Self-correcting visibility measurement method based on vision-scattering spectrum bimodal fusion

Abstract

The invention relates to the technical field of atmospheric visibility detection, in particular to a self-correcting visibility measurement method based on vision-scattering spectrum bimodal fusion, which comprises the following steps of: synchronously acquiring and preprocessing bimodal data; performing scene classification based on vision; performing visibility inversion based on a scattering spectrum; and bimodal dynamic gating fusion is continuously optimized based on a distributed model of a cloud-edge collaborative architecture. According to the method, the accuracy of atmospheric visibility detection can be improved through inversion operation, and the interpretability of model behaviors and the following of physical laws can be improved by introducing physical priori; visual information and scattering spectrum information are subjected to deep and dynamic interactive fusion in the model through fusion operation, and the macroscopic scene understanding capability of vision and the microscopic physical detection capability of a scattering spectrum are fully utilized, so that the limitation of a single sensor is overcome, and the accuracy of atmospheric visibility detection is improved.

Description

Technical Field

[0001] This invention relates to the field of atmospheric visibility detection technology, and in particular to a self-calibrating visibility measurement method based on visual-scattering spectrum dual-modal fusion. Background Technology

[0002] Atmospheric visibility, as a core physical quantity characterizing atmospheric transparency, directly impacts the safety and efficiency of modern society. In highly automated transportation systems, visibility data is a crucial decision-making basis for ensuring precise takeoffs and landings in civil aviation, smooth navigation of waterways, and highway management. Whether it's instrument landing guidance for aircraft under complex weather conditions, ship navigation in busy port waterways, or real-time early warning of sudden fog on highways, extremely high-frequency and accurate visibility feedback is essential to avoid mass casualties caused by limited visibility. This need for atmospheric perception extends beyond traffic safety to in-depth monitoring and scientific assessment of the atmospheric environment. Fluctuations in atmospheric visibility are often accompanied by drastic changes in aerosol particle concentration, particle size distribution, and physical properties, making it the most intuitive scientific indicator for assessing urban smog levels and tracking pollution source diffusion trends, providing indispensable underlying data support for environmental protection decisions. Within meteorological forecasting and disaster prevention and mitigation systems, long-term visibility monitoring can effectively identify different weather phenomena such as fog, haze, and dust storms, thus providing critical early warning windows for social operations. In special scenarios such as military operations, open-pit mining, and the operation of precision optical instruments, visibility directly impacts the effectiveness of visual detection and operational safety. Although various industries are increasingly reliant on visibility detection, existing single-method detection approaches remain insufficiently robust in complex atmospheric environments. Physical scattering methods are prone to fundamental biases in non-uniform atmospheres or mixed precipitation scenarios, while pure visual methods, lacking the constraints of underlying physical laws, struggle to guarantee long-term stability of detection results in scenarios with drastic changes in illumination or similar characteristics. This contradiction between high-standard application requirements and existing technological bottlenecks makes the development of a detection method capable of integrating multimodal information and possessing adaptive evolutionary capabilities a pressing issue in the field of atmospheric detection.

[0003] Currently, methods for detecting atmospheric visibility include: 1. Physical detection method: This method uses an instrument (e.g., a forward scattering instrument) to emit a modulated beam of light. By measuring the degree of forward scattering of the beam by a small volume of air, the atmospheric extinction coefficient is calculated. Then, the atmospheric visibility is obtained by inversion according to Koschmieder's law or Allard's law. 2. Vision-based image processing method: This method uses a camera to acquire scene images and continuously improves atmospheric visibility by analyzing features such as contrast, texture, color, and sharpness of the scene images. Early vision-based image processing methods relied on atmospheric illumination models (e.g., dark channel priors) and manual feature extraction. In recent years, with the development of deep learning, convolutional neural networks (CNNs) or Transformers have become research hotspots. These methods (e.g., using forward scattering instrument detection values ​​as training labels for visual models or as simple post-processing calibration references) are trained end-to-end on large datasets of "image-atmospheric visibility ground truth" and can learn more complex nonlinear relationships between image features and visibility. However, these methods are merely loose combinations and do not achieve deep, real-time, and dynamic fusion of the two modalities at the algorithm model level. Essentially, they still rely on a single detection instrument (e.g., forward scattering instrument) and fail to solve the fundamental problem of bias in both detection instruments under complex scenes. Furthermore, they lack the ability to adapt and self-correct under long-term operation. Both methods have the following drawbacks: A. The physical detection method is overly simplistic and has poor environmental adaptability: The forward scattering instrument is based on the ideal assumptions of "uniform atmosphere" and "specific particle size spectrum". When there are complex particle distributions such as rain, snow and fog in the real atmosphere, the assumptions do not hold, which leads to systematic bias in the inversion and a sharp drop in detection accuracy. B. Lack of physical constraints: Visual methods only process pixel information in images and lack direct physical measurement basis. When encountering scenes with similar visual features but different physical causes, the visual method memory becomes confused, leading to classification errors or jumps in atmospheric visibility estimation values, which affects the accuracy of atmospheric visibility detection. C. The fusion mechanism is simple and the complementarity of information is not fully utilized: Existing visual methods only use physical detection values ​​as static labels or post-calibration values ​​for the poetry model. This "one-way" information flow cannot use physical signals to dynamically guide visual analysis during real-time inference, nor can it use visual scene classification to know the selection of physical models. The fusion level is too shallow. When there is uncertainty in both modal data, the system will still fail, affecting the accuracy of atmospheric visibility detection. D. The system lacks adaptive evolution capabilities and has poor long-term reliability: The algorithm model deployed at the edge is fixed and unchanging. The hardware of the detection instruments (e.g., forward scatterers) will age and the lenses will become contaminated. At the same time, climate change can bring new weather phenomena, and the distribution of the model's input data will deviate from that during training, causing its performance to "silently decay" over time. The final detection results are unreliable and affect the accuracy of atmospheric visibility detection. Summary of the Invention

[0004] The technical problem to be solved by this invention is: in order to solve the problem of low detection accuracy of existing atmospheric visibility detection methods, this invention provides a self-calibrating visibility measurement method based on visual-scattering spectrum dual-modal fusion. By improving the atmospheric visibility detection method, the interaction and fusion of scattering spectrum information and visual information are realized, thereby improving the accuracy of atmospheric visibility detection.

