A severe weather target detection method based on a variational physical prior

Abstract

The application provides a severe weather target detection method based on a variational physical prior, and relates to the technical field of target detection. The method first acquires a degraded scene image, calculates a color saturation decay representation and an edge gradient saliency representation, and constructs a physical guiding feature. Then, the physical guiding feature and a visual feature are weighted and fused to obtain a physical enhancement feature. Next, candidate regions are generated according to the physical enhancement feature, and a target candidate region set is screened through the collaborative scoring of a physical saliency value and a visual confidence. Then, a probability distribution of the degradation degree is constructed for the target candidate region set, and an uncertainty perception feature is obtained through random sampling. Finally, a loss weight is constructed according to the uncertainty perception feature, and the target classification result and the bounding box regression result are weighted and optimized, and the target detection result is output. The technical scheme of the method solves the problem of detection performance decline caused by image feature degradation under extreme weather conditions.

Description

Technical Field

[0001] This application relates to the field of target detection technology, and more specifically, to a method for target detection in severe weather based on variational physics priors. Background Technology

[0002] Object detection is one of the core technologies in computer vision, widely used in intelligent transportation, autonomous driving, security monitoring, and other scenarios. In recent years, deep learning-based object detection algorithms, such as Faster R-CNN, the YOLO series, and DETR, have achieved significant detection accuracy under ideal weather conditions. However, in severe weather conditions such as dense fog, heavy rain, and snow, light is scattered and absorbed by atmospheric media, such as fog particles and raindrops, during propagation. This leads to severe color collapse, decreased contrast, and blurred edges in the acquired images, resulting in feature degradation. These degradations make it difficult for detection algorithms to accurately extract the discriminative features of targets, leading to a large number of missed and false detections, which seriously restricts the reliable operation of intelligent transportation systems in extreme weather environments.

[0003] For the problem of target detection in severe weather, existing improvement solutions are mainly divided into two categories.

[0004] The first category is image enhancement-based preprocessing methods. These methods first use algorithms such as dehazing and deraining to restore the degraded image, and then input the restored image into a standard detector. Commonly used enhancement algorithms include dehazing methods based on atmospheric scattering models and image translation methods based on deep learning. However, these methods have inherent drawbacks: on the one hand, the enhancement process may introduce additional artifacts or distortions, which may interfere with subsequent detection tasks; on the other hand, since enhancement and detection are optimized as two independent stages, the optimality of the enhancement result for the detection task cannot be guaranteed.

[0005] The second category is purely data-driven domain adaptation methods. These methods transfer the model from a clear image domain to a foggy image domain through domain adaptation techniques, or utilize multi-task learning to simultaneously optimize image restoration and object detection. Representative works include domain adaptation detectors based on adversarial learning and pseudo-label methods based on self-training. Although these methods do not rely on physical models, due to the lack of explicit modeling of the physical processes of atmospheric degradation, the models struggle to learn stable feature representations in extreme degradation scenarios such as non-uniform fog distribution and dense fog, thus limiting their generalization ability.

[0006] In recent years, some research has attempted to incorporate physical priors into detection networks, such as using atmospheric scattering models to estimate transmittance or atmospheric light values ​​and injecting them as auxiliary information into feature extraction or loss functions. However, most existing schemes employ a static injection approach, treating physical priors as deterministic truth inputs and ignoring the random fluctuations and estimation errors of physical parameters in complex environments. Under extreme conditions, physical priors themselves may be severely distorted. If a rigid weighting strategy is forcibly applied, the model will fit noisy features, leading to gradient oscillations or even semantic illusions. Therefore, how to adaptively reduce the impact of unreliable physical priors while amplifying their guiding role in high-confidence regions has become a pressing technical challenge in target detection during severe weather.

[0007] Therefore, there is an urgent need for a target detection method based on variational physics priors in severe weather to solve one of the aforementioned technical problems. Summary of the Invention

[0008] The purpose of this application is to provide a target detection method for severe weather based on variational physics priors, which can solve at least one of the technical problems mentioned above. The specific solution is as follows: According to a specific embodiment of this application, this application provides a method for detecting targets in severe weather based on variational physics priors, including the following steps: A degraded scene image is acquired, and feature calculations are performed on the degraded scene image to obtain a color saturation attenuation representation and an edge gradient saliency representation. Based on the color saturation attenuation characterization and the edge gradient saliency characterization, physical guidance features are constructed. The physical guidance features are then weighted and fused with visual features to obtain physical enhancement features. Candidate regions are generated based on the physical enhancement features. For each candidate region, a physical saliency value and a visual confidence value are calculated. A comprehensive score is obtained based on the physical saliency value and the visual confidence value. Candidate regions are then filtered based on the comprehensive score to obtain a set of target candidate regions. For the set of target candidate regions, a probability distribution of degradation degree is constructed, and random sampling is performed according to the probability distribution to obtain uncertainty perception features; The target classification results and bounding box regression results are weighted and optimized based on the multi-constraint loss function to output the target detection results.

