Sun azimuth estimation method and system based on multi-channel feature enhancement and regional perception attention

Abstract

The invention relates to a solar azimuth estimation method and system based on multichannel feature enhancement and regional perception attention, and belongs to the technical field of intelligent navigation of a low-altitude economic unmanned system. In order to solve the problem that the precision of solar azimuth estimation based on a polarization image is reduced under the condition of complex cloud layer coverage, the invention provides a deep learning framework fusing multi-channel features and direction perception attention. The method comprises the following steps: firstly, constructing a three-channel composite input feature consisting of a polarization intensity graph, a self-adaptive threshold gradient graph and high-frequency residual information based on polarization light field information acquired by a polarization beam splitting focal plane camera; secondly, a ResNet backbone network of an embedded compression excitation mechanism is adopted, a direction perception polarization attention module is introduced, and adaptive fusion of multi-scale features is realized through brightness guidance, depth feature enhancement and gradient edge branches; thirdly, dynamic fusion of multi-branch features is achieved through the learnable Softmax weight, and a direction constraint mechanism is introduced into an output layer to explicitly optimize estimation results of the solar azimuth angle and the solar elevation angle; and finally, in combination with a solar triangulation module based on a physical model, high-precision solar azimuth inversion is realized. According to the invention, the sun direction information can be stably extracted in a cloud layer interference environment, the robustness and precision of sun direction estimation are obviously improved, and the method has the advantages of compact structure, strong generalization, suitability for autonomous navigation and positioning in a GPS-free environment and the like.

Description

Technical Field

[0001] This invention belongs to the field of intelligent navigation technology for low-altitude economic unmanned systems, specifically involving a method and system for estimating solar azimuth based on multi-channel feature enhancement and regional perception attention. Background Technology

[0002] Achieving continuous and accurate autonomous navigation for unmanned systems in complex environments is a current research challenge and key focus. Global Navigation Satellite Systems (GNSS) are prone to signal attenuation or interruption in indoor environments, canyons, dense forests, or environments subject to interference, while Inertial Navigation Systems (INS) suffer from error accumulation over time. Navigation methods relying on environmental features, such as visual SLAM, lack stability in dynamic scenes, weak textures, or drastic changes in lighting. Therefore, developing auxiliary or alternative navigation schemes that do not rely on external signals and are suitable for complex weather conditions is of great value. Polarized light navigation, as a biomimetic navigation technology, provides orientation and attitude references for a vehicle by detecting stable polarization distribution patterns formed by sunlight scattering through the atmosphere. It has advantages such as being passive, having no accumulated errors, and being resistant to interference. Existing polarized light orientation methods mainly rely on point sensors or imaging polarization cameras to acquire polarization information, and then calculate the angle of polarization (AoP) and degree of polarization (DoP) to invert the sun's azimuth. However, in practical applications, especially under adverse weather conditions such as cloudy or hazy conditions, cloud cover and aerosol scattering can significantly reduce the degree of sky polarization, disrupt the symmetry and consistency of polarization patterns, and cause a sharp decline in the performance of solar azimuth estimation methods based on traditional physical models or extreme value search.

[0003] In recent years, some studies have attempted to introduce deep learning to improve the robustness of polarization navigation. For example, some works have used network structures such as SE-ResNet to fuse AoP, DoP, and light intensity information for solar vector estimation, but their generalization ability is limited when faced with complex degradation caused by clouds in the real sky. Other studies have used deep learning to enhance the quality of polarization images to improve heading estimation accuracy, but most of these studies focus on azimuth estimation and fail to fully utilize imaging polarization information to achieve high-precision joint estimation of the entire solar vector (azimuth and elevation). In addition, existing methods are insufficient in distinguishing weak effective signals from complex noise and edge features in polarization images under cloudy conditions at the feature extraction level, and lack dedicated network structure designs for polarization characteristics and direction awareness, which limits their navigation accuracy and application scope under complex weather conditions.