[0005] The technical solution adopted by this invention to solve its technical problem is: a self-calibrating visibility measurement method based on visual-scattering spectrum dual-modal fusion, comprising the following steps: S1. Dual-modal data synchronous acquisition and preprocessing: The image acquisition unit captures a frame of image, and the detector array synchronously samples to obtain the multidimensional scattering spectrum vector. And analyze the acquired image frame and multidimensional scattering spectrum vector. Preprocessing is required; S2. Vision-based scene classification: Obtain the scene images from S1 and utilize a lightweight scene classification network to output the scene probability distribution vector. ; S3. Visibility inversion based on scattering spectrum: Obtain the multidimensional scattering spectrum vector in S1. It also utilizes a KAN-enhanced physical prior to guide the residual correction network, outputting multiple preliminary visibility prediction vectors. ; S4. Dual-modal dynamic gating fusion: Obtain the scene probability distribution vector in S2. Multiple preliminary visibility prediction vectors in S3 To calculate the final output visibility value. ; S5. Continuous optimization of the distributed model based on cloud-edge collaborative architecture: Obtain the visibility value of the final output in S4. And continuously optimize it.

[0006] Therefore, inversion operations can improve the accuracy of atmospheric visibility detection. Furthermore, by introducing physical priors, the interpretability of model behavior and adherence to physical laws can be enhanced. Fusion operations enable deep and dynamic interaction between visual and scattering spectrum information within the model, fully utilizing the macroscopic scene understanding capabilities of vision and the microscopic physical detection capabilities of scattering spectra to overcome the limitations of a single sensor and improve the accuracy of atmospheric visibility detection. Continuous optimization operations enable the entire device to self-diagnose detection times with high uncertainty and proactively report them to the cloud. The cloud then analyzes and optimizes the model using a more precise source, feeding back the optimization capabilities to edge devices. This enables continuous learning and self-optimization of the entire system, ensuring the long-term accuracy of the entire device's detection.

[0007] Further, S2 includes the following steps: S2-1: Obtain the scene images from S1, and use a lightweight asymmetric backbone network to output feature pyramids with different resolutions and semantic levels. ; S2-2, Obtain the feature pyramid from S2-1 And a semantic-texture decoupling enhancement module is used to obtain the final enhanced feature map. ; S2-3, Obtain the final enhanced feature map from S2-2. It also utilizes a classification head to output a scene probability distribution vector. .

[0008] Furthermore, in S2-1, the lightweight asymmetric backbone network consists of multiple stages connected in series. Each stage is composed of N stacked asymmetric residual models. The asymmetric residual models are the basic building blocks of the lightweight asymmetric backbone network. The asymmetric residual models employ a large convolutional kernel integration strategy internally. Convolution decomposed into a and convolution; For an input feature map Its output The calculation formula is: ; in: This indicates a modified linear unit activation function. Indicates the normalization layer. express Convolution operation, express Convolution operation; Downsampling is performed between each stage using convolutions with a stride of 2, gradually expanding the receptive field to extract higher-level semantic information, ultimately forming a feature pyramid containing rich spatial and semantic information. .

[0009] Further, S2-2 includes the following steps: S2-2-1, Multi-scale Feature Preprocessing: Obtaining the Feature Pyramid from S2-1 And through bilinear interpolation of deep feature maps Upsampling is performed to make its spatial size match the feature map. To maintain consistency, then compare the upsampled feature map with the original feature map. The data is stitched together along the channel dimension to form a fusion feature map that aggregates multi-scale information. ; S2-2-2, Obtain the fusion feature map from S2-2-1 and the fused feature map Through a Convolutional layers to obtain semantic channel groups and texture channel groups ; S2-2-3, Obtain the semantic channel group in S2-2-2 and texture channel groups And perform enhancement branching to obtain enhanced semantic features. and enhanced texture features ; S2-2-4, Obtain the enhanced semantic features from S2-2-3 and enhanced texture features Through enhanced semantic features Generate dynamic gating signals The dynamic gating signal Enhanced texture features for adaptive adjustment The contribution of each feature map is then analyzed and fused to obtain the final enhanced feature map. ; In S2-2-1, feature maps are fused. The expression is: ; in: This indicates a splicing operation at the channel level. This indicates a bilinear interpolation upsampling operation; In S2-2-3, semantic channel groups Enter the channel attention module to calculate the channel attention weight vector. and the channel attention weight vector Apply to semantic channel group The above is used to obtain enhanced semantic features. ; Channel attention weight vector The expression is: ; in: This represents a layer consisting of two fully connected layers (with weights respectively). , A multilayer perceptron composed of ) This indicates a global average pooling operation. This represents the Sigmoid activation function; After enhancing semantic features The expression is: ; in: This represents a broadcast-based element-wise multiplication operation, utilizing a broadcast mechanism to distribute the channel attention weight vector. Applied to semantic channel groups At each spatial location; Texture Channel Group The spatial attention module then analyzes the importance of different spatial locations on the feature map to calculate the spatial attention weight map. And spatial attention weight map Apply to texture channel group Above, to obtain enhanced texture features. ; Spatial attention weight map The expression is: ; in: This indicates a kernel size of Convolutional layers, This indicates a splicing operation. This represents the average pooling operation along the channel dimension. This represents the max pooling operation along the channel dimension; Enhanced texture features The expression is: ; In S2-2-4, dynamic gating signal The expression is: ; in: This indicates a kernel size of Convolutional layers; Semantic-texture gated fusion feature map The expression is: ; in: This indicates element-wise addition; Final Enhanced Feature Map The expression is: ; The gating mechanism will enhance semantic features. The core information is retained and controlled via dynamic gating signals. Modulation of enhanced texture features The weights are ultimately determined through a... The convolutional layers yield the final enhanced feature map. .

[0010] Furthermore, in S2-3, the classification head includes: Global average pooling layer, fully connected layer, and Softmax activation function; Global average pooling layers are used to compress the spatial dimension of feature maps and extract global semantic information; Fully connected layers are used to output an N-dimensional vector. N-dimensional vector The expression is: ; The Softmax activation function is used to convert an N-dimensional vector into an N-dimensional vector. Convert to scene probability distribution vector ; Among them: the The probability of each category The calculation formula is: .