[0009] Furthermore, the method for obtaining the color saturation attenuation characterization includes: Calculate the color purity index based on the color components of the image pixels; Logarithmic mapping is performed on the color purity index to obtain a color saturation attenuation characterization that reflects the degree of spectral degradation.

[0010] Furthermore, the method for obtaining the saliency representation of the edge gradient includes: Gradient calculation is performed based on the degraded scene image to obtain the gradient response; The gradient response is standardized and nonlinearly mapped based on local regional statistics to obtain a characterization of the saliency of the edge gradient.

[0011] Furthermore, the method for constructing the physical guidance feature includes: The color saturation attenuation characterization and the edge gradient saliency characterization are concatenated by channels to obtain a joint physical tensor; Perform convolution mapping on the joint physical tensor to obtain fusion weights consistent with the visual feature dimensions, and use the fusion weights as physical guidance features.

[0012] Furthermore, the screening of the candidate regions includes: The candidate region is mapped onto the physical guidance feature, and the physical salience value within the candidate region is calculated; Obtain the visual confidence score corresponding to the candidate region; The physical saliency value and the visual confidence level are fused based on the learnable parameters to obtain the comprehensive score; Non-maximum suppression is applied to the comprehensive score to obtain the target candidate region set.

[0013] Furthermore, the severe weather target detection method also includes: Based on the set of target candidate regions, local regions are extracted from the physical guidance features to obtain local physical features, which are then used for subsequent probability modeling.

[0014] Furthermore, the method for constructing the probability distribution includes: The mean and variance parameters of the degradation degree are calculated based on the characteristics of the candidate region. Logarithmic constraints are applied to the variance parameter to ensure that the variance is non-negative.

[0015] Furthermore, the random sampling includes: Introduce a standard normal distribution variable; The uncertainty-aware feature is obtained by performing reparameterization calculations based on the mean parameter, the variance parameter, and the standard normal distribution variable.

[0016] Furthermore, the method for constructing the autonomously guided loss of the multi-constraint loss function includes: The dynamic gating weights are calculated based on the mean and variance parameters of the degradation level. The dynamic gate weight is increased when the variance parameter decreases, and decreased when the variance parameter increases.

[0017] Furthermore, the output target detection result includes: The target category is determined based on the weighted and optimized classification results; The location of the target bounding box is determined based on the weighted and optimized regression results.

[0018] Compared with the prior art, the above-described solutions of this application have at least the following beneficial effects: 1. This application discloses a target detection method for severe weather based on variational physics priors. It constructs color saturation attenuation representation and edge gradient saliency representation from the spectral and spatial dimensions, respectively. By calculating the color saturation attenuation representation and edge gradient saliency representation of the degraded scene image, it constructs physical guided features and fuses them with visual features in a weighted manner. It achieves joint description and compensation of severe weather feature degradation in the spectral and spatial dimensions, effectively alleviating the problem of decreased detection performance caused by the damage of single-dimensional features in traditional methods.

[0019] 2. This application discloses a target detection method for severe weather based on variational physics priors. It calculates the physical saliency value of candidate regions based on physical guided features and fuses it with visual confidence to obtain a comprehensive score. It filters the target candidate region set through non-maximum suppression, which solves the proposal dependency trap in which region candidate networks driven by pure vision cannot effectively generate foreground candidate boxes in extreme degradation scenarios, and significantly improves the recall rate of hidden hard cases.

[0020] 3. This application discloses a target detection method for severe weather based on variational physics priors. It constructs a probability distribution of degradation degree for the target candidate region set, obtains uncertainty perception features through random sampling, and constructs loss weights to weight and optimize the target classification results and bounding box regression results. The technical solution of this application realizes the quantification and adaptive modulation of the random uncertainty of physical priors. When the physical priors are distorted, the loss weights are automatically reduced to avoid noise interference, thereby enhancing the robustness and detection accuracy of the model under severe weather conditions. Attached Figure Description

[0021] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort. In the drawings: Figure 1 This is a flowchart illustrating a severe weather target detection method based on variational physics priors, as shown in an embodiment of this application.

[0022] Figure 2This is a schematic diagram of the model structure of the severe weather target detection method based on variational physics priors, as shown in an embodiment of this application. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0024] The terminology used in the embodiments of this application is for the purpose of describing particular embodiments only and is not intended to limit the application. The singular forms “a,” “said,” and “the” used in the embodiments of this application and the appended claims are also intended to include the plural forms, and “multiple” generally includes at least two unless the context clearly indicates otherwise.

[0025] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0026] It should be understood that although the terms first, second, third, etc., may be used in the embodiments of this application, these descriptions should not be limited to these terms. These terms are only used to distinguish the descriptions. For example, first may also be referred to as second without departing from the scope of the embodiments of this application, and similarly, second may also be referred to as first.