[0004] Therefore, there is an urgent need for a high-precision solar azimuth estimation method that can effectively cope with complex sky conditions such as cloudy skies, make full use of multi-dimensional information from polarization images, and perform intelligent feature enhancement and selection, so as to improve the navigation reliability of unmanned systems in environments without reliable GNSS signals. Summary of the Invention

[0005] To address the issue of decreased accuracy in solar azimuth estimation based on polarization images under complex cloud cover conditions, a deep learning framework fusing multi-channel features and direction-aware attention is proposed. First, based on polarization field information acquired by a polarization-splitting focal plane camera, a three-channel composite input feature is constructed, consisting of a polarization intensity map, an adaptive threshold gradient map, and high-frequency residual information. Second, a ResNet backbone network with an embedded compression excitation mechanism is employed, and a direction-aware polarization attention module is introduced to achieve adaptive fusion of multi-scale features through brightness guidance, depth feature enhancement, and gradient edge branches. Third, learnable Softmax weights are used to dynamically fuse multi-branch features, and a direction constraint mechanism is introduced in the output layer to explicitly optimize the estimation results of the solar azimuth and elevation angles.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] A method and system for estimating the sun's azimuth based on multi-channel feature enhancement and region-aware attention includes the following steps:

[0008] The first step is to establish a three-channel polarization image feature construction module, including the original brightness channel, the brightness contrast enhancement channel, and the high-frequency residual edge channel;

[0009] The second step involves using ResNet-50 (Residual Network-50) as the backbone network for the modules in the first step, and establishing a compression-excitation-enhanced residual backbone network architecture.

[0010] The third step is to design a solar region perception and attention module based on the second step.

[0011] The fourth step is to design the output regression layer and continuous angle prediction module based on the third step.

[0012] The fifth step is to design the loss function.

[0013] Furthermore, in the first step,

[0014] The original polarization-degree-based polarization image is expanded into a three-channel tensor to encode orientation cues. Input image (H, W represent height and width), containing: (1) a normalized original polarization image, (2) an adaptive threshold gradient channel to highlight brightness and polarization variations, and (3) a high-frequency residual edge channel to preserve small-scale polarization variations by suppressing low-frequency illumination.

[0015] The adaptive threshold gradient channel is calculated by combining the Gaussian adaptive threshold with the Sobel magnitude edge response:

[0016]

[0017] in, This represents the intensity value of the input image at pixel coordinates (u, v). This represents the intensity value after the adaptive threshold gradient operation. This represents the gradient magnitude of the image calculated based on the Sobel operator; This indicates the fusion of gradient information; This indicates that it provides block-based brightness and contrast enhancement and binarizes the image; This indicates a normalization operation.

[0018] The channel uses Gaussian subtraction to remove low-frequency illumination, thereby enhancing directional cues at a fine scale.

[0019]

[0020] in Indicates the relationship with the Gaussian kernel convolution, It is the standard deviation. It is a scaling factor. This process preserves edge variations caused by polarization, making these variations detectable even if clouds suppress the global polarization degree. This represents the intensity value after high-frequency residual edge channel operation.

[0021] Furthermore, in the second step, during the initial feature extraction stage, the input polarization image features are... Convolution (Conv1), batch normalization (BN), and rectified linear activation (ReLU) operations are performed sequentially to extract basic light intensity and polarization mode features; subsequently, an early orientation-aware polarization attention module is introduced. This is used to enhance the spatial response of polarization modes; after max pooling, early enhancement features are obtained. Its structural relationship is expressed as:

[0022]

[0023] in: Represents the early feature tensor of the input image; Conv1 represents the first convolutional operation, used to extract local brightness and polarization pattern distribution; BN represents the batch normalization operation, used to stabilize network training and accelerate convergence; ReLU represents the rectified linear unit, used to introduce non-linear feature representation capabilities; represents the early direction-aware polarization attention module, which enhances polarization direction features through direction-selective weight calculation; MaxPool represents the max pooling operation, used for downsampling and preserving salient feature regions.

[0024] After the aforementioned early feature enhancement, the feature map The deep polarization feature extraction is performed in the backbone network. The backbone network adopts a residual structure (ResNet-50) and embeds a mid-stage polarization attention module in the intermediate stage (after Layer 1). This is to further enhance the hierarchical response of the polarization structure.

[0025] Its structural relationship is expressed as follows:

[0026]

[0027] in: This represents the intermediate feature tensor input to the residual network; Layer1–Layer4 represent the four layer modules in the residual network. This represents the mid-range orientation-aware polarization attention module, which is used to selectively enhance polarization orientation features and suppress noise interference within the mid-level feature space.

