A physical guided unmanned aerial vehicle fog image target detection method and system

Abstract

The application discloses a kind of physical guiding unmanned aerial vehicle fog image target detection method and system, belong to computer vision and unmanned aerial vehicle remote sensing technical field.The method includes: the image is input to physical guiding target detection model, edge enhancement network in model is extracted by local high-frequency residual and adaptively weighted amplification, with the boundary recovery under the scale of multiple receptive fields;Attention reprogrammable normalization Transformer module is serialized after feature map is encoded with position, globally modeled by multi-head self-attention, and then reprogrammable normalization is carried out by calculating mean and variance of local window;Multi-branch structure is used in training when reparameterization multi-head detection head, and multiple branches are equivalent to single-layer convolution in inference to complete classification and regression;Output target class and position.The application also constructs the detection system for realizing the above method, including image acquisition module, edge enhancement module, statistical heterogeneity elimination module, efficient detection module and output module.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and UAV remote sensing technology, and in particular to a physically guided UAV target detection method and system in foggy weather images. Background Technology

[0002] Unmanned aerial vehicles (UAVs) have revolutionized remote sensing applications with their ability to capture high-resolution imagery from unique perspectives, significantly enhancing critical tasks such as vehicle detection and agricultural plant protection. Despite these remarkable capabilities, the reliability of UAV-based remote sensing systems remains severely affected by adverse weather conditions, particularly fog, which introduces complex atmospheric scattering effects, fundamentally weakening the visual features necessary for accurate object detection.

[0003] Given the increasing deployment of drones in critical applications including search and rescue, traffic monitoring, and infrastructure inspection, reliable detection under adverse weather conditions is crucial for operational safety and mission success.

[0004] From a physics perspective, fog poses a significant challenge to aerial remote sensing tasks through its interaction with electromagnetic radiation. When light passes through fog particles, the Mie scattering process causes non-uniform attenuation of high-frequency spatial information. This scattering acts as a natural low-pass filter, severely damaging edge information crucial for accurate target localization. Even more serious is the highly non-uniform spatial distribution of fog in aerial scenarios, leading to extreme differences in visibility conditions within a single image. These characteristics present two fundamental challenges: (1) systematic attenuation of edge information, resulting in blurred target boundaries; and (2) spatially heterogeneous feature degradation within the image domain. Furthermore, UAV platforms place additional demands on computational efficiency, requiring detection algorithms to maintain high detection accuracy even under strict computational resource constraints.

[0005] These physical degradation effects significantly impact the performance of existing object detection frameworks. Mainstream object detection methods widely used in UAV image analysis exhibit significant limitations when faced with fog-induced image degradation. Classic CNN-based detectors (such as Cascade R-CNN and the YOLO series) rely on convolutional operations, which presuppose that the feature quality of the entire image remains consistent. This assumption makes them highly susceptible to spatial degradation caused by non-uniform fog distribution. Although recent Transformer-based detectors (such as Deformable DETR and Conditional DETR) possess stronger global context modeling capabilities, they still struggle to cope with the inherent systematic low-pass filtering effects of atmospheric scattering. Even methods specifically designed for foggy scenes typically treat fog as uniform additive noise rather than addressing its physical nature and spatial heterogeneity, resulting in poor object detection performance.

[0006] Dehazing images before target detection, while seemingly an intuitive preprocessing approach, faces significant challenges in drone scenarios. First, this sequential processing mode incurs substantial computational overhead, as each network must independently complete a full forward inference, making real-time processing difficult on resource-constrained drone platforms. Second, general dehazing algorithms primarily aim to improve visual quality rather than preserve features crucial for downstream detection tasks, potentially losing information essential for accurate target identification. Third, most existing dehazing methods are designed for ground-based shooting perspectives and cannot adequately adapt to the unique characteristics of aerial images, including the extremely wide viewing angles, dramatic scale variations, and distance-dependent fog effects characteristic of drone data. Summary of the Invention

[0007] This invention provides a physically guided UAV target detection method and system in foggy weather to solve three core problems in the prior art: edge information attenuation caused by atmospheric scattering, statistical heterogeneity caused by non-uniform fog, and limited computing resources.

[0008] To solve the above-mentioned technical problems, the technical solution proposed by this invention is as follows:

[0009] A physically guided method for target detection in foggy images using a drone includes the following steps:

[0010] Step S1: Acquire aerial images of the target area taken by the drone in foggy weather.

[0011] Step S2: Input the aerial image of the foggy day to be detected into a pre-built physically guided target detection model, the model including:

[0012] An edge enhancement network is used to perform edge enhancement operations on an input feature map based on an atmospheric scattering physics model. The edge enhancement operations include: adaptive high-frequency amplification of the feature map through an edge enhancer, and recovery of boundary information at multiple spatial scales through a boundary adaptive recovery module.

[0013] The attention-based reprogrammable normalization Transformer module is used to perform global context modeling and local statistical normalization on the edge-enhanced feature map, eliminating the statistical heterogeneity caused by non-uniform fog distribution.

[0014] The reparameterized multi-head detection head is used to perform object classification and bounding box regression on the normalized feature map and output the detection results.

[0015] Step S3: Output the target category and location information in the drone aerial image.

[0016] A further improvement to the above technical solution is as follows:

[0017] Preferably, the edge enhancer process specifically involves: subtracting the feature map obtained by 3×3 average pooling with a stride of 1 and padding of 1 from the input feature map to obtain local detail difference; passing the local detail difference through a convolutional layer with a Sigmoid activation function to generate adaptive weights; and finally adding the input feature map and the adaptive weights pixel by pixel to obtain the edge-enhanced feature map.