[0027] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that an article or device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such an article or device. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or device that includes said element.

[0028] The optional embodiments of this application are described in detail below with reference to the accompanying drawings.

[0029] In target detection tasks in harsh environments, the damage to image features caused by environmental degradation does not occur independently, but is caused by a combination of factors. According to atmospheric optics theory, photons undergo two key physical processes when passing through a scattering medium, eroding semantic information in the spectral and spatial dimensions, respectively. Background scattering introduces stray light into the atmosphere, directly diluting the original radiation of the target, manifesting as nonlinear color collapse at the spectral level; forward scattering causes spatial deflection of light, resulting in energy dissipation of the originally clear geometric contours on the projection plane, manifesting as exponential smoothing of gradients at the spatial level. Based on these physical laws, this embodiment uses color saturation attenuation as a priori to lock in the loss of features in the spectral dimension, and edge gradient saliency as a priori to lock in structural damage in the spatial dimension, jointly defining the detectability boundary of targets in complex environments.

[0030] The adverse weather target detection method based on variational physics priors described in this application is implemented through an end-to-end target detection network, which is a robust target detection framework based on variational physics priors. The network includes a backbone feature extraction network, a dual-domain physics prior module, a physics-driven active candidate generation mechanism, a variational physics uncertainty modeling module, a classifier, and a regressor. Specifically, the dual-domain physics prior module is used to calculate color saturation decay representations and edge gradient saliency representations and generate physics-guided features; the physics-driven active candidate generation mechanism is used to generate a set of target candidate regions based on the physics-guided features and visual features; the variational physics uncertainty modeling module is used to construct a probability distribution of degradation degree for the set of target candidate regions and sample uncertainty-perceived features; the classifier and regressor are used to perform weighted optimization based on the loss weights constructed from the uncertainty-perceived features and output the target detection result.

[0031] like Figure 1 and Figure 2 As shown, a method for detecting targets in severe weather based on variational physics priors includes the following steps: S1. Obtain a degraded scene image, perform feature calculation on the degraded scene image to obtain a color saturation attenuation representation and an edge gradient saliency representation, and construct physical guidance features based on the color saturation attenuation representation and the edge gradient saliency representation.

[0032] Degraded scene images refer to images affected by severe weather, such as fog, haze, rain, snow, etc. This application uses foggy day observation images as an example for illustration.

[0033] In target detection tasks in harsh environments, the damage to image features caused by environmental degradation does not occur independently, but follows strict physical laws. According to atmospheric optics theory, photons undergo two key physical processes when passing through a scattering medium, eroding semantic information from both the spectral and spatial dimensions. Background scattering introduces stray light into the atmosphere, directly diluting the target's original radiation, manifesting as nonlinear color collapse at the spectral level; forward scattering causes spatial deflection of light, resulting in energy dissipation of the originally clear geometric contours on the projection plane, manifesting as exponential smoothing of gradients at the spatial level. Based on these physical laws, this application's embodiments use color saturation attenuation as a priori to lock in spectral feature loss and edge gradient saliency as a priori to lock in spatial structural damage, jointly defining the detectability boundary of targets in complex environments.

[0034] The expression for the atmospheric physics model is: in, The image referred to is a degraded scene image, i.e., a foggy observation image, in the technical solution of this application embodiment. Represents a clear image; Indicates transmittance; A represents atmospheric light value; This indicates the background light item.

[0035] As can be seen from the atmospheric physics model, the image received by the sensor consists of two parts: one part is the target signal after atmospheric attenuation. The other part is the background light signal. As scene depth increases, transmittance... Exponential decay, background light term Gradually gaining dominance. In the RGB color space, this manifests as the pixel's color vector continuously shrinking towards the brightness axis, meaning color saturation gradually decreases, while overall brightness gradually approaches the atmospheric light value A. In other words, the denser the smog, the more severely the target's original color information is covered by the background light, and the deeper the erosion of color energy.

[0036] Therefore, in the technical solution of this application embodiment, the remaining degree of color energy is quantified by calculating the product of the saturation and brightness of the pixel.

[0037] S101. This application embodiment provides a preferred technical solution, the method for obtaining color saturation attenuation characterization includes the following steps: The color purity index is calculated based on the color components of the image pixels.

[0038] The input degraded image tensor is .in, H represents the three color channels RGB; H represents the height; W represents the width.

[0039] Define the color purity measure of pixel x Color purity measurement From RGB components The decision, expressed as: in, It represents a measure of color purity, reflecting the color saturation of pixel x, and indicating the degree of saturation attenuation caused by background light superposition. These represent the intensity values ​​of pixel x in the three color channels; max indicates the maximum value among the RGB channels; min indicates the minimum value among the RGB channels. This represents a very small constant used to prevent numerical anomalies.