[0028] A channel attention module (SE, Squeeze-and-Excitation) is further embedded in each residual block from Layer 1 to Layer 4 to achieve adaptive recalibration of channel-level features. This is represented as follows:

[0029]

[0030] in: Represents the input feature tensor; This represents a global average pooling operation, used to extract global statistical features for each channel; and These represent the weight matrices of the two fully connected layers, respectively. Represents the ReLU activation function; This represents the Sigmoid activation function; This represents the multiplication operation (Hadamard product). This represents the output feature tensor after channel recalibration.

[0031] Furthermore, in the third step, a solar region sensing attention module is designed. This module consists of a brightness guidance branch, an orientation recalibration branch, and a polarization gradient sensing branch, which is used to achieve joint enhancement of polarization features and spatial structure features.

[0032] The input feature tensor is denoted as . C represents the number of channels, and H and W represent the feature map height and width, respectively; the outputs of the three branches are denoted as... , and Its core calculation form is as follows:

[0033]

[0034] in, This represents a brightness-guided channel weighting function used to enhance global illumination response; This represents a direction-sensitive channel recalibration function used to improve the angular resolution of local features; This represents the polarization gradient sensing function, used to capture spatial variations in the polarization angle.

[0035] The final output of the module is denoted as

[0036]

[0037] in, This represents a multi-branch fusion function used to synthesize different feature responses to generate a region-aware enhanced feature map, thereby improving the accuracy and stability of solar azimuth estimation.

[0038] Furthermore, in the fourth step, an output regression layer and a continuous angle prediction module are designed to extract global orientation information from the fused high-dimensional feature map and output continuous estimates of the solar azimuth and elevation angles; the processing includes the following steps:

[0039] (1) Feature maps after region-aware attention enhancement Perform global average pooling (GAP) to compress the spatial feature distribution and generate a compact global feature representation:

[0040]

[0041] Where d is the number of channels.

[0042] (2) The global feature vector The input is fed into a nonlinear regression network consisting of multiple fully connected layers to achieve joint continuous prediction of azimuth and elevation angles:

[0043]

[0044] in, This represents the activation function. , , The weight matrix is ​​a learnable matrix. , These represent the estimated values ​​of the solar azimuth and elevation angles, respectively.

[0045] In the fifth step, a continuous angle regression loss function is designed for the network training phase to minimize the deviation between the predicted output and the true sun position.

[0046] The advantages of this invention compared to the prior art are:

[0047] This invention introduces a multi-branch feature enhancement module that incorporates brightness guidance, orientation recalibration, and polarization gradient perception. This module simultaneously captures global brightness information, local directional features, and polarization angle gradient changes in polarized images, thereby fully exploiting relevant features of the solar region and improving the accuracy of solar azimuth extraction. The invention employs a region-aware attention module to selectively enhance the polarization features of key regions, helping to suppress cloud and illumination interference, improve response to the solar region, and achieve robust azimuth estimation. By combining polarization information, brightness, and local gradient features, this invention maintains high-precision solar azimuth estimation even under cloudy, illumination-varying, and complex sky background conditions, significantly outperforming traditional methods based on single brightness or color features. The proposed multi-channel feature enhancement and region-aware attention framework features a modular design, allowing for easy integration into different UAVs or ground navigation systems. It is also compatible with other multimodal sensor data fusion, improving the overall positioning robustness of the system. Attached Figure Description

[0048] Figure 1 This is a flowchart illustrating the design of a solar azimuth estimation method based on multi-channel feature enhancement and region-aware attention according to the present invention. Detailed Implementation

[0049] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

[0050] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0051] Reference Figure 1 A method and system for estimating the sun's azimuth based on multi-channel feature enhancement and region-aware attention includes the following steps:

[0052] The first step is to establish a three-channel polarization image feature construction module, including the original brightness channel, the brightness contrast enhancement channel, and the high-frequency residual edge channel;

[0053] The second step involves using ResNet-50 (Residual Network-50) as the backbone network for the modules in the first step, and establishing a compression-excitation-enhanced residual backbone network architecture.

[0054] The third step is to design a solar region perception and attention module based on the second step.

[0055] The fourth step is to design the output regression layer and continuous angle prediction module based on the third step.

[0056] The fifth step is to design the loss function.