[0018] Preferably, the processing procedure of the boundary adaptive restoration module is as follows: the input feature map is processed at multiple different receptive field scales. At each scale, adaptive average pooling downsampling, 1×1 convolution dimensionality reduction, and grouped 3×3 convolution feature extraction are performed sequentially. Then, the original spatial size is upsampled back by bilinear interpolation and fed into the corresponding edge enhancer. At the same time, local features are extracted by an independent 3×3 convolution. Finally, the edge enhancement features and local features obtained from all scales are concatenated in the channel dimension, and the multi-scale fused boundary restoration features are output through a final convolutional layer.

[0019] Preferably, the processing steps of the attention-based reprogrammable normalized Transformer module are as follows: First, the input feature map is simultaneously flattened and rearranged in dimensions to generate a two-dimensional positional encoding; then, the flattened features are added to the positional encoding and fed into the Transformer encoder, which captures global feature dependencies through a multi-head self-attention mechanism; finally, a reprogrammable normalized layer is applied to the output of the Transformer encoder. The reprogrammable normalized layer calculates the mean and variance in the local spatial region and performs an affine transformation on the normalized features using learnable scaling and offset parameters to output a feature map after statistical heterogeneity elimination.

[0020] Preferably, in the training phase, the multi-branch module of the reparameterized multi-head detector includes the following four branches in parallel: a standard 3×3 convolution branch, sequentially connected 1×1 and 3×3 convolution branches, sequentially connected 1×1 and dilated 3×3 convolution branches, and an identity mapping branch; in the inference phase, the weights and biases of all branches of the multi-branch module are equivalent to the weights and biases of a single-layer convolution through mathematical fusion, so that feature processing is completed in a single convolution operation.

[0021] The present invention also provides a physically guided UAV target detection system in foggy weather images, the system comprising:

[0022] The image acquisition module is used to acquire aerial images of the target drone taken in foggy conditions.

[0023] The edge enhancement module is used to perform edge enhancement operations on the input feature map based on an atmospheric scattering physics model;

[0024] The statistical heterogeneity elimination module is used to perform global context modeling and local statistical normalization on the edge-enhanced feature map;

[0025] The high-efficiency detection module performs object classification and bounding box regression on the normalized feature map and outputs the detection results;

[0026] The output module is used to output the target category and location information in the drone aerial images.

[0027] Preferably, the edge enhancement module is embedded in the backbone network and neck network of the target detection model; the feature maps output from multiple stages of the backbone network are processed by the edge enhancement module and then sent to the neck network, where they are further enhanced by the edge enhancement module at multiple pyramid levels of the neck network.

[0028] Preferably, the edge enhancement module includes an edge enhancer unit and a boundary adaptive restoration unit. The edge enhancer unit extracts local details and adaptively weights and amplifies them. The boundary adaptive restoration unit processes feature maps in parallel at multiple receptive field scales and restores the boundary information of targets at different depths through bilinear interpolation and feature fusion.

[0029] Preferably, the statistical heterogeneity elimination module includes: a location encoding generation unit for adding two-dimensional spatial location encoding to the input feature map; a Transformer encoder unit, comprising a multi-head self-attention layer and a feedforward network layer, with each sub-layer followed by a residual connection; and a reprogrammable normalization unit for calculating the mean and variance of the features within a local window and performing an adaptive affine transformation.

[0030] Preferably, the high-efficiency detection module includes a multi-branch structure during the training phase, which includes a standard 3×3 convolution path, a 1×1 convolution cascaded 3×3 convolution path, a 1×1 convolution cascaded dilated 3×3 convolution path, and an identity mapping path; during the inference phase, the multi-branch structure is equivalently fused into a single standard 3×3 convolution layer, thereby reducing computational overhead.

[0031] The physically guided UAV target detection method and system in foggy weather provided by this invention has the following advantages compared with the prior art:

[0032] (1) The physical-guided UAV fog image target detection method and system of the present invention is an object detection framework based on physical laws. This framework systematically addresses the meteorological scattering problem faced in fog images taken by UAVs by clearly modeling the physical characteristics of fog, rather than treating it as uniform noise. The present invention directly addresses the fundamental physical causes, namely the attenuation of edge information and the statistical heterogeneity caused by the uneven spatial distribution of fog, and its detection performance is significantly better than that of traditional methods.

[0033] (2) The physical-guided UAV fog image target detection method and system of the present invention employs three collaborative components to target specific aspects of atmospheric attenuation while maintaining computational efficiency. The EENet edge enhancement network is proposed, which suppresses multiplicative attenuation through adaptive detail enhancement, providing a physically-guided solution for the loss of edge information in aerial images; an ATRN module with reprogrammable normalization is constructed to cope with the statistical heterogeneity caused by changes in fog concentration, and robust feature processing can be achieved under different visibility conditions; the RPM-Head reparameterized detection head is introduced, which reduces the computational load by 40.9% while maintaining detection accuracy, proving its feasibility for deployment on resource-constrained platforms. Attached Figure Description

[0034] Figure 1 This is a diagram of the edge enhancement module in this invention.

[0035] Figure 2 This is a diagram of the boundary adaptive recovery module in this invention.

[0036] Figure 3 This is the structure of the reprogrammable normalized attention Transformer of this invention.

[0037] Figure 4 This invention relates to a reparameterized multi-head detection head structure.

[0038] Figure 5 This is the structure of the multi-branch module of the present invention.

[0039] Figure 6 This is the original fog map from the training set used in the experimental verification.

[0040] Figure 7 To verify the distribution of target categories in the dataset in the experiment.

[0041] Figure 8 To verify the TIDE error analysis in the experiment. Detailed Implementation

[0042] The following provides a detailed description of specific embodiments of the present invention. It should be understood that the specific embodiments described herein are for illustrative and explanatory purposes only and are not intended to limit the scope of the invention.