[0040] Logarithmic mapping was performed on the color purity index to obtain a color saturation attenuation characterization that reflects the degree of spectral degradation.

[0041] As environmental degradation deepens, the product space of saturation S and luminance V undergoes nonlinear contraction. To capture this spectral degradation trajectory, embodiments of this application define a parameter-free single-channel tensor. The specific expression for measuring color energy erosion is as follows: in, It represents the color saturation decay characterization of a single-channel tensor, measuring the feature loss in the spectral dimension; It represents the product of saturation and brightness; The mapping function represents the product of saturation and brightness; S represents saturation; V represents brightness. This represents the saturation of the pixel at spatial coordinate y; This represents the brightness of the pixel at spatial coordinate y; This represents the brightness of the pixel at spatial coordinate x. 'c' indicates iterating through the three color channels of RGB. y represents the intensity value of the c-th color channel at x; y represents the iteration variable. It represents the global spatial coordinate set, which is the position of all pixels in the entire image; It represents the natural logarithm.

[0042] The technical solution of this application embodiment is achieved through... The introduction of this method simulates the nonlinear sensitivity of the human visual system to color degradation. At low degradation levels, the human eye is highly sensitive to subtle color changes, and the logarithmic mapping amplifies this difference; at high degradation levels, the sensitivity of the human eye decreases, and the logarithmic mapping correspondingly compresses the response range. The color saturation decay characterization... A mask coefficient is provided from the spectral dimension, indicating the degree of unreliability of color information in a local area, thus providing a physical basis for feature compensation. The higher the value, the more color energy the pixel has lost relative to the clearest area in the image; that is, the more the background light covers that area, the less reliable the color information. Conversely, a lower value indicates a lower value. The lower the value, the better the color energy is preserved.

[0043] S102. This application embodiment provides a preferred technical solution, the method for obtaining the saliency characterization of edge gradients includes the following steps: The saliency characterization of edge gradients is responsible for addressing the problem of forward-scattered blurred targets. Due to the inhomogeneity of the scattering medium, the point spread function broadens, resulting in a low-pass filtering effect that impairs the key high-frequency gradients for classifying targets.

[0044] Gradient calculation is performed based on the degraded scene image to obtain the gradient response.

[0045] To quantify structural damage, the technical solution of this application utilizes discrete differential operators: horizontal and vertical differential operators, to capture the first-order energy distribution of the image. The specific expression is as follows: in, This represents the gradient magnitude of a degraded scene image, used to measure the edge intensity or local contrast of the degraded scene at pixel x. The larger the value, the greater the likelihood that the pixel is located in an edge or textured area. The smaller the value, the smoother the region; Represents the horizontal differential operator; This represents the vertical differential operator.

[0046] The gradient response is standardized and nonlinearly mapped based on local regional statistics to obtain a characterization of the saliency of the edge gradient.

[0047] To eliminate reliance on artificial hyperparameters, the technical solution of this application utilizes the expectation and standard deviation within a local topological window centered at x to construct a standardized local response ratio, and maps it to a nonlinear exponential space to obtain a characterization of the marginal gradient significance. The specific expression is as follows: in, This represents the significance of the marginal gradient. Represents a local topological window centered at x; Represents a local topological window centered at x. The gradient mean within; This represents the standard deviation of the gradient within a local window; This represents an exponential function.

[0048] The technical solution of this application embodiment utilizes parameterless standardized design to ensure strict uniformity of mathematical dimensions, and accurately locates the geometric risk area caused by forward scattering by using the relative persistence intensity of spatial gradients. The higher the value, the weaker the gradient response in that region is relative to its neighborhood, meaning the more severe the structural damage. The lower the value, the better the edge structure is preserved.

[0049] S2. Construct physical guidance features based on the color saturation attenuation representation and the edge gradient saliency representation, and then perform weighted fusion of the physical guidance features and visual features to obtain physical enhancement features.

[0050] The coherence between color saturation decay representation and edge gradient saliency representation is built upon a full-dimensional description of physical degradation. In the overall technical architecture of the adverse weather target detection method based on variational physics priors proposed in this application, the two priors are logically deeply coupled through a joint physical attention mechanism.

[0051] S201. Construct physical guidance features based on the color saturation attenuation characterization and the edge gradient saliency characterization.

[0052] Characterizing color saturation decay Characterization of marginal gradient significance The data is concatenated along the channel dimension to form a physical guidance tensor, the specific expression of which is: in, Represents the physical guidance tensor; The physical guidance tensor has a dimension of 2 channels, a height of H, and a width of W.