[0057] Furthermore, in the first step,

[0058] The original polarization-degree-based polarization image is expanded into a three-channel tensor to encode orientation cues. Input image (H, W represent height and width), containing: (1) a normalized original polarization image, (2) an adaptive threshold gradient channel to highlight brightness and polarization variations, and (3) a high-frequency residual edge channel to preserve small-scale polarization variations by suppressing low-frequency illumination.

[0059] The adaptive threshold gradient channel is calculated by combining the Gaussian adaptive threshold with the Sobel magnitude edge response:

[0060]

[0061] in, This represents the intensity value of the input image at pixel coordinates (u, v). This represents the intensity value after the adaptive threshold gradient operation. This represents the gradient magnitude of the image calculated based on the Sobel operator; This indicates the fusion of gradient information; This indicates that it provides block-based brightness and contrast enhancement and binarizes the image; This indicates a normalization operation.

[0062] The channel uses Gaussian subtraction to remove low-frequency illumination, thereby enhancing directional cues at a fine scale.

[0063]

[0064] in Indicates the relationship with the Gaussian kernel convolution, It is the standard deviation. It is a scaling factor. This process preserves edge variations caused by polarization, making these variations detectable even if clouds suppress the global polarization degree. This represents the intensity value after high-frequency residual edge channel operation.

[0065] Furthermore, in the second step, during the initial feature extraction stage, the input polarization image features are... Convolution (Conv1), batch normalization (BN), and rectified linear activation (ReLU) operations are performed sequentially to extract basic light intensity and polarization mode features; subsequently, an early orientation-aware polarization attention module is introduced. This is used to enhance the spatial response of polarization modes; after max pooling, early enhancement features are obtained. Its structural relationship is expressed as:

[0066]

[0067] in: Represents the early feature tensor of the input image; Conv1 represents the first convolutional operation, used to extract local brightness and polarization pattern distribution; BN represents the batch normalization operation, used to stabilize network training and accelerate convergence; ReLU represents the rectified linear unit, used to introduce non-linear feature representation capabilities; represents the early direction-aware polarization attention module, which enhances polarization direction features through direction-selective weight calculation; MaxPool represents the max pooling operation, used for downsampling and preserving salient feature regions.

[0068] After the aforementioned early feature enhancement, the feature map The deep polarization feature extraction is performed in the backbone network. The backbone network adopts a residual structure (ResNet-50) and embeds a mid-stage polarization attention module in the intermediate stage (after Layer 1). This is to further enhance the hierarchical response of the polarization structure.

[0069] Its structural relationship is expressed as follows:

[0070]

[0071] in: This represents the intermediate feature tensor input to the residual network; Layer1–Layer4 represent the four layer modules in the residual network. This represents the mid-range orientation-aware polarization attention module, which is used to selectively enhance polarization orientation features and suppress noise interference within the mid-level feature space.

[0072] A channel attention module (SE, Squeeze-and-Excitation) is further embedded in each residual block from Layer 1 to Layer 4 to achieve adaptive recalibration of channel-level features. This is represented as follows:

[0073]

[0074] in: Represents the input feature tensor; This represents a global average pooling operation, used to extract global statistical features for each channel; and These represent the weight matrices of the two fully connected layers, respectively. Represents the ReLU activation function; This represents the Sigmoid activation function; This represents the multiplication operation (Hadamard product). This represents the output feature tensor after channel recalibration.

[0075] Furthermore, in the third step, a solar region sensing attention module is designed. This module consists of a brightness guidance branch, an orientation recalibration branch, and a polarization gradient sensing branch, which is used to achieve joint enhancement of polarization features and spatial structure features.

[0076] The input feature tensor is denoted as . C represents the number of channels, and H and W represent the feature map height and width, respectively; the outputs of the three branches are denoted as... , and Its core calculation form is as follows:

[0077]

[0078] in, This represents a brightness-guided channel weighting function used to enhance global illumination response; This represents a direction-sensitive channel recalibration function used to improve the angular resolution of local features; This represents the polarization gradient sensing function, used to capture spatial variations in the polarization angle.

[0079] The final output of the module is denoted as

[0080]

[0081] in, This represents a multi-branch fusion function used to synthesize different feature responses to generate a region-aware enhanced feature map, thereby improving the accuracy and stability of solar azimuth estimation.