[0043] This invention presents a physics-guided UAV target detection method and system for foggy images, employing a UAV-FogDet detection framework. This framework, guided by physical mechanisms, directly addresses the root causes of atmospheric degradation while maintaining computational efficiency for UAV deployment. Based on an understanding of atmospheric scattering physics, the method integrates three collaborative modules to fundamentally solve the detection failure problem caused by fog, rather than merely addressing surface phenomena. First, the Edge Enhancement Network (EENet) directly counteracts the low-pass filtering effect caused by atmospheric scattering through adaptive high-frequency amplification, systematically recovering boundary information lost due to Mie scattering. EENet operates at multiple spatial scales to adapt to the distance-dependent characteristics of atmospheric degradation, ensuring robust edge recovery regardless of target depth in aviation scenarios. Second, the Transformer (ATRN) utilizes Transformer-based global context modeling combined with adaptive normalization to handle statistical heterogeneity caused by spatial variations in fog concentration. It calculates local correlation statistics rather than relying on global assumptions that fail under non-uniform atmospheric conditions. Third, the reparameterized multi-head detector (RPM-Head) adopts a paradigm of "complex training and simple inference," enabling it to be practically deployed on resource-constrained UAV platforms. While maintaining robustness to variable atmospheric conditions, it significantly reduces computational overhead through innovative structural reparameterization.

[0044] The physically guided UAV target detection method in foggy weather images of the present invention specifically includes the following steps:

[0045] Step S1: Acquire aerial images of the target foggy weather captured by the drone in foggy conditions.

[0046] During flight, drones use their onboard cameras to continuously capture images of the ground scene from a top-down or oblique perspective. Due to the presence of scattering media such as fog and haze in the atmosphere, the captured images may exhibit degradation phenomena such as decreased contrast, blurred edges, and color distortion.

[0047] Step S2: Input the aerial image of the foggy day to be detected into the pre-built physical-guided target detection model for feature processing and target detection.

[0048] The model employs a backbone network—neck network—detector head architecture. The backbone network integrates an Edge Enhancement Network (EENet) and an Attention-Based Reprogrammable Normalized Transformer Module (ATRN): EENet recovers high-frequency spatial details significantly attenuated by atmospheric scattering low-pass filtering, while ATRN handles the statistical heterogeneity introduced by spatial variations in fog concentration through adaptive normalization. Features output from different stages of the backbone network are processed by the neck network, which continues to use EENet for further feature enhancement at multiple pyramid levels (P3, P4, P5). The detector head utilizes a reparameterized multi-head detector (RPM-Head). Through structural reparameterization, it significantly reduces computational overhead while maintaining detection robustness, ensuring the practical deployment feasibility of the framework.

[0049] Specifically, it includes the following:

[0050] S2-1, Preliminary feature extraction and edge enhancement of the backbone network.

[0051] First, the input image is fed into the backbone network. The backbone network employs a multi-stage convolutional structure, with each stage outputting feature maps at different spatial resolutions, denoted as C3, C4, and C5. An edge enhancement network, EENet, is then connected after the feature maps output from each stage.

[0052] S2-1-1, Design an edge enhancement network.

[0053] Atmospheric scattering in foggy weather follows a well-established atmospheric scattering model, expressed as:

[0054] (1)

[0055] in, Images of foggy weather observed. For scene radiation, Atmospheric light, Represents distance-dependent transmission mapping. These are the pixel coordinates of the image. ,in It is a natural constant. Atmospheric scattering coefficient, This represents the scene depth (or the distance from a scene point to the camera). Multiplication terms. This can lead to a systematic degradation of the original scene content, and because Mie scattering is wavelength-dependent, high-frequency components (such as edges and fine details) are particularly affected.

[0056] EENet directly counteracts multiplicative attenuation through adaptive high-frequency enhancement, systematically recovering boundary information lost due to atmospheric scattering. Unlike existing image enhancement methods that employ uniform processing, this invention uses a learnable filter to adapt to local atmospheric conditions, selectively enhancing degraded spatial frequencies while preserving intact information.

[0057] like Figure 1 As shown, the core of EENet is the EdgeEnhancer module, which specifically addresses the contrast reduction and detail loss caused by multiplicative terms in the atmospheric scattering model. It is represented as:

[0058] (2)

[0059] in, This represents the output feature map after the edge enhancement operation. Indicates the input feature map; For convolution operations with sigmoid activation function, This represents the Sigmoid activation function; It is a 3×3 average pooling operation with a step size of 1 and a padding size of 1. Used to extract local detail information, it can accurately characterize atmospheric models. This results in the attenuation of high-frequency spatial components. Through sigmoid-activated convolution operations, the system can adaptively weight these details according to the degree of local degradation, thereby achieving selective enhancement that balances detail recovery and noise suppression.

[0060] S2-1-2, Design boundary adaptive recovery module.

[0061] like Figure 2 As shown, considering atmospheric effects To address the distance-dependent characteristics, this invention introduces a boundary adaptive recovery module (BARM). Objects at different depths experience varying degrees of degradation, thus requiring processing at multiple spatial scales to handle the heterogeneous transmission values ​​in drone-captured scenes.

[0062] BARM processes edge information across multiple receptive fields, with receptive field sizes ranging from [3, 6, 9, 12] (the output size of adaptive average pooling), to capture degradation patterns at multiple scales from fine to coarse.