[0053] To enable physical priors to accurately guide the purification of high-dimensional semantic features, the technical solution of this application constructs a physical projection layer based on a gating mechanism. The physical projection layer utilizes a set of learnable priors... The convolutional kernel maps the number of channels from 2 to a dimension consistent with the features of the backbone network, generating channel-aware spatial feature fusion weights, specifically expressed as: in, This represents the fusion weight at pixel position x, which is also known as the physical guidance feature. C represents the dimension of the backbone network features, which is the number of channels; Indicates learnable Convolution kernel parameters; This represents the learnable bias parameter; This represents the Sigmoid activation function.

[0054] S202. The physical guidance features and visual features are weighted and fused to obtain physical enhancement features.

[0055] Based on the fusion weights, also known as physical guidance features, the visual features are weighted positionally to construct the physical enhancement features. The deep feature tensors extracted from the backbone network are weighted positionally to obtain weighted features, which are then sequentially processed... After convolution operations and batch normalization, the deep feature tensors are weighted and processed accordingly. Perform position-by-position weighting, then sum the results of the two branches to construct the physical enhancement feature, the specific expression of which is: in, Indicates physical enhancement features; This represents the deep feature tensor extracted by the backbone network; This indicates a batch normalization operation; express Convolution operation.

[0056] The technical solution of this application embodiment is that when the color feature fails due to background scattering, that is... A high value forces a shift to cues that depend on spatial structure, which means... Guiding; when the structure is completely smoothed, If the value is high, target discrimination is performed through color residue. The bidirectional compensation mechanism of spectrum and space solves the lack of robustness of traditional methods in extreme environments due to the single feature dimension, and provides a highly physically consistent input manifold for subsequent Bayesian uncertainty modeling, that is, variational physical uncertainty modeling.

[0057] In target detection tasks in harsh environments, existing physics-guided paradigms generally fall into a proposal dependency trap. Traditional models heavily rely on purely vision-driven region candidate networks for initial foreground target screening, only introducing physical priors within the candidate boxes later. However, in extremely degraded scenarios, the optical features of targets are severely eroded, causing standard region candidate networks to fail to generate effective foreground responses, leading to fatal false negatives and false positives. Once the front-end candidate box generation fails, all subsequent Bayesian uncertainty inferences and physical autonomy constraints completely lose their basis for effectiveness. To address the above problems, the technical solution of this application implements physics-driven active candidate generation through the following steps.

[0058] S3. Generate candidate regions based on the physical enhancement features, calculate the physical saliency value and visual confidence for each candidate region, obtain a comprehensive score based on the physical saliency value and the visual confidence, and filter candidate regions based on the comprehensive score to obtain a set of target candidate regions.

[0059] S301. Generate candidate regions based on the physical enhancement features, calculate the physical saliency value and visual confidence for each candidate region, and obtain a comprehensive score based on the physical saliency value and the visual confidence.

[0060] Physical enhancement features Input the region candidate network to generate an initial set of anchor boxes. Where N represents the total number of anchor frames; This represents the spatial coordinates of the i-th anchor frame. These represent the center coordinates, These represent the width and height, respectively.

[0061] In the initial screening stage of candidate boxes, an adaptive region feature alignment operator is introduced. RoI Align, also known as RoI Align, uses physical guidance features. With each anchor frame Perform spatial mapping and calculate local expectation to extract the physical saliency confidence at the anchor frame level. The specific expression is as follows: in, Indicates the confidence level of physical significance; This indicates that the average value of all pixels within the region is taken. This indicates the RoI Align adaptive region feature alignment operation, where i represents the index and N represents the total number of anchor boxes.

[0062] Obtain the physical confidence vector covering all initial anchor boxes: .in, Indicates the 1st Physical significance confidence of N anchor frames; Indicates transpose; Let N be the number of rows and 1 column of real vector space.

[0063] The standard region candidate network outputs a visual confidence vector for the anchor box set as follows: .in, Represents the visual confidence vector; Indicates the 1st Visual confidence of N anchor frames.

[0064] To perform nonlinear calibration on the foreground probability, the physical saliency value and the visual confidence value are fused based on learnable parameters to obtain the comprehensive score, specifically expressed as: in, Represents the overall score vector; This represents a learnable adaptive adjustment scalar.

[0065] The technical solution of this application embodiment, when a potential target When visual features fail due to extreme scattering, if the region exhibits a strong anomalous degradation response in the two-domain prior, it transcends the limitations of visual features and passes through a scalar... The adjustment forcefully boosts the overall score of the anchor box, retaining extremely difficult samples that would otherwise be discarded as valid candidate boxes.

[0066] The comprehensive score Perform non-maximum suppression and output the set of target candidate regions. , of which M .

[0067] According to the target candidate region set In the physical guiding features A local area is extracted to obtain local physical features, which are then used for subsequent probabilistic modeling. Using M candidate boxes, the physical guiding features are further analyzed. The execution space is truncated, and the specific expression is: in, represents the local physical features corresponding to the j-th candidate box; K represents the spatial resolution of the output.