[0011] Further, S3 includes the following steps: S3-1, Regarding the multidimensional scattering spectrum vector Weighted integrals are performed to calculate the atmospheric extinction coefficient. Then, the basic visibility is calculated according to Koschmieder's law. ; S3-2, KAN Network Enhanced Residual Correction Module: The KAN network enhanced residual correction module corrects model biases to obtain low visibility corrected residuals. High visibility correction residuals ; S3-3, Output Fusion and Vector Construction: Obtaining the Basic Visibility from S3-1 Low visibility correction residuals in S3-2 High visibility correction residuals To generate prediction vectors ; Model training and loss function design for S3-4 and K-PIRC-Net: Constructing a composite loss function; In S3-1, the atmospheric extinction coefficient The calculation formula is: ; in: These represent the weighting coefficients for sensitive presets at different wavelengths and angles. Represents a multidimensional scattering spectral vector Intensity value for the corresponding dimension; Basic visibility The calculation formula is: ; in: Represents Koschmieder's constant; In S3-3, the prediction vector The calculation formula is: ; Low visibility forecast The calculation formula is: ; High visibility forecast The calculation formula is: ; In S3-4, the parameters of the visual classification network and the scattering spectrum inversion network are jointly optimized simultaneously, and the total loss function is... The expression is: ; in: Used to penalize visibility output results ground truth The differences between them The hyperparameter represents the balance between final accuracy and internal consistency. As a regularization term for retrieving the internal behavior of the network, an interval-based ranking loss is used. Penalty visibility output results ground truth Differences between The calculation formula is: ; In low-visibility scenarios, the loss function The expression is: ; In high visibility scenarios, the loss function The expression is: ; Final consistency loss The expression for dynamic selection based on the scenario is: .

[0012] Further, S3-2 includes the following steps: S3-2-1, Local Feature Extraction: Obtaining the Multidimensional Scattering Spectrum Vector in S1 And through a one-dimensional convolutional network, the multi-dimensional scattering spectral vector is transformed. Compression into spectral eigenvectors ; S3-2-2, Relationship Discovery: Obtaining the spectral feature vector from S3-2-1 and the spectral eigenvector Input into the KAN layer; S3-2-3, Dual-head output: The output of the KAN layer is sent to two independent linear layers as dual output heads to obtain the low visibility correction residual. High visibility correction residuals ; In S3-2-2, for a set of input features KAN layer output The expression is: ; in: Indicates the first A one-dimensional function that can be learned on each input edge.

[0013] Furthermore, in S4, the dynamic gating unit includes: Fully connected layer network; The fully connected layer network uses the scene probability distribution vector in S2. As input, learn and output a preliminary visibility prediction vector for S3. The fusion weight vector of each component fusion weight vector The expression is: ; in: The sum of all weights of the function is 1. Indicates a fully connected layer; fusion weight vector Compared with the preliminary visibility prediction vector Perform a weighted sum to obtain the final visibility value. The final output visibility value The expression is: .

[0014] Further, step S5 includes the following steps: S5-1 Intelligent identification of edge performance anomalies: Models deployed in edge computing units execute efficient intelligent strategies and trigger data upload when the model itself exhibits uncertainty or internal conflicts; S5-2, Generation and Upload of Problem Sample Information Package: If a sample is judged to be a problem sample, the edge computing unit will generate a structured problem sample data package and upload the structured problem sample data package to the cloud intelligent center. After receiving a problem sample data packet from one or more edge devices, the various functional modules within the S5-3 Cloud Intelligence Center work together. S5-4, Lightweight Model Update Generation and Distribution: After the model incremental optimization engine is adjusted, the adjusted model is handed over to the model library and deployment module for processing. Therefore...

[0015] Furthermore, in S5-1, the triggering conditions include: bimodal decision conflict, low confidence in visual classification, and severe fluctuations in the output result; in bimodal decision conflict, when the fusion weight vector generated in S4... The occurrence of a high-entropy state indicates a high degree of uncertainty in multiple preliminary predictions in S3. Therefore, a fusion weight vector is calculated. Shannon entropy To quantify this uncertainty, the Shannon entropy of the fused weight vector is used. An upload is triggered when the threshold is exceeded, and the Shannon entropy of the fused weight vector is used. The expression is: ; In low-confidence visual classification, when the scene probability vector in S2... The maximum confidence level is lower than the preset threshold Time (i.e.) This indicates that the visual classification network itself lacks confidence in its judgment of the current scene; in the case of drastic fluctuations in the output results, when the system's input signal changes smoothly within a short period of time, but the final output visibility value remains high. When violent, high-frequency oscillations that do not conform to physical laws occur, it indicates that the model has entered an unstable state. In S5-2, the information package includes: raw sensor data, intermediate model states, final output results, and device and environmental metadata. Among the raw sensor data, there are single-frame or multi-frame key image data at the trigger time and the synchronized multidimensional scattering spectrum vector. The intermediate states of the model include: scene probability distribution vector. Preliminary visibility prediction vector fusion weight vector The final output result is: the final visibility value. Device and environment metadata includes: device ID, GPS location, and timestamp. In S5-3, collaborative work performs the following operations: data aggregation and labeling, and incremental model adjustment. In data aggregation and labeling, firstly, the uploaded problem sample data package is sent to the data storage and management module and then merged into the problematic sample database. Subsequently, the system calls a high-precision benchmark data source to process the problem sample data package. In incremental model adjustment, problematic samples with high-quality labels, along with the benchmark model maintained in the cloud and isomorphic to the edge model, are sent to the incremental model optimization engine. In S5-4, the model library and deployment module is responsible for performing the following key tasks: model version management, update package generation, and secure deployment. In model version management, the internal model library archives old and new models, records performance indicators, and performs version control to ensure the traceability of model iterations. In update package generation, lightweight model update packages are generated. In secure deployment, the generated update packages are distributed to designated edge computing units through a secure network channel. Thus...

[0016] Compared with the prior art, the beneficial effects of the present invention are: This invention improves the accuracy of atmospheric visibility detection through inversion operations and enhances the interpretability of model behavior and adherence to physical laws by introducing physical priors. Through fusion operations, visual and scattering spectrum information are deeply and dynamically integrated within the model, fully utilizing the macroscopic scene understanding capabilities of vision and the microscopic physical detection capabilities of scattering spectra to overcome the limitations of single sensors and improve the accuracy of atmospheric visibility detection. Continuous optimization operations enable the entire device to self-diagnose detection times with high uncertainty and proactively report them to the cloud. The cloud then analyzes and optimizes the model using a more precise source and feeds the optimization capabilities back to edge devices, enabling continuous learning and self-optimization of the entire system, thereby ensuring the long-term accuracy of the entire device's detection. Attached Figure Description