[0063] (3)

[0064] (4)

[0065] in, This represents the edge enhancement feature at the i-th scale. This represents the edge enhancer at the i-th scale. This indicates that the feature at scale i is subjected to bilinear upsampling. This represents the feature extraction result at the i-th scale. This indicates the output of the boundary adaptive recovery module. This indicates the final fused convolution. This indicates concatenation along the channel dimension. This represents local 3×3 convolutional features. Represents the set of edge enhancement features across all scales. Indicates the total number of scales. Indicates scale index.

[0066] Local features are processed using a standard 3×3 convolution operation. Features are extracted at scale i, while Interp uses bilinear interpolation to restore the feature map to its original spatial dimensions. The feature extraction function for a specific scale is defined as follows:

[0067] (5)

[0068] in, This represents a 3×3 grouped convolution with g groups. Indicates the number of groups. This represents a 1×1 dimension-reduced convolution. This represents the adaptive average pooling at the i-th scale.

[0069] This invention’s multi-scale approach ensures robust boundary recovery in aerial remote sensing scenarios, even when faced with complex and variable atmospheric densities. It also directly addresses the core issue of the inherent distance dependence of atmospheric degradation, which is often overlooked by traditional enhancement methods.

[0070] S2-2, Attention-based reprogrammable normalized Transformer module (ATRN).

[0071] The spatiotemporal distribution of fog in aerial images varies depending on the scene depth at different aerial viewpoints. Inherent differences exist, resulting in an uneven degradation pattern. Due to... Objects at different distances experience different atmospheric propagation path lengths, leading to spatial differences in transmission coefficients. This depth-dependent variation means that features of nearby objects retain higher fidelity, while features of distant objects suffer greater attenuation, resulting in statistical heterogeneity, which contradicts the assumptions of traditional normalization methods.

[0072] ATRN addresses this statistical heterogeneity through a Transformer-based architecture, incorporating a reprogrammable normalization mechanism (e.g., Equation (7)). RepNorm(x) is a reprogrammable normalized mathematical expression to accommodate spatially varying degradation characteristics. Unlike traditional batch normalization, which assumes consistent feature distributions, the method of this invention can calculate locally correlated statistics, thus fully considering the differences in atmospheric pollution levels between different spatial regions.

[0073] The architecture of this invention processes input features through a designed pathway. Let the input feature be... ,in For batch size, For the number of channels, and These represent the height and width of the feature map, respectively. ATRN includes two parallel processing paths: the first path flattens and rearranges the input features into... Used for Transformer encoding; the second path generates two-dimensional positional encoding. This is used to preserve spatial relationships. The output feature map of the backbone network after EENet edge enhancement has the shape [B, C, H, W]. It can retain spatial relationship information even under degraded conditions such as fog. Figure 3 As shown, the input features are processed through two parallel branches: one for flattening / rearranging the Transformer computation, and the other for two-dimensional position encoding to maintain spatial awareness. The Transformer encoder achieves its global receptive field through a self-attention mechanism.

[0074] (6)

[0075] in, This indicates the output of multi-head self-attention. Represents the query matrix. Represents the key matrix. Represents a value matrix, Represents the normalization function. This represents the attention head dimension. This indicates the matrix transpose.

[0076] Using global context modeling allows each spatial location to gather information from the entire feature map, which is crucial for dealing with non-uniform degradation—in severely degraded regions, the information preserved by clear regions can actually be beneficial.

[0077] The reprogrammable normalization mechanism effectively solves the statistical heterogeneity problem caused by changes in spatial transport values. This reprogrammable normalization mechanism calculates the mean and standard deviation in local regions and uses learnable parameters to perform affine transformations, so that the feature distributions of different spatial regions are adaptively aligned, thereby reducing the statistical heterogeneity caused by non-uniform fog distribution and improving the robustness of classification and localization.

[0078] (7)

[0079] in, This represents the reprogrammable normalized output feature. and Indicates learnable parameters, Indicates input features, This represents the mean calculated within a local window. This represents the standard deviation calculated for a local window. and It can adaptively adjust based on local features; and This indicates a locally calculated statistic, rather than a global batch statistic.

[0080] Different spatial regions have different levels of atmospheric pollution due to varying scene depths, thus enabling robust feature processing under variable visibility conditions.

[0081] S2-3, multi-scale feature fusion and further enhancement of the neck network.

[0082] The C3, C4, and C5 feature maps output from each stage of the backbone network, as well as the feature maps output from the ATRN module, are fed into the neck network. The neck network employs a feature pyramid structure, achieving multi-scale feature fusion through upsampling, downsampling, and lateral connections. After each pyramid output level (denoted as P3, P4, and P5) of the neck network, an EENet module is embedded again to perform a second edge enhancement on the fused features, further sharpening the target boundaries and improving the detection consistency between small and large targets.

[0083] S2-4, Reparameterize the output of the multi-head detection head to obtain the detection results.

[0084] While EEENet and ATRN address the physical challenges posed by atmospheric scattering, practical UAV deployments introduce crucial computational constraints. The variability in feature quality within foggy sky imagery suggests that sophisticated adaptive processing techniques can contribute to robust detection performance. However, the inherent resource limitations of UAV platforms restrict algorithm complexity, creating a fundamental conflict between detection robustness and computational efficiency.

[0085] RPM-Head resolves this contradiction by employing a paradigm that is complex to train but simple to reason, which maintains robustness in detection under different atmospheric conditions while also achieving computational efficiency suitable for UAV deployment.

[0086] like Figure 4 and Figure 5 As shown, the architecture of this invention consists of three main functional modules: (1) initial GNCov group normalized convolution operation, used to normalize features from different pyramid levels (P3, P4, P5); (2) shared multi-branch module (MBM) layer, which processes features through multiple different computational paths; and (3) dedicated box regression and classification branches, which generate the final detection output with optimized computational efficiency.