[0068] The technical solution of this application embodiment achieves multiple functions by repeatedly using physical guidance features at different stages of the detection process. First, in the candidate region screening stage, physical guidance features are used to guide spatial localization. By mapping each initial anchor box to the physical guidance feature and calculating the local expectation, the physical saliency value at the anchor box level is extracted, thereby forming a physical confidence vector. Second, after obtaining the target candidate region set, the physical guidance features are used again. Through spatial truncation operations, the local physical features corresponding to each candidate box are extracted from the physical guidance features. These local physical features are used for subsequent probabilistic modeling. In this way, physical guidance features are used to calculate physical saliency values ​​to guide the selection of candidate boxes in the candidate region screening stage, and to truncate local physical features in the probabilistic modeling stage to support the construction of the probability distribution of degradation degree. This achieves multiple uses of the same feature in front-end guided spatial localization and back-end probabilistic modeling. This design allows physical guidance features to run through multiple key links in the detection process, forming a complete closed loop from physical prior calculation to candidate box generation and uncertainty modeling, enhancing the logical self-consistency of the entire-link physical guidance target detection.

[0069] S4. For the set of target candidate regions, construct a probability distribution of degradation degree, and perform random sampling according to the probability distribution to obtain uncertainty perception features.

[0070] In engineering practice of target detection in harsh environments, injecting reconstructed physical priors deep into neural networks is an effective means of compensating for visibility loss. However, existing paradigms generally treat physical priors as deterministic truth inputs, ignoring their inherent random fluctuations in complex scenarios. Affected by sensor thermal noise, non-uniform atmospheric turbulence, and errors in physical parameter estimation, physical priors inevitably carry accidental uncertainties. To address this issue, the technical solution of this application embodiment achieves variational physical uncertainty modeling through the following steps.

[0071] The technical solution of this application embodiment no longer characterizes the degradation degree of each candidate region as a single scalar, but defines the degradation degree of each candidate region as a latent random variable constrained by local observations. Based on variational inference theory, it is assumed that the true degradation level follows a parameterized Gaussian distribution in the latent manifold space, specifically expressed as: in, Represents the conditional probability density function; Indicates the degradation intensity of the i-th candidate region; This represents the local physical features corresponding to the i-th candidate region. ; The set of learnable parameters representing the variation head; Indicates a Gaussian distribution; The mean parameter represents the core estimate of the degradation intensity from the physical prior. This represents the variance parameter, which quantifies the random uncertainty of the estimation results under the current environmental noise.

[0072] To achieve automatic learning of the distribution, the technical solution of this application constructs a lightweight variational head and utilizes a multilayer perceptron to perform heteroscedasticity decomposition at the feature extraction layer. To ensure the numerical stability of the training process and satisfy the physical constraint of non-negativity of variance, the network is designed to predict the log-variance term, specifically expressed as: in, This represents the predicted logarithmic variance term. ; This represents a variational multilayer perceptron network.

[0073] To ensure the probability sampling process is end-to-end differentiable, this embodiment introduces a standard normal distribution auxiliary variable. .in, This represents a Gaussian distribution. Through reparameterization techniques, the random mapping is transformed into a deterministic function output, specifically expressed as: in, This indicates the characteristics of uncertainty perception.

[0074] The technical solution of this application establishes an effective probability buffer along the physical injection path through reparameterized computation. When a decrease in prior input quality is detected, a larger prediction variance increases the randomness of sampling, thereby triggering the suppression mechanism in the subsequent loss function to protect the feature representation of the backbone network from being contaminated by degraded signals. The variational physical uncertainty modeling mechanism achieves adaptive soft switching between physical guidance and pure semantic detection modes through probabilistic inference, enabling the physical prior to evolve from a single static constraint into an intelligent guide with self-awareness and correction capabilities.

[0075] S5. Based on the multi-constraint loss function, the target classification results and bounding box regression results are weighted and optimized to output the target detection results.

[0076] In target detection in harsh environments, converting physical priors into loss weights is a common practice for mining difficult samples. However, existing weighting mechanisms generally suffer from deterministic blind following: under extreme conditions, physical priors are easily distorted. If a rigid amplification strategy is forcibly applied, the system will fit noisy features, leading to gradient oscillations or even semantic illusions. To overcome this problem, the technical solution of this application implements an uncertainty-modulated physically autonomous guided loss through the following steps.

[0077] The technical solution of this application adopts an end-to-end joint training strategy and designs a multi-constraint loss function. The specific expression is: in, Represents a multi-constraint loss function; Indicates the regional candidate network loss; This represents the bounding box regression loss; This indicates losses caused by self-discipline. This represents a hyperparameter used to balance the autonomous guidance loss. The weighting coefficients for the contribution to the overall objective function.