[0017] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0018] Figure 1 This is a flowchart of the self-calibrating visibility measurement method based on visual-scattering spectrum dual-modal fusion of the present invention; Figure 2 This is a system diagram of the cloud-edge collaborative architecture of the present invention; Figure 3 This is a flowchart illustrating the dual-modal data processing and fusion within the edge nodes of the present invention. Figure 4 This is an architecture diagram of the scene-adaptive lightweight classification network of the present invention; Figure 5This is a diagram of the architecture of the KAN-enhanced physical prior-guided residual correction network of the present invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," and "circumferential," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, features defined with "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0021] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0022] like Figures 1 to 5 The diagram shows the preferred embodiment of the present invention. This embodiment of the self-calibrating visibility measurement method based on visual-scattering spectrum dual-modal fusion includes the following steps: S1. Dual-modal data synchronous acquisition and preprocessing: The image acquisition unit captures a frame of image, and the detector array synchronously samples to obtain the multidimensional scattering spectrum vector. And analyze the acquired image frame and multidimensional scattering spectrum vector. Preprocessing is required; S2. Vision-based scene classification: Obtain the scene images from S1 and utilize the Scene-Adaptive Lightweight Classification Network (SALC-Net) to output the scene probability distribution vector. ; S3. Visibility inversion based on scattering spectrum: Obtain the multidimensional scattering spectrum vector in S1. It also utilizes a KAN-enhanced Physics-Informed Residual Correction Network (K-PIRC-Net) to output multiple preliminary visibility prediction vectors. ; S4. Dual-modal dynamic gating fusion: Obtain the scene probability distribution vector in S2. Multiple preliminary visibility prediction vectors in S3 To calculate the final output visibility value. ; S5. Continuous optimization of the distributed model based on cloud-edge collaborative architecture: Obtain the visibility value of the final output in S4. The system is continuously optimized. Inversion operations improve the accuracy of atmospheric visibility detection, and the introduction of physical priors enhances the interpretability of model behavior and adherence to physical laws. Fusion operations enable deep and dynamic interaction between visual and scattering spectrum information within the model, fully utilizing the macroscopic scene understanding capabilities of vision and the microscopic physical detection capabilities of scattering spectra to overcome the limitations of a single sensor and improve the accuracy of atmospheric visibility detection. Continuous optimization allows the entire device to self-diagnose detection times with high uncertainty and proactively report them to the cloud. The cloud then analyzes and optimizes the model using a more precise source, feeding back the optimization capabilities to edge devices. This enables continuous learning and self-optimization of the entire system, ensuring the long-term accuracy of the entire device's detection.

[0023] In other words, this application guides and modulates the scattering inversion process in real time and dynamically through visual scene classification results. It can adaptively select the optimal fusion strategy according to the actual atmospheric conditions, so that the entire system can maintain high robustness and all-weather adaptability even when facing challenging scenarios such as mixed precipitation and non-uniform atmosphere that are difficult to handle by traditional technologies. By using a KAN-enhanced physical prior-guided residual correction network to combine the physical prior module with the data-driven residual correction module, the physical prior module provides strong theoretical constraints for the model output, ensuring the physical interpretability of the detection results and effectively avoiding predictions that violate physical common sense from purely data-driven models. Meanwhile, KAN's enhanced residual correction module can accurately compensate for the deviation between the simplified physical model and complex reality, thereby improving the detection accuracy to a new level while ensuring physical consistency. By constructing a cloud-edge collaborative remote update mechanism, the performance degradation problem of fixed models caused by factors such as sensor aging and environmental drift is solved. This mechanism establishes a closed-loop optimization feedback loop from the edge to the cloud by intelligently identifying detection events with uncertainties. This enables the entire system to have the ability to continuously learn and self-optimize, adapt to long-term changes, ensure the long-term consistency and reliability of detection data quality, and significantly reduce the frequency and cost of on-site manual maintenance.

[0024] For example, the image acquisition unit uses an industrial-grade CMOS or CCD camera with high dynamic range imaging capabilities to cope with complex lighting conditions such as day and night cycles, strong light, and weak light.

[0025] Specifically, the scattering spectrum measurement unit includes: a multi-wavelength light source (using at least two different wavelengths (e.g., 650nm in the visible light band and 850nm in the near-infrared band)) laser diodes as the light source (different wavelengths have different scattering sensitivities to aerosols of different particle sizes, providing a physical basis for distinguishing visibility reduction caused by different factors such as fog and haze), a multi-angle detector array (arranging highly sensitive photodetectors at multiple specific forward scattering angles (e.g., 35°, 42°, 50°), and polarization information acquisition (setting polarization beam splitting elements in the optical path to measure the depolarization ratio of the scattered light).

[0026] Specifically, multidimensional scattering spectral vector The expression is: ; in: Indicates the intensity of scattered light. Indicates wavelength. Indicates the scattering angle. Indicates the depolarization ratio; Multidimensional scattering spectrum vector It can precisely characterize the particle size distribution, phase state, and concentration information of suspended particles in the atmosphere.

[0027] Specifically, at the beginning of each measurement cycle, the module sends a synchronous hardware trigger signal to the image acquisition unit and the scattering spectrum measurement unit. Upon receiving the signal, the image acquisition unit captures a frame of image, while the multi-wavelength light source performs pulse emission and the detector array performs synchronous sampling.

[0028] Specifically, image preprocessing includes, but is not limited to: lens distortion correction, image denoising; and multidimensional scattering spectral vector analysis. Preprocessing includes, but is not limited to: dark current subtraction, ambient light compensation, and physical quantity calibration.