[0087] The multi-branch module effectively addresses the challenge of maintaining robust detection performance while controlling computational costs:

[0088] (8)

[0089] in, Indicates the output of a multi-branch module. This represents the weight of the i-th branch. This represents the calculation result of the i-th branch.

[0090] The multi-branch module of this invention comprises four complementary branches, each responsible for addressing different aspects of atmospheric degradation: (1) a standard 3×3 convolution for basic processing; (2) a sequence of 1×1 convolutions followed by 3×3 convolutions for efficient feature transformation; (3) a sequence of 1×1 convolutions followed by dilated 3×3 convolutions to capture contextual information at different atmospheric scales; and (4) an identity mapping to preserve original information. This multi-branch design helps the network learn robust representations during the training phase, thereby effectively addressing various image quality changes introduced by atmospheric effects.

[0091] Mathematically merge the branches of multiple training phases into a single inference phase convolution:

[0092] (9)

[0093] (10)

[0094] in, and These represent the equivalent weights and bias parameters of the convolution during a single inference iteration, respectively. (Function) A mathematical approach is used to fuse consecutive convolution operations. This indicates that the identity mapping is converted into a compatible convolutional form. This represents the standard 3×3 convolution weights. This represents the 1×1 convolution weight of the first branch. This represents the 3×3 convolution weights of the first branch. This represents the 1×1 convolution weight of the second branch. This represents the dilated 3×3 convolution weights. Indicates the hole convolution dilation rate. This represents the standard 3×3 convolution bias. This indicates the bias of the 3×3 convolution in the first branch. This indicates a dilated 3×3 convolution bias. The reparameterization technique in this step ensures that the detector head can operate with the computational efficiency of a single convolution during deployment, while maintaining the robustness learned during multi-branch training, thus directly addressing the resource constraints that limit the use of traditional methods in UAV applications.

[0095] Step S3: Output the detection results.

[0096] The RPM-Head output is post-processed to obtain the final detection result, which is then output.

[0097] This invention also provides a physically guided UAV target detection system for foggy images, which implements the above-described method. The system comprises the following five modules:

[0098] (1) Image acquisition module: responsible for acquiring aerial images in foggy conditions in real time during the flight of the UAV, or reading pre-acquired foggy images from the onboard storage device, and converting the images into a uniform size for subsequent modules to process.

[0099] (2) Edge enhancement module (EENet): Embedded in the backbone network and neck network, it is used to perform edge enhancement operation based on the atmospheric scattering physical model on the input feature map, to counteract the low-pass filtering effect of fog and restore the high-frequency boundary information lost due to Mie scattering.

[0100] The edge enhancement module consists of two sub-units: the edge enhancer unit and the boundary adaptive recovery unit.

[0101] Edge enhancer unit: Enhances edge and texture information in feature maps through local detail extraction and adaptive weighted amplification.

[0102] Boundary Adaptive Recovery Unit: Processes feature maps in parallel at multiple receptive field scales (3, 6, 9, 12) and recovers boundary information of targets at different depths through bilinear interpolation and feature fusion.

[0103] The adaptive boundary restoration unit comprises a local feature extraction branch and multiple parallel multi-scale processing branches. In the multi-scale branches, the input feature map is first downsampled to different sizes (e.g., 3×3, 6×6, 9×9, 12×12) using adaptive average pooling, then sequentially subjected to 1×1 convolutional dimensionality reduction and grouped 3×3 convolutional feature extraction. The processed feature maps are then upsampled back to their original size using bilinear interpolation and fed into independent edge enhancers for high-frequency enhancement. Simultaneously, a separate 3×3 convolutional path is responsible for extracting local features at the original scale. Finally, the outputs of all branches are concatenated along the channel dimension and fused through a final convolutional layer to output a feature map that restores the multi-scale boundary information.

[0104] (3) Statistical heterogeneity elimination module: global context modeling and local statistical normalization are performed on the deep feature map after edge enhancement to eliminate the spatial statistical differences caused by non-uniform fog distribution, so that the fog degradation feature distribution in different regions tends to be consistent, thereby improving the robustness of classification and localization.

[0105] The statistical heterogeneity elimination module first receives an input feature map of size B×C×H×W (batch size × number of channels × height × width). This feature map is processed in two ways: one way involves flattening and permutation to transform it into a sequence format acceptable to the Transformer encoder; the other way generates a 2D positional encoding to preserve spatial structure information. The two are then added together and fed into a Transformer encoder composed of stacked multi-head self-attention and feedforward networks (FFNs), with residual connections and normalization layers after each sub-layer. After the encoder output is reshaped to its original spatial dimensions, it undergoes an adaptive affine transformation through a reprogrammable normalization layer, outputting a feature map of the same size as the input that has eliminated statistical heterogeneity.

[0106] The statistical heterogeneity elimination module includes a position encoding generation unit, a Transformer encoder unit, and a reprogrammable normalization unit.

[0107] The location encoding generation unit adds a two-dimensional spatial location code to the input feature map while preserving spatial structure information.

[0108] The Transformer encoder unit contains a multi-head self-attention layer and a feedforward network layer. Each sub-layer is followed by residual connections and normalization to capture global dependencies.

[0109] The reprogrammable normalization unit calculates the mean and variance of features within a local window and performs affine transformations using learnable scaling and offset factors, replacing traditional batch normalization.

[0110] (4) High-efficiency detection module: Performs target classification and bounding box regression on the normalized multi-scale feature maps and outputs the detection results. This module adopts a design that is complex to train but simple to infer. In the training stage, a multi-branch structure is used to enhance robustness, and in the inference stage, the multi-branch structure is fused into a single-layer convolution, which greatly reduces the computational overhead and is suitable for real-time deployment on UAV platforms.