[0078] and It ensures basic spatial positioning and coordinate regression capabilities. This completely replaces the original classification cross-entropy loss. The technical solution of this application, through end-to-end multi-task design, enables the model to adaptively resist extreme prior noise while maintaining stable localization, comprehensively improving detection performance under harsh conditions. By using a self-guided loss to weight and optimize the target classification results, the model can adaptively adjust the weights of the classification loss based on the reliability of the physical prior during training.

[0079] This application provides a preferred technical solution. The method for constructing the self-regulating loss includes calculating dynamic gating weights based on the mean and variance parameters of the degradation degree. The specific expression is: in, Indicates dynamic gating weights; This represents a learnable parameter used to control the degree of influence of the mean degradation intensity on the dynamic gating weights.

[0080] When variance parameter When decreasing, Approaching 1, increase dynamic gating weights When the variance parameter When it increases, Approaching 0, dynamic gating weight reduce.

[0081] By combining the predicted probabilities of candidate regions, an autonomous guidance loss is constructed. The specific expression is: in, This represents the probability that the i-th candidate region output by the classifier belongs to the target category; Represents the regularization coefficient; Indicates the candidate region; This represents the set of candidate regions containing the target.

[0082] Based on the multi-constraint loss function The weighted and optimized classification results determine the target category. After end-to-end training of the network using a multi-constraint loss function, the robust target detection framework based on variational physics priors can be used for inference. In the inference phase, the degraded scene image to be tested is input into the network. The backbone feature extraction network extracts visual features, the dual-domain physics prior module calculates color saturation decay representation and edge gradient saliency representation and generates physical guidance features, the physics-driven active candidate generation mechanism generates a set of target candidate regions based on the physical guidance features and visual features, the variational physics uncertainty modeling module constructs a probability distribution of degradation degree for the set of target candidate regions and samples to obtain uncertainty-aware features, and the classifier and regressor are weighted and optimized based on the loss weights constructed from the uncertainty-aware features.

[0083] The target category is determined based on the weighted optimization classification results: For each candidate region in the target candidate region set, the classifier output is processed by a multi-constraint loss function. The weighted and optimized classification probability distribution selects the category with the highest probability as the target category for the candidate region. Categories include common targets in traffic scenarios such as pedestrians, cars, buses, bicycles, motorcycles, riders, and trucks. The target bounding box location is determined based on the weighted and optimized regression results: for each candidate region in the target candidate region set, the regressor outputs the bounding box regression loss. The optimized bounding box coordinates include the x-coordinate and y-coordinate of the center point, the width of the bounding box, and the height of the bounding box. This completes the severe weather target detection based on variational physics priors, outputting the category label and bounding box location for each detected target.

[0084] This application's embodiments verify the effectiveness of the adverse weather target detection method based on variational physics priors through comparative experiments. The specific method and data in the RTTS dataset are shown in Table 1 below: Table 1 Comparison results of RTTS datasets Table 1 shows the experimental comparison results, indicating that the proposed target detection method based on variational physics priors in severe weather achieves a mAP of 75.21% on the RTTS fog dataset, significantly outperforming existing mainstream methods. Simultaneously, the inference speed reaches 154 FPS with only 18.48 M parameters, demonstrating significant advantages in detection accuracy, real-time performance, and model lightweighting. In particular, the robustness of this method for detecting small targets such as pedestrians, bicycles, and motorcycles in foggy conditions is significantly improved, validating its practical value in intelligent transportation systems.

[0085] The specific methods and data for the Foggy Cityscapes dataset are shown in Table 2 below: Table 2 Comparison results of the Foggy Cityscapes dataset Table 2 shows the experimental comparison results, demonstrating the performance of our proposed method against several existing methods on the Foggy Cityscapes dataset. The results show that our proposed method achieves leading or near-optimal detection accuracy across seven categories: people, vehicles, buses, bicycles, motorcycles, riders, and trucks. Specifically, the vehicle detection accuracy reaches 66.57%, bus 51.74%, rider 62.4%, and truck 40.4%, all significantly outperforming most of the compared methods. The overall mean accuracy (mAP@0.5) reaches 30.95%, ranking highest among all methods. Furthermore, our proposed method achieves an inference speed of 154.5 FPS with only 18.01M parameters, achieving an optimal balance between detection accuracy, real-time performance, and lightweight model, fully validating its robustness and practicality in complex foggy scenarios.

