, systematic experiments were conducted to evaluate its performance in terms of detection accuracy, inference speed, model complexity, fog adaptability, and embedded deployment performance. Experiments were performed on a hybrid foggy weather dataset constructed using VisDrone2019 and a self-collected foggy small target dataset. The training environment used an RTX 4070 GPU, and deployment and speed testing were performed on a Jetson Orin NX embedded device. The method in this embodiment was compared with mainstream lightweight detection models such as YOLOv5s, YOLOv7-tiny, PP-YOLOE-s, and YOLOv8n. Figure 3 Images (a) and (b) in the figure show the detection results of the AERO-YOLO object detection model in Example 1 of the baseline model; from Figure 3 As shown in (a) and (b), the method in this embodiment can stably detect small targets such as pedestrians and vehicles at a distance under various visibility conditions, including dense fog and light fog, while maintaining a similar detection capability for large targets such as buses. Compared with the baseline model, it has a lower false negative rate for small targets, and the detection speed is improved by about 30% to 40% compared with YOLOv8n, achieving a good balance between detection accuracy and real-time performance, and exhibiting better robustness in low-contrast categories in foggy weather.

[0102] like Figure 4 As shown in (a) and (b) in the figure; the method of this embodiment achieves an inference speed of about 20 FPS while maintaining a similar mAP@0.5 (about 30%), which is about 30% to 40% higher than YOLOv8n. It achieves a good balance between detection accuracy and real-time performance, especially with better detection accuracy in low contrast categories and better generalization ability in foggy environments.

[0103] As shown in Table 2, the AERO-YOLO target detection model in this embodiment maintains a detection accuracy similar to that of mainstream models (mAP@0.5=29.4%) while significantly reducing the number of parameters and FLOPs, and the inference speed reaches 19 FPS, which is about 35.7% higher than YOLOv8n, achieving a good balance between detection accuracy and real-time performance.

[0104] Table 2. Detection speed and accuracy of different models

[0105]

[0106] Example 2

[0107] like Figure 5 As shown, this embodiment provides a real-time small target detection system for UAVs in foggy environments, including:

[0108] The image acquisition module is used to acquire aerial images at different distances and with different fog densities and to preprocess them;

[0109] In this embodiment, aerial images taken at different distances and with different fog densities are first obtained using a visible light camera device mounted on a drone.

[0110] The model building module is used to build and train the AERO-YOLO object detection model, which consists of a backbone feature extraction network, an image adaptive feature enhancement module, and a feature fusion module.

[0111] In this embodiment, an image adaptive feature enhancement module is embedded between the backbone feature extraction network and the feature fusion module to suppress fog interference and enhance the feature representation of small targets. Through this module, the features of small targets and high-contrast regions are amplified, while the noise response in foggy regions is suppressed, improving the model's detection performance in foggy environments. Specifically, the image adaptive feature enhancement module employs a feature enhancement mechanism based on the fusion of channel attention and spatial attention to suppress fog noise interference and amplify the edge and texture features of small targets, thus improving the detection performance of small targets in foggy conditions. The fusion module is used to achieve multi-scale feature interaction and detection.

[0112] The image adaptive feature enhancement module includes a channel global statistical extraction unit, an attention weight generation unit, and a feature saliency enhancement unit; wherein:

[0113] The aforementioned channel global statistical extraction unit is used to perform global average pooling on the input multi-scale feature map to extract the saliency distribution of fog noise, thereby obtaining the sensitivity of different channels to small target information.

[0114] ; (3)

[0115] In the formula, The multi-scale feature map output by the backbone feature extraction network; This indicates a global average pooling operation; This is a channel statistics vector;

[0116] The attention weight generation unit generates channel attention weights based on channel statistical vectors through a two-layer mapping network to distinguish between fog noise regions and effective target regions, thereby achieving feature selective enhancement. Its calculation formula is as follows:

[0117] ; (4)

[0118] in, For trainable weights, It is the ReLU activation function. It is the Sigmoid activation function. The generated channel attention weights.

[0119] The feature saliency enhancement unit is used to broadcast the channel attention weights and apply them to the multi-scale feature map. The channel enhancement feature map is obtained. Channels dominated by fog interference are suppressed, while channels containing small target features are enhanced to improve the feature representation capability of small target detection. The calculation formula is as follows:

[0120] (5)

[0121] Where ⊙ represents a channel-by-channel multiplication operation;

[0122] In channel enhancement feature map Above, average pooling and max pooling are performed along the channel dimension respectively to obtain two single-channel spatial feature maps. and The data is then stitched together along the channel dimension and then passed through a convolutional layer. Spatial attention map generated with Sigmoid activation function ,Right now:

[0123] ; (6)

[0124] The spatial attention map assigns lower weights to large, low-information fog background regions and higher weights to salient regions containing small targets.