[0029] In this embodiment, S2 includes the following steps: S2-1: Obtain the scene image from S1 and use a lightweight asymmetric backbone network (SAC-Backbone) to output feature pyramids with different resolutions and semantic levels. ; S2-2, Obtain the feature pyramid from S2-1 The final enhanced feature map is obtained by utilizing the Semantic-Textural Decoupling and Enhancement Module (STDEM). ; S2-3, Obtain the final enhanced feature map from S2-2. It also utilizes a classification head to output a scene probability distribution vector. ; In S2-1, the lightweight asymmetric backbone network consists of multiple stages connected in series. Each stage is composed of N asymmetric residual blocks (ARBs) stacked together. The asymmetric residual block is the basic building block of the lightweight asymmetric backbone network. Internally, the asymmetric residual block employs a large convolutional kernel integration strategy. Convolution decomposed into a and convolution; For an input feature map Its output The calculation formula is: ; in: This indicates a modified linear unit activation function. Indicates the normalization layer. express Convolution operation, express Convolution operation; Downsampling is performed between each stage using convolutions with a stride of 2, gradually expanding the receptive field to extract higher-level semantic information, ultimately forming a feature pyramid containing rich spatial and semantic information. ; S2-2 includes the following steps: S2-2-1, Multi-scale Feature Preprocessing: Obtaining the Feature Pyramid from S2-1 And through bilinear interpolation of deep feature maps Upsampling is performed to make its spatial size match the feature map. To maintain consistency, then compare the upsampled feature map with the original feature map. The data is stitched together along the channel dimension to form a fusion feature map that aggregates multi-scale information. ; S2-2-2, Obtain the fusion feature map from S2-2-1 and the fused feature map Through a Convolutional layers to obtain semantic channel groups and texture channel groups ; S2-2-3, Obtain the semantic channel group in S2-2-2 and texture channel groups And perform enhancement branching to obtain enhanced semantic features. and enhanced texture features ; S2-2-4, Obtain the enhanced semantic features from S2-2-3 and enhanced texture features Through enhanced semantic features Generate dynamic gating signals The dynamic gating signal Enhanced texture features for adaptive adjustment The contribution of each feature map is then analyzed and fused to obtain the final enhanced feature map. ; In S2-2-1, feature maps are fused. The expression is: ; in: This indicates a splicing operation at the channel level. This indicates a bilinear interpolation upsampling operation; In S2-2-3, semantic channel groups Enter the channel attention module to calculate the channel attention weight vector. and the channel attention weight vector Apply to semantic channel group The above is used to obtain enhanced semantic features. ; Channel attention weight vector The expression is: ; in: This represents a layer consisting of two fully connected layers (with weights respectively). , A multilayer perceptron composed of ) This indicates a global average pooling operation. This represents the Sigmoid activation function; After enhancing semantic features The expression is: ; in: This represents a broadcast-based element-wise multiplication operation, utilizing a broadcast mechanism to distribute the channel attention weight vector. Applied to semantic channel groups At each spatial location; Texture Channel Group The spatial attention module then analyzes the importance of different spatial locations on the feature map to calculate the spatial attention weight map. And spatial attention weight map Apply to texture channel group Above, to obtain enhanced texture features. ; Spatial attention weight map The expression is: ; in: This indicates a kernel size of Convolutional layers, This indicates a splicing operation. This represents the average pooling operation along the channel dimension. This represents the max pooling operation along the channel dimension; Enhanced texture features The expression is: ; In S2-2-4, dynamic gating signal The expression is: ; in: This indicates a kernel size of Convolutional layers; Semantic-texture gated fusion feature map The expression is: ; in: This indicates element-wise addition; Final Enhanced Feature Map The expression is: ; The gating mechanism will enhance semantic features. The core information is retained and controlled via dynamic gating signals. Modulation of enhanced texture features The weights are ultimately determined through a... The convolutional layers yield the final enhanced feature map. ; In S2-3, the classification head includes: a Global Average Pooling Layer (GAP), a Fully Connected Layer (FC), and a Softmax activation function; the Global Average Pooling Layer is used for spatial dimension compression of the feature map and extraction of global semantic information; Fully connected layers are used to output an N-dimensional vector. N-dimensional vector The expression is: ; The Softmax activation function is used to convert an N-dimensional vector into an N-dimensional vector. Convert to scene probability distribution vector ; Among them: the The probability of each category The calculation formula is: .

[0030] Specifically, the lightweight asymmetric backbone network has low computational cost and efficiently extracts low- and mid-level visual features of images; the semantic-texture decoupling enhancement module combines the physical characteristics of meteorological scenes with the attention mechanism of deep learning to deeply decouple and specifically enhance the mixed visual features.

[0031] In this embodiment, S3 includes the following steps: S3-1, Regarding the multidimensional scattering spectrum vector Weighted integrals are performed to calculate the atmospheric extinction coefficient. Then, the basic visibility is calculated according to Koschmieder's law. ; S3-2, KAN Network Enhanced Residual Correction Module: The KAN network enhanced residual correction module corrects model biases to obtain low visibility corrected residuals. High visibility correction residuals ; S3-3, Output Fusion and Vector Construction: Obtaining the Basic Visibility from S3-1 Low visibility correction residuals in S3-2 High visibility correction residuals To generate prediction vectors ; Model training and loss function design for S3-4 and K-PIRC-Net: Constructing a composite loss function; In S3-1, the atmospheric extinction coefficient The calculation formula is: ; in: These represent the weighting coefficients for sensitive presets at different wavelengths and angles. Represents a multidimensional scattering spectral vector Intensity value for the corresponding dimension; Basic visibility The calculation formula is: ; in: Represents Koschmieder's constant; In S3-3, the prediction vector The calculation formula is: ; Low visibility forecast The calculation formula is: ; High visibility forecast The calculation formula is: ; S3-2 includes the following steps: S3-2-1, Local Feature Extraction: Obtaining the Multidimensional Scattering Spectrum Vector in S1 And through a one-dimensional convolutional network (1D-CNN), the multi-dimensional scattering spectral vector is processed. Compression into spectral eigenvectors (More compact, higher information density); S3-2-2, Relationship Discovery: Obtaining the spectral feature vector from S3-2-1 and the spectral eigenvector Input into the KAN layer; S3-2-3, Dual-head output: The outputs of the KAN layer (where activation functions parameterized by B-strip curves are placed on the network's connection edges) are fed into two independent linear layers as dual output heads to obtain low visibility correction residuals. High visibility correction residuals ; In S3-2-2, for a set of input features KAN layer output The expression is: ; in: Indicates the first A one-dimensional function that can be learned on each input edge; In S3-4, the parameters of the visual classification network (SALC-Net) and the scattering spectrum inversion network (K-PIRC-Net) are jointly optimized simultaneously, and the total loss function is... The expression is: ; in: Used to penalize visibility output results ground truth The differences between them The hyperparameter represents the balance between final accuracy and internal consistency. As a regularization term for retrieving the internal behavior of the network, an interval-based ranking loss is used. Penalty visibility output results ground truth Differences between The calculation formula is: ; In low visibility scenarios (i.e.) < , The desired visibility correction value is set to a preset threshold. Higher visibility correction value (closer to the true value), loss function The expression is: ; In high visibility scenarios (i.e.) > , The desired visibility correction value is set to a preset threshold. Higher visibility correction value (closer to the true value), loss function The expression is: ; Final consistency loss The expression for dynamic selection based on the scenario is: .