[0086] In summary, the severe weather target detection method based on variational physics priors provided in this embodiment first acquires a degraded scene image, calculates color saturation attenuation representation and edge gradient saliency representation, and constructs physical guidance features accordingly. Then, the physical guidance features are weighted and fused with visual features to obtain physical enhancement features. Next, candidate regions are generated based on the physical enhancement features, and the physical saliency value and visual confidence of each candidate region are calculated. A comprehensive score is obtained by fusing learnable parameters, and a set of target candidate regions is selected through non-maximum suppression. Local physical features are extracted from the physical guidance features. Then, for the set of target candidate regions, the degradation degree is modeled as a Gaussian distribution, and uncertainty perception features are obtained by sampling using reparameterization techniques. Finally, dynamic gating weights are constructed based on the degradation mean and variance to form an autonomous guidance loss, which is jointly optimized with the region candidate network loss and bounding box regression loss. The target classification results and bounding box regression results are weighted and optimized to output the target category and bounding box position. This method was validated on the RTTS and FoggyCityscapes datasets, achieving an inference speed of over 154 frames per second. It effectively reduces the missed detection of small targets such as pedestrians, bicycles, and motorcycles obscured by dense fog, and significantly improves the robustness and real-time performance of target detection under adverse weather conditions.

[0087] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0088] The units described in the embodiments of this application can be implemented in software or hardware. The names of the units are not, in some cases, limiting the scope of the unit itself.

Claims

1. A method for target detection in severe weather based on variational physics priors, characterized in that, Includes the following steps: A degraded scene image is acquired, and feature calculations are performed on the degraded scene image to obtain a color saturation attenuation representation and an edge gradient saliency representation. Based on the color saturation attenuation characterization and the edge gradient saliency characterization, physical guidance features are constructed. The physical guidance features are then weighted and fused with visual features to obtain physical enhancement features. Candidate regions are generated based on the physical enhancement features. For each candidate region, a physical saliency value and a visual confidence value are calculated. A comprehensive score is obtained based on the physical saliency value and the visual confidence value. Candidate regions are then filtered based on the comprehensive score to obtain a set of target candidate regions. For the set of target candidate regions, a probability distribution of the degree of degradation is constructed, and random sampling is performed according to the probability distribution to obtain uncertainty perception features; the uncertainty perception features include mean parameters and variance parameters; A dynamic gating weight is constructed based on the mean parameter and the variance parameter. An autonomous guidance loss is constructed based on the dynamic gating weight. A multi-constraint loss function is constructed based on the autonomous guidance loss and the bounding box regression loss. The target classification result and the bounding box regression result are weighted and optimized based on the multi-constraint loss function, and the target detection result is output.

2. The method for detecting targets in severe weather according to claim 1, characterized in that, The method for obtaining the color saturation attenuation characterization includes: Calculate the color purity index based on the color components of the image pixels; Logarithmic mapping is performed on the color purity index to obtain a color saturation attenuation characterization that reflects the degree of spectral degradation.

3. The method for detecting targets in severe weather according to claim 1, characterized in that, The method for obtaining the significance representation of the edge gradient includes: Gradient calculation is performed based on the degraded scene image to obtain the gradient response; The gradient response is standardized and nonlinearly mapped based on local regional statistics to obtain a characterization of the saliency of the edge gradient.

4. The method for detecting targets in severe weather according to claim 1, characterized in that, The method for constructing the physical guidance feature includes: The color saturation attenuation characterization and the edge gradient saliency characterization are concatenated by channels to obtain a joint physical tensor; Perform convolution mapping on the joint physical tensor to obtain fusion weights consistent with the visual feature dimensions, and use the fusion weights as physical guidance features.

5. The method for detecting targets in severe weather according to claim 1, characterized in that, The screening of the candidate regions includes: The candidate region is mapped onto the physical guidance feature, and the physical salience value within the candidate region is calculated; Obtain the visual confidence score corresponding to the candidate region; The physical saliency value and the visual confidence level are fused based on the learnable parameters to obtain the comprehensive score; Non-maximum suppression is applied to the comprehensive score to obtain the target candidate region set.

6. The method for detecting targets in severe weather according to claim 5, characterized in that, The severe weather target detection method also includes: Based on the set of target candidate regions, local regions are extracted from the physical guidance features to obtain local physical features, which are then used for subsequent probability modeling.

7. The method for detecting targets in severe weather according to claim 1, characterized in that, The method for constructing the probability distribution includes: The mean and variance parameters of the degradation degree are calculated based on the characteristics of the candidate region. Logarithmic constraints are applied to the variance parameter to ensure that the variance is non-negative.

8. The method for detecting targets in severe weather according to claim 1, characterized in that, The random sampling includes: Introduce a standard normal distribution variable; The uncertainty-aware feature is obtained by performing reparameterization calculations based on the mean parameter, variance parameter, and the standard normal distribution variable.

9. The method for detecting targets in severe weather according to claim 1, characterized in that, The method for constructing the autonomously guided loss of the multi-constraint loss function includes: The dynamic gating weights are calculated based on the mean and variance parameters of the degradation level. The dynamic gate weight is increased when the variance parameter of the degradation degree decreases, and decreased when the variance parameter of the degradation degree increases.

10. The method for detecting targets in severe weather according to claim 1, characterized in that, The output target detection results include: The target category is determined based on the weighted and optimized classification results; The location of the target bounding box is determined based on the weighted and optimized regression results.

Images