[0125] Spatial attention map Broadcast along the channel dimension and then applied to the channel-enhanced feature map. And through residual methods and multi-scale feature maps Add them together to obtain the final enhanced features. ,Right now:

[0126] ; (7)

[0127] Where ⊙ represents element-wise multiplication;

[0128] The final enhancement feature While suppressing fog noise interference, it significantly improves the contrast and distinguishability of small targets in foggy weather on the feature map, and sends them as input to the feature fusion module for detection.

[0129] The model training module trains the AERO-YOLO object detection model. During the training process, the image adaptive feature enhancement module participates in forward and backward propagation together with the AERO-YOLO object detection model.

[0130] During the inference phase, the image adaptive feature enhancement module applies a joint weighting of channel attention and spatial attention to the input feature map using fixed parameters obtained from training. The fused feature enhancement mechanism suppresses fog noise interference and amplifies the edge and texture features of small targets.

[0131] The model optimization module prunes and lightweights the trained AERO-YOLO object detection model, and then deploys the optimized and lightweight AERO-YOLO object detection model to the UAV edge computing platform.

[0132] In this embodiment, the model pruning and lightweight optimization unit includes a channel importance evaluation unit, a redundant channel removal unit, and a convolution and normalization fusion unit; wherein:

[0133] The channel importance evaluation unit calculates the contribution of each convolutional channel to the object detection task using a scaling factor based on the Batch Normalization layer, i.e.:

[0134] ; (8)

[0135] In the formula, For the first Key metrics for each channel; This refers to the scaling parameters for the corresponding Batch Normalization layer.

[0136] The redundant channel removal unit prunes channels according to a preset pruning ratio based on channel importance indicators. The channels are sorted and low-contribution channels are removed, with the following pruning conditions:

[0137] ; (9)

[0138] in, The threshold is dynamically determined based on the pruning ratio; below the threshold... The channels are marked as redundant channels.

[0139] The convolution and normalization fusion unit fuses the convolutional layer and the batch normalization layer to reduce inference computation overhead. The fusion computation method is as follows:

[0140] ; (10)

[0141] ; (11)

[0142] in, The weights and biases of the merged convolutional layers, These are the original convolution weights. For the mean and variance; Scaling and translation parameters for Batch Normalization It is the stability constant.

[0143] The real-time detection module calls the AERO-YOLO target detection model in foggy remote sensing environments to detect the target's category, location, and confidence level in real time, and outputs the detection box to the UAV flight control system for target localization and path correction.

[0144] The embodiments and descriptions above are merely illustrative of the principles and preferred embodiments of the present invention. Various changes and modifications may be made to the present invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed.

Claims

1. A real-time detection method for small targets by unmanned aerial vehicles (UAVs) in foggy environments, characterized in that, Includes the following steps: S1) Acquire aerial images at different distances and fog densities and preprocess them; S2) Construct an AERO-YOLO target detection model consisting of a backbone feature extraction network, an image adaptive feature enhancement module, and a feature fusion module; S3) Train the AERO-YOLO object detection model. During the training process, the image adaptive feature enhancement module and the AERO-YOLO object detection model participate in forward propagation and backward propagation together. During the inference phase, the image adaptive feature enhancement module applies a joint weighting of channel attention and spatial attention to the input feature map using fixed parameters obtained from training. The fused feature enhancement mechanism suppresses fog noise interference and amplifies the edge and texture features of small targets. S4) The trained AERO-YOLO object detection model is pruned and lightweighted using the model pruning and lightweighting optimization unit; S5) Deploy the optimized lightweight AERO-YOLO target detection model to the UAV edge computing platform to detect the target category, location and confidence level in real time in foggy remote sensing environment, and output the detection box to the UAV flight control system for target localization and path correction.

2. The method for real-time detection of small targets by UAVs in foggy environments according to claim 1, characterized in that: In step S2), the image adaptive feature enhancement module is embedded between the backbone feature extraction network and the feature fusion module. The backbone feature extraction network is used to extract multi-scale feature maps of aerial images. The image adaptive feature enhancement module suppresses fog noise interference and amplifies the edge and texture features of small targets by adopting a feature enhancement mechanism based on the fusion of channel attention and spatial attention, thereby improving the detection performance of small targets in foggy weather. The fusion module is used to realize multi-scale feature interaction and detection.