[0111] The high-efficiency detection module includes a multi-branch training unit, a reparameterized fusion unit, and classification and regression sub-networks.

[0112] The multi-branch training unit runs four branches in parallel during training: standard convolution, serial convolution, dilated convolution, and identity mapping.

[0113] The reparameterized fusion unit effectively merges the weights and biases of multiple branches into a single-layer convolution during inference.

[0114] The classification and regression sub-networks output the target category probability and bounding box coordinates, respectively.

[0115] like Figure 4 and Figure 5 As shown, the efficient detection module receives feature maps from layers P3, P4, and P5 of the neck network and first normalizes them through group-normalized convolutional layers. Then, the feature maps enter a shared multi-branch module. This module contains four parallel branches during the training phase: the first branch is a standard 3×3 convolution for basic feature processing; the second branch is a serial structure of 1×1 convolution followed by 3×3 convolutions for efficient feature transformation; the third branch is a serial structure of 1×1 convolution followed by dilated 3×3 convolutions to capture contextual information at different scales; and the fourth branch is an identity mapping path to preserve original information. The outputs of each branch are summed to obtain the enhanced feature map. During the inference phase, through structural reparameterization, the weights and biases of the four branches are mathematically fused into an equivalent standard 3×3 convolutional layer, achieving equivalent feature processing capabilities to the training phase with extremely low computational overhead. Finally, the processed features are fed into the classification and regression branches respectively to generate the final detection results of the target category and bounding box.

[0116] (5) Output module: Perform non-maximum suppression post-processing on the original candidate boxes output by the high-efficiency detection module to remove redundant detection, and output the final target category, bounding box position and confidence in the form of a visual image or structured data for use by users or downstream tasks.

[0117] Experimental verification:

[0118] To fully validate the effectiveness of the physically-guided UAV-FogDet framework of this invention, a systematic evaluation scheme was designed. This scheme directly addresses the three core challenges identified in the research: edge information attenuation, statistical heterogeneity caused by spatially uneven fog distribution, and limitations in computational efficiency. Experimental methods employed both synthetic and real-world datasets, demonstrating that by explicitly modeling the physical laws of atmospheric scattering, not only can superior detection performance be achieved, but high feasibility can also be maintained in deployment on practical UAV platforms.

[0119] As the basis for experimental evaluation, HazyDet is the first large-scale UAV view detection dataset specifically designed for severe weather conditions. This dataset contains 11,000 synthetic images with 364,433 annotated instances, systematically divided into training, validation, and test sets in a 7:1:2 ratio. The dataset was constructed using a carefully calibrated atmospheric scattering model to ensure realistic simulation of haze effects, accurately reflecting the atmospheric scattering physics discussed within the theoretical framework of this invention, thus providing an ideal testing platform for evaluating physics-guided methods.

[0120] The HazyDet dataset contains a variety of highly challenging features that closely match the degradation patterns that drone-fog detection frameworks aim to address. For example... Figure 6 As shown, drone images in foggy weather exhibit systematic edge information attenuation and spatially uneven degradation, which are precisely the problems that the EENet and ATRN components of this invention aim to address. Furthermore, the dataset also suffers from added complexity due to its pronounced long-tail distribution (e.g., Figure 7 As shown in the figure, the number of car instances (273,757 in the training / validation set and 60,026 in the test set) far exceeds that of trucks (13,134 and 2,553) and buses (12,220 and 2,743). This class imbalance corresponds to real-world traffic scenarios and provides a rigorous test of the ability of the method of this invention to simultaneously cope with atmospheric degradation and dataset bias.

[0121] To verify the generalization ability of the method of this invention under real atmospheric conditions beyond synthetic data, a Real Haze Drone Detection Test Set (RDDTS) was introduced. This dataset contains 600 carefully selected real-world hazy images. These images were obtained through systematic field photography and online resource acquisition, covering various environmental scenarios such as urban, rural, and coastal areas, with varying flight altitudes and atmospheric conditions. During the annotation process, a rigorous method was employed, combining semi-automatic annotation with extensive manual verification to ensure accuracy. Comparative analysis using Fréchetinception distance (FID) and kernel inception distance (KID) metrics confirmed that RDDTS can effectively simulate real-world hazy conditions captured by drones, providing important validation for the practical application value of the physical guidance method of this invention.

[0122] The experimental verification and evaluation strategy systematically covers the three key dimensions identified in the work: model complexity, computational efficiency, and detection accuracy. This multi-dimensional evaluation ensures that the physics-based solution maintains excellent detection performance while meeting the practical needs of real-world UAV applications.

[0123] Model complexity evaluation primarily focuses on the total number of parameters, and its calculation formula is as follows:

[0124] (11)

[0125] in, The parameter indicates the number of parameters in the convolutional layer, while Params represents the total number of parameters in the model. Indicates the height of the convolution kernel. Indicates the kernel width. This indicates the number of input channels. This parameter directly determines memory usage and has a significant impact on the deployment feasibility on resource-constrained drone platforms.

[0126] Computational efficiency is evaluated using floating-point operations (FLOPs), and the formula is as follows:

[0127] (12)

[0128] Where H represents the feature map height and W represents the feature map width.

[0129] A thorough analysis of the computational resources required for the inference phase is crucial for drone deployment.

[0130] Regarding detection accuracy, the average accuracy (AP) is used for evaluation according to the standard COCO evaluation protocol, and the specific formula is as follows:

[0131] (13)

[0132] in, Indicates average precision. This represents the average accuracy when the IoU threshold is t.