[0032] Specifically, the core function of the visual classification network is to efficiently extract image features and perform scene discrimination; the core function of the scattering spectrum inversion model is a parallel structure of physical priors and data-driven correction.

[0033] Specifically, the Kolmogorov-Arnold network (KAN) is introduced as its core engine to enhance the model's interpretability and learning efficiency; in S3-2-1, the role of the one-dimensional convolutional network is to preprocess the original spectral lines, capturing local shape features (e.g., the sharpness of the scattering spectrum, the width of the valleys, etc.).

[0034] Specifically, low visibility correction residuals Correction of model biases in low-visibility scenarios such as fog and dense haze, and correction of residuals in high-visibility scenarios. Correction is performed on model biases in high visibility scenarios such as light fog and clear skies.

[0035] Specifically, in S3-4, samples in the fuzzy intermediate zone (i.e. ), In this way, the model is strongly constrained only in scenarios where the physical prior is very clear, while in complex scenarios the learning is dominated by the final accuracy loss, which can greatly improve the stability of training and the final generalization ability of the model.

[0036] In this embodiment, in S4, the dynamic gating unit includes: a fully connected layer network; the fully connected layer network uses the scene probability distribution vector in S2. As input, learn and output a preliminary visibility prediction vector for S3. The fusion weight vector of each component fusion weight vector The expression is: ; in: The sum of all weights of the function is 1. Indicates a fully connected layer; fusion weight vector Compared with the preliminary visibility prediction vector Perform a weighted sum to obtain the final visibility value. The final output visibility value The expression is: .

[0037] Specifically, the final output visibility value Used to calculate the final accuracy loss This allows the gradient to propagate in reverse, enabling end-to-end joint optimization of the entire bimodal system.

[0038] In this embodiment, step S5 includes the following steps: S5-1 Intelligent identification of edge performance anomalies: Models deployed in edge computing units execute efficient intelligent strategies and trigger data upload when the model itself exhibits uncertainty or internal conflicts; S5-2, Generation and Upload of Problem Sample Information Package: If a sample is judged to be a problem sample, the edge computing unit will generate a structured problem sample data package and upload the structured problem sample data package to the cloud intelligent center. After receiving a problem sample data packet from one or more edge devices, the various functional modules within the S5-3 Cloud Intelligence Center work together. S5-4, Generation and Distribution of Lightweight Model Updates: After the incremental optimization engine for the model is adjusted, the adjusted model is handed over to the model library and deployment module for processing; In S5-1, the triggering conditions include: bimodal decision conflict, low confidence in visual classification, and severe fluctuations in the output result; in bimodal decision conflict, when the fusion weight vector generated in S4... The occurrence of a high-entropy state indicates a high degree of uncertainty in multiple preliminary predictions in S3. Therefore, a fusion weight vector is calculated. Shannon entropy To quantify this uncertainty, the Shannon entropy of the fused weight vector is used. An upload is triggered when the threshold is exceeded, and the Shannon entropy of the fused weight vector is used. The expression is: ; In low-confidence visual classification, when the scene probability vector in S2... The maximum confidence level is lower than the preset threshold Time (i.e.) This indicates that the visual classification network itself lacks confidence in its judgment of the current scene; in the case of drastic fluctuations in the output results, when the system's input signal changes smoothly within a short period of time, but the final output visibility value remains high. When violent, high-frequency oscillations that do not conform to physical laws occur, it indicates that the model has entered an unstable state. In S5-2, the information package includes: raw sensor data, intermediate model states, final output results, and device and environmental metadata. Among the raw sensor data, there are single-frame or multi-frame key image data at the trigger time and the synchronized multidimensional scattering spectrum vector. The intermediate states of the model include: scene probability distribution vector. Preliminary visibility prediction vector fusion weight vector The final output result is: the final visibility value. Device and environment metadata includes: device ID, GPS location, and timestamp. In S5-3, collaborative work performs the following operations: data aggregation and labeling, and incremental model adjustment. In data aggregation and labeling, firstly, the uploaded problem sample data package is sent to the data storage and management module and then merged into the problem sample database. Subsequently, the system calls a high-precision benchmark data source to process the problem sample data package. In incremental model adjustment, problem samples with high-quality labels, along with the benchmark model maintained in the cloud and isomorphic to the edge model, are sent to the incremental model optimization engine (the incremental model optimization engine uses these problem samples to incrementally adjust and train the benchmark model). In S5-4, the model library and deployment module is responsible for performing the following key tasks: model version management, update package generation, and secure deployment. In model version management, the internal model library archives old and new models, records performance metrics, and performs version control to ensure the traceability of model iteration. In update package generation, lightweight model update packages are generated. In secure deployment, the generated update packages are distributed to designated edge computing units through a secure network channel (the edge computing units periodically query the cloud for available model updates; once an update is detected, they automatically download and apply the update package when the system is idle, thus seamlessly completing self-evolution).

[0039] Specifically, through the remote parameter update mechanism, it is possible to assess the uncertainty of detection results, thereby enabling the proactive discovery and uploading of difficult samples at the edge, as well as the complete closed loop of incremental optimization and backhaul deployment of cloud models, giving the system the ability to adapt and evolve over a long period of time.

[0040] Specifically, by continuously optimizing a distributed model based on a cloud-edge collaborative architecture, the inherent defects of traditional fixed-parameter models in terms of performance degradation when faced with gradual environmental changes, seasonal changes, or rare weather phenomena that have never been seen before can be overcome.

[0041] Specifically, in S5-3, the high-precision data source includes: A) a human-computer interaction system that incorporates human experts for review and fine calibration, providing a high-precision ground truth visibility label for difficult samples; B) a high-precision "teacher model" deployed in the cloud, much larger in scale than edge models, and pre-trained with massive amounts of data, to perform offline analysis on difficult samples and generate high-quality pseudo-labels.

[0042] Specifically, in S5-4, the update package includes: A, parameter differential update (i.e., calculating the weight difference between the old and new baseline models, and only packaging and distributing this difference); B, knowledge distillation (i.e., using the optimized baseline model as the "teacher" to train a new "student model" of the same size as the edge on the cloud, and then packaging and distributing the parameters of this brand-new, better "student model").