3. The method for real-time detection of small targets by UAVs in foggy environments according to claim 2, characterized in that: The image adaptive feature enhancement module includes a channel global statistical extraction unit, an attention weight generation unit, and a feature saliency enhancement unit.

4. The method for real-time detection of small targets by UAVs in foggy environments according to claim 3, characterized in that: In step S2), the channel global statistical extraction unit is used to perform global average pooling on the input multi-scale feature map to extract the saliency distribution of fog noise, thereby obtaining the sensitivity of different channels to small target information; that is: ； （3） In the formula, Multi-scale features are output by the backbone feature extraction network; This indicates a global average pooling operation; This is the channel statistics vector.

5. A method for real-time detection of small targets by unmanned aerial vehicles (UAVs) in foggy environments according to claim 4, characterized in that: In step S2), the attention weight generation unit generates channel attention weights based on channel statistical vectors through a two-layer mapping network to distinguish between fog noise regions and effective target regions, thereby achieving feature selective enhancement. The calculation formula is as follows: ； （4） in, These are the trainable weights. It is the ReLU activation function. It is the Sigmoid activation function. The generated channel attention weights.

6. A method for real-time detection of small targets by unmanned aerial vehicles (UAVs) in foggy environments according to claim 5, characterized in that: In step S2), the feature saliency enhancement unit is used to broadcast the channel attention weights and apply them to multi-scale features. The channel enhancement feature map is obtained. : ； （5） Where ⊙ represents a channel-by-channel multiplication operation; In channel enhancement feature map The above methods perform average pooling and max pooling along the channel dimension to obtain two single-channel spatial feature maps. and The data is then stitched together along the channel dimension and then passed through a convolutional layer. Spatial attention map generated with Sigmoid activation function ,Right now: ； （6） Spatial attention map Broadcast along the channel dimension and then applied to the channel-enhanced feature map. And through residual methods and multi-scale feature maps Add them together to obtain the final enhanced features. ,Right now: ； （7） The final enhancement feature It is fed into the feature fusion module for detection.

7. A method for real-time detection of small targets by unmanned aerial vehicles (UAVs) in foggy environments according to claim 1, characterized in that: In step S4), the model pruning and lightweight optimization unit includes a channel importance evaluation unit, a redundant channel screening unit, and a convolution and normalization fusion unit.

8. A method for real-time detection of small targets by unmanned aerial vehicles (UAVs) in foggy environments according to claim 7, characterized in that: In step S4), the channel importance evaluation unit calculates the contribution of each convolutional channel to the target detection task using the scaling factor based on the Batch Normalization layer, i.e.: ； （8） In the formula, For the first Key metrics for each channel; These are the scaling parameters for the corresponding Batch Normalization layer; The redundant channel removal unit prunes channels according to a preset pruning ratio based on channel importance indicators. The channels are sorted and low-contribution channels are removed, with the following pruning conditions: ； （9） in, The threshold is dynamically determined based on the pruning ratio; below the threshold... The channels are marked as redundant channels.

9. A method for real-time detection of small targets by unmanned aerial vehicles (UAVs) in foggy environments according to claim 8, characterized in that: In step S4), the convolution and normalization fusion unit performs structural fusion of the convolutional layer and the Batch Normalization layer. The fusion calculation method is as follows: ； （10） ； （11） in, These represent the weights and biases of the fused convolutional layer. These are the original convolution weights. These are the mean and variance, respectively. These are the scaling and translation parameters for Batch Normalization; It is the stability constant.

10. A real-time detection system for small targets on unmanned aerial vehicles (UAVs) in foggy environments, characterized in that, include: The image acquisition module is used to acquire aerial images at different distances and with different fog densities and to preprocess them; The model building module is used to build and train the AERO-YOLO object detection model, which consists of a backbone feature extraction network, an image adaptive feature enhancement module, and a feature fusion module. The model training module trains the AERO-YOLO object detection model. During the training process, the image adaptive feature enhancement module participates in forward and backward propagation together with the AERO-YOLO object detection model. During the inference phase, the image adaptive feature enhancement module applies a joint weighting of channel attention and spatial attention to the input feature map using fixed parameters obtained from training. The fused feature enhancement mechanism suppresses fog noise interference and amplifies the edge and texture features of small targets. The model optimization module prunes and lightweights the trained AERO-YOLO object detection model, and then deploys the optimized and lightweight AERO-YOLO object detection model to the UAV edge computing platform. The real-time detection module calls the AERO-YOLO target detection model in foggy remote sensing environments to detect the target's category, location, and confidence level in real time, and outputs the detection box to the UAV flight control system for target localization and path correction.

Images