[0133] This comprehensive metric assesses both localization and classification accuracy. In addition to overall performance, detailed category-specific AP evaluations were conducted for different classes. Given the long-tailed distribution of the HazyDet dataset, this categorized AP evaluation is particularly important, as it helps to comprehensively evaluate the consistency of the method's performance across mainstream and underrepresented classes under atmospheric degradation conditions.

[0134] (a) Comparison with other models

[0135] A comprehensive performance comparison was conducted on state-of-the-art object detection models covering the three major architectural paradigms to verify the effectiveness of the physical law-based approach of this invention. As shown in Table 1, the evaluation included both the HazyDet and RDDTS datasets, thus achieving robust validation in both synthetic and real-world foggy environments.

[0136] Table 1. Comparison of detectors on the HazyDet and RDDTS datasets.

[0137]

[0138] Experimental results show that UAV-FogDet achieves superior performance on HazyDet, with a mean average accuracy (mAP) of 62.4% at 0.5:0.95, outperforming all comparable methods. Notably, compared to YOLOv12-S, the method of this invention still achieves a 0.4 percentage point improvement with fewer parameters; and in terms of computational resource consumption, UAV-FogDet significantly outperforms Mamba-YOLO-B, requiring only 8.3M parameters, while Mamba-YOLO-B requires 21.8M parameters. On the RDDTS dataset, UAV-FogDet performs even better: its mAP at 0.5:0.95 reaches 41.4%, 0.8 percentage points higher than the second-best algorithm (Hyper-YOLO-S); especially in the Bus category, UAV-FogDet's performance improvement is astonishing, reaching 1.4 percentage points, fully demonstrating its ability to handle class imbalance problems under real-world atmospheric conditions.

[0139] UAV-FogDet achieves these performance improvements while maintaining excellent computational efficiency. With only 8.3 million parameters and 23 billion floating-point operations, compared to Hyper-YOLO-S, the model of this invention significantly reduces the number of parameters by 43.9% and the number of floating-point operations by 40.9%, while still providing excellent detection accuracy. This achievement fully verifies the effectiveness of the RPM-Head design of this invention in resolving the contradiction between detection robustness and computational efficiency.

[0140] (II) Error Analysis

[0141] To further explore how the physical guidance design of this invention addresses specific detection challenges in foggy environments, a detailed error analysis was conducted using TIDE (Target Detection Error Analysis Tool). Figure 8 A systematic comparative analysis of the error characteristics of UAV-FogDet and YOLOv11-S was conducted, comparing the performance of UAV-Fog detection and YOLOv11-S in major and specific error categories. This reveals, at a detailed level, how the three architectural components of this invention effectively address the fundamental challenges identified in this invention.

[0142] Error analysis clearly demonstrates the effectiveness of the physical-guided method of this invention. UAV-FogDet achieves significant reductions in both false alarm rate and missed detection error, fully verifying that the EENet edge recovery capability of this invention can effectively offset the systematic information loss caused by atmospheric scattering low-pass filtering. The improvement in classification errors can be directly attributed to the fact that the ATRN module of this invention effectively preserves discriminative features through reprogrammable normalization techniques, thereby solving the statistical heterogeneity problem caused by differences in spatial fog density.

[0143] Importantly, the reduction in localization error demonstrates that EENet's systematic boundary recovery capability achieves more accurate target localization even when edges are degraded by atmospheric influences. The reduction in the number of duplicate detection errors indicates that the RPM-Head multi-branch training method of this invention generates more coherent detection outputs, thereby reducing collision detection between regions of different feature quality.

[0144] While YOLOv11-S has slightly better background error rates, this improvement comes at the cost of missing a significant number of real targets. For safety-critical drone applications, the substantial reduction in false alarm rates represents a favorable trade-off: the risk of failing to detect a target often far outweighs the risk of occasional false alarms.

[0145] (III) Ablation Experiment

[0146] To systematically verify whether each component truly addresses the challenges it is intended to solve within its theoretical framework, comprehensive ablation experiments were conducted on the HazyDet test set. Table 2 details the respective impacts of the gradual introduction of EENet, ATRN, and RPM-Head, demonstrating how each component helps solve problems such as edge information decay, statistical heterogeneity, and computational efficiency.

[0147] Table 2 Ablation Study of UAV-Fog Detection Components on the HazyDet Test Set

[0148]

[0149] The ablation experiments systematically validated the physical guidance design principles of this invention. EENet alone achieved a 0.9 percentage point performance improvement, verifying the core assumption of this invention: explicitly recovering edge information weakened by atmospheric scattering is crucial for fog detection. Furthermore, EENet significantly reduced computational overhead, demonstrating that targeted enhancements not only achieve high efficiency but also maintain good performance.

[0150] With the addition of ATRN, performance improved to 62.0% mAP, with the most significant improvement in truck detection. This verifies that the present invention's handling of statistical heterogeneity can maximize the benefits to the minority classes most severely affected by spatial variations in fog density. The reduction in inference speed reflects the computational cost of the adaptive processing mechanism.

[0151] The complete drone-fog detection model achieves optimal performance while also delivering the most significant efficiency improvement. The integration of RPM-Head reduces computational requirements to 23.0 GFLOPs (Giga Floating-Point Operations Per Second) and 8.3M (M represents millions) parameters, while increasing inference speed to 143 FPS, demonstrating the effectiveness of this invention in achieving complex training and simple inference.

[0152] The analysis by category shows that the complete model achieves a balanced improvement across all categories, with a particularly significant improvement in the truck category, confirming that the collaborative method of this invention can effectively address both atmospheric challenges and dataset bias.