[0043] In summary, this invention improves the accuracy of atmospheric visibility detection through inversion operations and enhances the interpretability of model behavior and adherence to physical laws by introducing physical priors. Through fusion operations, visual and scattering spectrum information are deeply and dynamically integrated within the model, fully utilizing the macroscopic scene understanding capabilities of vision and the microscopic physical detection capabilities of scattering spectra to overcome the limitations of single sensors and improve the accuracy of atmospheric visibility detection. Continuous optimization operations enable the entire device to self-diagnose detection times with high uncertainty and proactively report them to the cloud. The cloud then analyzes and optimizes the model using a more precise source and feeds the optimization capabilities back to edge devices, enabling continuous learning and self-optimization of the entire system, thereby ensuring the long-term accuracy of the entire device's detection.

[0044] The above description is based on the preferred embodiments of the present invention. Through the above description, those skilled in the art can make various changes and modifications without departing from the technical concept of the present invention. The technical scope of the present invention is not limited to the contents of the specification, but must be determined by the scope of the claims.

Claims

1. A self-calibrating visibility measurement method based on visual-scattering spectrum dual-modal fusion, characterized in that, Includes the following steps: S1. Dual-modal data synchronous acquisition and preprocessing: The image acquisition unit captures a frame of image, and the detector array synchronously samples to obtain the multidimensional scattering spectrum vector. And analyze the acquired image frame and multidimensional scattering spectrum vector. Preprocessing is required; S2. Vision-based scene classification: Obtain the scene images from S1 and utilize a lightweight scene classification network to output the scene probability distribution vector. ; S3. Visibility inversion based on scattering spectrum: Obtain the multidimensional scattering spectrum vector in S1. It also utilizes a KAN-enhanced physical prior to guide the residual correction network, outputting multiple preliminary visibility prediction vectors. ; S4. Dual-modal dynamic gating fusion: Obtain the scene probability distribution vector in S2. Multiple preliminary visibility prediction vectors in S3 To calculate the final output visibility value. ; S5. Continuous optimization of the distributed model based on cloud-edge collaborative architecture: Obtain the visibility value of the final output in S4. And continuously optimize it.

2. The self-calibrating visibility measurement method based on visual-scattering spectrum dual-modal fusion according to claim 1, characterized in that, S2 includes the following steps: S2-1: Obtain the scene images from S1, and use a lightweight asymmetric backbone network to output feature pyramids with different resolutions and semantic levels. ; S2-2, Obtain the feature pyramid from S2-1 And a semantic-texture decoupling enhancement module is used to obtain the final enhanced feature map. ; S2-3, Obtain the final enhanced feature map from S2-2. It also utilizes a classification head to output a scene probability distribution vector. .

3. The self-calibrating visibility measurement method based on visual-scattering spectrum dual-modal fusion according to claim 2, characterized in that, In S2-1, the lightweight asymmetric backbone network consists of multiple stages connected in series. Each stage is composed of N stacked asymmetric residual models. The asymmetric residual model is the basic building block of the lightweight asymmetric backbone network. The asymmetric residual model employs a large convolutional kernel integration strategy internally. Convolution decomposed into a and convolution; For an input feature map Its output The calculation formula is: ； in: This indicates a modified linear unit activation function. Indicates the normalization layer. express Convolution operation, express Convolution operation; Downsampling is performed between each stage using convolutions with a stride of 2, gradually expanding the receptive field to extract higher-level semantic information, ultimately forming a feature pyramid containing rich spatial and semantic information. .

4. The self-calibrating visibility measurement method based on visual-scattering spectrum dual-modal fusion according to claim 2, characterized in that, S2-2 includes the following steps: S2-2-1, Multi-scale Feature Preprocessing: Obtaining the Feature Pyramid from S2-1 And through bilinear interpolation of deep feature maps Upsampling is performed to make its spatial size match the feature map. To maintain consistency, then compare the upsampled feature map with the original feature map. The data is stitched together along the channel dimension to form a fusion feature map that aggregates multi-scale information. ; S2-2-2, Obtain the fusion feature map from S2-2-1 and the fused feature map Through a Convolutional layers to obtain semantic channel groups and texture channel groups ; S2-2-3, Obtain the semantic channel group in S2-2-2 and texture channel groups And perform enhancement branching to obtain enhanced semantic features. and enhanced texture features ; S2-2-4, Obtain the enhanced semantic features from S2-2-3 and enhanced texture features Through enhanced semantic features Generate dynamic gating signals The dynamic gating signal Enhanced texture features for adaptive adjustment The contribution of each feature map is then analyzed and fused to obtain the final enhanced feature map. ; In S2-2-1, feature maps are fused. The expression is: ； in: This indicates a splicing operation at the channel level. This indicates a bilinear interpolation upsampling operation; In S2-2-3, semantic channel groups Enter the channel attention module to calculate the channel attention weight vector. and the channel attention weight vector Apply to semantic channel group The above is used to obtain enhanced semantic features. ; Channel attention weight vector The expression is: ； in: This represents a layer consisting of two fully connected layers (with weights respectively). , A multilayer perceptron composed of ) This indicates a global average pooling operation. This represents the Sigmoid activation function; After enhancing semantic features The expression is: ； in: This represents a broadcast-based element-wise multiplication operation, utilizing a broadcast mechanism to distribute the channel attention weight vector. Applied to semantic channel groups At each spatial location; Texture Channel Group The spatial attention module then analyzes the importance of different spatial locations on the feature map to calculate the spatial attention weight map. And spatial attention weight map Apply to texture channel group Above, to obtain enhanced texture features. ; Spatial attention weight map The expression is: ； in: This indicates a kernel size of Convolutional layers, This indicates a splicing operation. This represents the average pooling operation along the channel dimension. This represents the max pooling operation along the channel dimension; Enhanced texture features The expression is: ； In S2-2-4, dynamic gating signal The expression is: ； in: This indicates a kernel size of Convolutional layers; Semantic-texture gated fusion feature map The expression is: ； in: This indicates element-wise addition; Final Enhanced Feature Map The expression is: ； The gating mechanism will enhance semantic features. The core information is retained and controlled via dynamic gating signals. Modulation of enhanced texture features The weights are ultimately determined through a... The convolutional layers yield the final enhanced feature map. .