[0082] Furthermore, in the fourth step, an output regression layer and a continuous angle prediction module are designed to extract global orientation information from the fused high-dimensional feature map and output continuous estimates of the solar azimuth and elevation angles; the processing includes the following steps:

[0083] (1) Feature maps after region-aware attention enhancement Perform global average pooling (GAP) to compress the spatial feature distribution and generate a compact global feature representation:

[0084]

[0085] Where d is the number of channels.

[0086] (2) The global feature vector The input is fed into a nonlinear regression network consisting of multiple fully connected layers to achieve joint continuous prediction of azimuth and elevation angles:

[0087]

[0088] in, This represents the activation function. , , The weight matrix is ​​a learnable matrix. , These represent the estimated values ​​of the solar azimuth and elevation angles, respectively.

[0089] In the fifth step, a continuous angle regression loss function is designed for the network training phase to minimize the deviation between the predicted output and the true sun position.

[0090] The advantages of this invention compared to the prior art are:

[0091] This invention introduces a multi-branch feature enhancement module that incorporates brightness guidance, orientation recalibration, and polarization gradient perception. This module simultaneously captures global brightness information, local directional features, and polarization angle gradient changes in polarized images, thereby fully exploiting relevant features of the solar region and improving the accuracy of solar azimuth extraction. The invention employs a region-aware attention module to selectively enhance the polarization features of key regions, helping to suppress cloud and illumination interference, improve response to the solar region, and achieve robust azimuth estimation. By combining polarization information, brightness, and local gradient features, this invention maintains high-precision solar azimuth estimation even under cloudy, illumination-varying, and complex sky background conditions, significantly outperforming traditional methods based on single brightness or color features. The proposed multi-channel feature enhancement and region-aware attention framework features a modular design, allowing for easy integration into different UAVs or ground navigation systems. It is also compatible with other multimodal sensor data fusion, improving the overall positioning robustness of the system.

[0092] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method and system for estimating the sun's azimuth based on multi-channel feature enhancement and region-aware attention, characterized in that, Includes the following steps: The first step is to establish a three-channel polarization image feature construction module, including the original brightness channel, the brightness contrast enhancement channel, and the high-frequency residual edge channel; The second step involves using ResNet-50 (Residual Network-50) as the backbone network for the modules in the first step, and establishing a compression-excitation-enhanced residual backbone network architecture. The third step is to design a solar region perception and attention module based on the second step. The fourth step is to design the output regression layer and continuous angle prediction module based on the third step. The fifth step is to design the loss function.

2. The solar azimuth estimation method and system based on multi-channel feature enhancement and region-aware attention as described in claim 1, characterized in that: In the first step, the original polarization-degree-based polarization image is expanded into a three-channel tensor to encode orientation cues. Input image (H, W represent height and width), including: (1) the normalized original polarization image (2) An adaptive threshold gradient channel , used to highlight brightness and polarization variations, and (3) a high-frequency residual edge channel. Small-scale polarization changes are preserved by suppressing low-frequency illumination. The adaptive threshold gradient channel is calculated by combining the Gaussian adaptive threshold with the Sobel magnitude edge response: in, This represents the intensity value of the input image at pixel coordinates (u, v). This represents the intensity value after the adaptive threshold gradient operation. This represents the gradient magnitude of the image calculated based on the Sobel operator; This indicates the fusion of gradient information; This indicates that it provides block-based brightness and contrast enhancement and binarizes the image; This indicates a normalization operation. The channel uses Gaussian subtraction to remove low-frequency illumination, thereby enhancing directional cues at a fine scale. in Indicates the relationship with the Gaussian kernel convolution, It is the standard deviation. It is a scaling factor. This process preserves edge variations caused by polarization, making these variations detectable even if clouds suppress the global polarization degree. This represents the intensity value after high-frequency residual edge channel operation.