[0153] Experimental verification fully demonstrates the superior performance of the physically guided solution of this invention. UAV-FogDet achieves industry-leading performance on HazyDet, with an mAP of 62.4%, and on RD-DTS, it reaches 41.4%. Compared with leading models, it requires only 8.3 million parameters and 23 billion floating-point operations, reducing computational cost by 40.9%. Systematic ablation experiments confirm that each module effectively addresses its respective core challenges, and the synergistic integration of these modules significantly improves both accuracy and efficiency.

[0154] The above embodiments are merely preferred examples of the present invention and are not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Therefore, any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention should fall within the protection scope of the present invention.

Claims

1. A physically guided UAV target detection method in foggy weather images, characterized in that, Includes the following steps: Step S1: Acquire aerial images of the target area taken by the drone in foggy weather. Step S2: Input the aerial image of the foggy day to be detected into a pre-built physically guided target detection model, the model including: An edge enhancement network is used to perform edge enhancement operations on an input feature map based on an atmospheric scattering physics model. The edge enhancement operations include: adaptive high-frequency amplification of the feature map through an edge enhancer, and recovery of boundary information at multiple spatial scales through a boundary adaptive recovery module. The attention-based reprogrammable normalization Transformer module is used to perform global context modeling and local statistical normalization on the edge-enhanced feature map, eliminating the statistical heterogeneity caused by non-uniform fog distribution. The reparameterized multi-head detection head is used to perform object classification and bounding box regression on the normalized feature map and output the detection results. Step S3: Output the target category and location information in the drone aerial image; The edge enhancer process is as follows: subtract the feature map obtained after average pooling from the input feature map to obtain the local detail difference; pass the local detail difference through a convolutional layer with a sigmoid activation function to generate adaptive weights; finally, add the input feature map and the adaptive weights pixel by pixel to obtain the edge-enhanced feature map.

2. The physically guided UAV target detection method in foggy weather according to claim 1, characterized in that, The boundary adaptive restoration module processes the input feature map at multiple different receptive field scales as follows: at each scale, adaptive average pooling downsampling, 1×1 convolution dimensionality reduction, and grouped 3×3 convolution feature extraction are performed sequentially. Then, the original spatial size is upsampled back through bilinear interpolation and fed into the corresponding edge enhancer. At the same time, local features are extracted through an independent 3×3 convolution. Finally, the edge enhancement features and local features obtained from all scales are concatenated along the channel dimension, and a final convolutional layer is used to output the multi-scale fused boundary restoration features.

3. The physically guided UAV target detection method in foggy weather according to claim 1, characterized in that, The processing steps of the attention-based reprogrammable normalized Transformer module are as follows: First, the input feature map is simultaneously flattened and rearranged in dimensions to generate a two-dimensional positional encoding; then, the flattened features are added to the positional encoding and fed into the Transformer encoder to capture global feature dependencies; finally, the output of the Transformer encoder is applied to the reprogrammable normalized layer, which calculates the mean and variance in the local spatial region and performs an affine transformation on the normalized features using learnable scaling and offset parameters to output a feature map after statistical heterogeneity elimination.

4. The physically guided UAV target detection method in foggy weather according to claim 3, characterized in that, During the training phase, the reparameterized multi-head detector's multi-branch module includes the following four branches in parallel: a standard 3×3 convolution branch, sequentially connected 1×1 and 3×3 convolution branches, sequentially connected 1×1 and dilated 3×3 convolution branches, and an identity mapping branch. During the inference phase, the weights and biases of all branches of the multi-branch module are equivalent to the weights and biases of a single-layer convolution through mathematical fusion, so that feature processing is completed in a single convolution operation.

5. A physically guided UAV target detection system in foggy weather images, characterized in that, The system is used to implement the physically guided UAV target detection method in foggy images according to any one of claims 1-4, the system comprising: The image acquisition module is used to acquire aerial images of the target drone taken in foggy conditions. The edge enhancement module is used to perform edge enhancement operations on the input feature map based on an atmospheric scattering physics model; The statistical heterogeneity elimination module is used to perform global context modeling and local statistical normalization on the edge-enhanced feature map; The high-efficiency detection module performs object classification and bounding box regression on the normalized feature map and outputs the detection results; The output module is used to output the target category and location information in the drone aerial images.

6. The physically guided UAV fog image target detection system according to claim 5, characterized in that, The edge enhancement module is embedded in the backbone network and neck network of the target detection model. The feature maps output from multiple stages of the backbone network are processed by the edge enhancement module and then sent to the neck network. The feature maps are then enhanced again by the edge enhancement module at multiple pyramid levels of the neck network.

7. The physically guided UAV fog image target detection system according to claim 6, characterized in that, The edge enhancement module includes an edge enhancer unit and a boundary adaptive restoration unit. The edge enhancer unit extracts local details and adaptively weights and amplifies them. The boundary adaptive restoration unit processes feature maps in parallel at multiple receptive field scales and restores the boundary information of targets at different depths through bilinear interpolation and feature fusion.

8. The physically guided UAV fog image target detection system according to claim 7, characterized in that, The statistical heterogeneity elimination module includes: a location encoding generation unit for adding two-dimensional spatial location encoding to the input feature map; a Transformer encoder unit, comprising a multi-head self-attention layer and a feedforward network layer, with each sub-layer followed by a residual connection; and a reprogrammable normalization unit for calculating the mean and variance of features within a local window and performing an adaptive affine transformation.

9. The physically guided UAV fog image target detection system according to claim 8, characterized in that, The efficient detection module includes a multi-branch structure during the training phase, which includes a standard 3×3 convolution path, a 1×1 convolution cascaded 3×3 convolution path, a 1×1 convolution cascaded dilated 3×3 convolution path, and an identity mapping path; during the inference phase, the multi-branch structure is equivalently fused into a single standard 3×3 convolution layer.

Images