5. The self-calibrating visibility measurement method based on visual-scattering spectrum dual-modal fusion according to claim 2, characterized in that, In S2-3, the classification header includes: Global average pooling layer, fully connected layer, and Softmax activation function; Global average pooling layers are used to compress the spatial dimension of feature maps and extract global semantic information; Fully connected layers are used to output an N-dimensional vector. N-dimensional vector The expression is: ； The Softmax activation function is used to convert an N-dimensional vector into an N-dimensional vector. Convert to scene probability distribution vector ; Among them: the The probability of each category The calculation formula is: 。 6. The self-calibrating visibility measurement method based on visual-scattering spectrum dual-modal fusion according to claim 1, characterized in that, S3 includes the following steps: S3-1, Regarding the multidimensional scattering spectrum vector Weighted integrals are performed to calculate the atmospheric extinction coefficient. Then, the basic visibility is calculated according to Koschmieder's law. ; S3-2, KAN Network Enhanced Residual Correction Module: The KAN network enhanced residual correction module corrects model biases to obtain low visibility corrected residuals. High visibility correction residuals ; S3-3, Output Fusion and Vector Construction: Obtaining the Basic Visibility from S3-1 Low visibility correction residuals in S3-2 High visibility correction residuals To generate prediction vectors ; Model training and loss function design for S3-4 and K-PIRC-Net: Constructing a composite loss function; In S3-1, the atmospheric extinction coefficient The calculation formula is: ； in: These represent the weighting coefficients for sensitive presets at different wavelengths and angles. Represents a multidimensional scattering spectral vector Intensity value for the corresponding dimension; Basic visibility The calculation formula is: ； in: Represents Koschmieder's constant; In S3-3, the prediction vector The calculation formula is: ； Low visibility forecast The calculation formula is: ； High visibility forecast The calculation formula is: ； In S3-4, the parameters of the visual classification network and the scattering spectrum inversion network are jointly optimized simultaneously, and the total loss function is... The expression is: ； in: Used to penalize visibility output results ground truth The differences between them The hyperparameter represents the balance between final accuracy and internal consistency. As a regularization term for retrieving the internal behavior of the network, an interval-based ranking loss is used. Penalty visibility output results ground truth Differences between The calculation formula is: ； In low-visibility scenarios, the loss function The expression is: ； In high visibility scenarios, the loss function The expression is: ； Final consistency loss The expression for dynamic selection based on the scenario is: 。 7. The self-calibrating visibility measurement method based on visual-scattering spectrum dual-modal fusion according to claim 1, characterized in that, S3-2 includes the following steps: S3-2-1, Local Feature Extraction: Obtaining the Multidimensional Scattering Spectrum Vector in S1 And through a one-dimensional convolutional network, the multi-dimensional scattering spectral vector is transformed. Compression into spectral eigenvectors ; S3-2-2, Relationship Discovery: Obtaining the spectral feature vector from S3-2-1 and the spectral eigenvector Input into the KAN layer; S3-2-3, Dual-head output: The output of the KAN layer is sent to two independent linear layers as dual output heads to obtain the low visibility correction residual. High visibility correction residuals ; In S3-2-2, for a set of input features KAN layer output The expression is: ； in: Indicates the first A one-dimensional function that can be learned on each input edge.

8. The self-calibrating visibility measurement method based on visual-scattering spectrum dual-modal fusion according to claim 1, characterized in that, In S4, the dynamic gating unit includes: Fully connected layer network; The fully connected layer network uses the scene probability distribution vector in S2. As input, learn and output a preliminary visibility prediction vector for S3. The fusion weight vector of each component fusion weight vector The expression is: ； in: The sum of all weights of the function is 1. Indicates a fully connected layer; fusion weight vector Compared with the preliminary visibility prediction vector Perform a weighted sum to obtain the final visibility value. The final output visibility value The expression is: 。 9. The self-calibrating visibility measurement method based on visual-scattering spectrum dual-modal fusion according to claim 1, characterized in that, Step S5 includes the following steps: S5-1 Intelligent identification of edge performance anomalies: Models deployed in edge computing units execute efficient intelligent strategies and trigger data upload when the model itself exhibits uncertainty or internal conflicts; S5-2, Generation and Upload of Problem Sample Information Package: If a sample is judged to be a problem sample, the edge computing unit will generate a structured problem sample data package and upload the structured problem sample data package to the cloud intelligent center. After receiving a problem sample data packet from one or more edge devices, the various functional modules within the S5-3 Cloud Intelligence Center work together. S5-4 Lightweight Model Update Generation and Distribution: After the model incremental optimization engine is adjusted, the adjusted model is handed over to the model library and deployment module for processing.

10. The self-calibrating visibility measurement method based on visual-scattering spectrum dual-modal fusion according to claim 9, characterized in that, In S5-1, the triggering conditions include: Bimodal decision conflict, low confidence in visual classification, and severe output jitter; In bimodal decision conflict, when the fusion weight vector generated in S4 The occurrence of a high-entropy state indicates a high degree of uncertainty in multiple preliminary predictions in S3. Therefore, a fusion weight vector is calculated. Shannon entropy To quantify this uncertainty, the Shannon entropy of the fused weight vector is used. An upload is triggered when the threshold is exceeded, and the Shannon entropy of the fused weight vector is used. The expression is: ； In low-confidence visual classification, when the scene probability vector in S2... The maximum confidence level is lower than the preset threshold Time (i.e.) This indicates that the visual classification network itself lacks confidence in its judgment of the current scene; In cases of severe output fluctuations, when the input signal changes smoothly for a short period, but the final output visibility value... When violent, high-frequency oscillations that do not conform to the laws of physics occur, it indicates that the model has entered an unstable state; In S5-2, the information packet includes: Raw sensor data, intermediate model states, final output results, and device and environmental metadata; In the raw sensing data, key image data of a single frame or multiple frames at the trigger time, and the synchronized multidimensional scattering spectrum vector. ; The intermediate states of the model include: Scene probability distribution vector Preliminary visibility prediction vector fusion weight vector ; The final output result is: the final visibility value. ; Device and environment metadata includes: Device ID, GPS location, timestamp; In S5-3, collaborative work performs the following operations: Data aggregation and annotation, incremental model adjustment; In the data aggregation and annotation process, firstly, the uploaded problem sample data package is sent to the data storage and management module and then imported into the difficult sample database; subsequently, the system calls a high-precision benchmark data source to process the problem sample data package. In the incremental model adjustment, difficult samples with high-quality labels, along with the baseline model maintained in the cloud and isomorphic to the edge model, are sent into the incremental model optimization engine. In S5-4, the model library and deployment module are responsible for performing the following key tasks: Model version management, update package generation, and secure deployment; In model version management, new and old versions of models are archived, performance metrics are recorded, and version control is implemented in the internal model library to ensure the traceability of model iterations. In the package generation process, a lightweight model package is generated. In a secure deployment, the generated update package is distributed to the designated edge computing unit via a secure network channel.

Images