3. The solar azimuth estimation method and system based on multi-channel feature enhancement and region-aware attention as described in claim 2, characterized in that: The method includes a multi-layer polarization feature enhancement process based on the fusion of convolutional feature extraction and attention mechanism, and its structure is as follows: In the initial feature extraction stage, the input polarization image features are... Convolution (Conv1), batch normalization (BN), and linear rectified activation (ReLU) operations are performed sequentially to extract the basic light intensity and polarization mode features; Subsequently, an early orientation-aware polarization attention module was introduced. This is used to enhance the spatial response of polarization modes; after max pooling, early enhancement features are obtained. Its structural relationship is expressed as: in: Represents the early feature tensor of the input image; Conv1 represents the first convolutional operation, used to extract local brightness and polarization pattern distribution; BN represents the batch normalization operation, used to stabilize network training and accelerate convergence; ReLU represents the rectified linear unit, used to introduce non-linear feature representation capabilities. This represents the early direction-aware polarization attention module, which enhances polarization direction features through direction-selective weight calculation; MaxPool represents the max pooling operation, used for downsampling and preserving salient feature regions. After the aforementioned early feature enhancement, the feature map The deep polarization feature extraction is performed in the backbone network. The backbone network adopts a residual structure (ResNet-50), and a mid-stage polarization attention module is embedded in the intermediate stage (after Layer 1). This is to further enhance the hierarchical response of the polarization structure. Its structural relationship is expressed as follows: in: This represents the intermediate feature tensor input to the residual network; Layer1–Layer4 represent the four layer modules in the residual network. This represents the mid-range orientation-aware polarization attention module, which is used to selectively enhance polarization orientation features and suppress noise interference within the mid-level feature space. A channel attention module (SE, Squeeze-and-Excitation) is further embedded in each residual block from Layer 1 to Layer 4 to achieve adaptive recalibration of channel-level features. This is represented as follows: in: Represents the input feature tensor; This represents a global average pooling operation, used to extract global statistical features for each channel; and These represent the weight matrices of the two fully connected layers, respectively. Represents the ReLU activation function; This represents the Sigmoid activation function; Indicates the multiplication operation (Hadamard product); This represents the output feature tensor after channel recalibration. Through the synergistic effect of the above modules, the system enhances and dynamically weights the multi-level features of polarization images, enabling it to accurately extract solar azimuth-related features even in complex cloud environments, thereby improving the robustness and accuracy of solar azimuth estimation.

4. The solar azimuth estimation method and system based on multi-channel feature enhancement and region-aware attention as described in claim 3, characterized in that: In the third step, based on the multi-scale features extracted in the second step, a solar region sensing attention module is designed. This module consists of a brightness guidance branch, an orientation recalibration branch, and a polarization gradient sensing branch, which is used to achieve joint enhancement of polarization features and spatial structure features. The input feature tensor is denoted as . C represents the number of channels, and H and W represent the feature map height and width, respectively; the outputs of the three branches are denoted as... , and Its core calculation form is as follows: in, This represents a brightness-guided channel weighting function used to enhance global illumination response; This represents a direction-sensitive channel recalibration function used to improve the angular resolution of local features; This represents the polarization gradient sensing function, used to capture spatial variations in the polarization angle. The final output of the module is denoted as in, This represents a multi-branch fusion function used to synthesize different feature responses to generate a region-aware enhanced feature map, thereby improving the accuracy and stability of solar azimuth estimation.

5. The solar azimuth estimation method and system based on multi-channel feature enhancement and region-aware attention as described in claim 4, characterized in that: The fourth step involves designing an output regression layer and a continuous angle prediction module to extract global direction information from the fused high-dimensional feature map and output continuous estimation results of the solar azimuth and elevation angles. The processing procedure includes the following steps: (1) Feature maps after region-aware attention enhancement Perform global average pooling (GAP) to compress the spatial feature distribution and generate a compact global feature representation: Where d is the number of channels. (2) The global feature vector The input is fed into a nonlinear regression network consisting of multiple fully connected layers to achieve joint continuous prediction of azimuth and elevation angles: in, This represents the activation function. , , The weight matrix is ​​a learnable matrix. , These represent the estimated values ​​of the solar azimuth and elevation angles, respectively. Through the above structural design, the modeling of the global correlation of polarization characteristics and the continuous mapping of spatial angles are realized, thereby improving the accuracy and stability of azimuth estimation under complex lighting and cloud interference conditions.

6. The solar azimuth estimation method and system based on multi-channel feature enhancement and region-aware attention as described in claim 5, characterized in that: The fifth step involves designing a continuous angle regression loss function for the network training phase to minimize the deviation between the predicted output and the true sun position. Its definition is as follows: This represents the network prediction result for the i-th input sample. Let represent the corresponding true solar azimuth and elevation angles, and N be the total number of samples in the training batch. This loss design achieves joint regression optimization of azimuth and elevation angles, enhancing the model's learning stability in continuous angular space and its adaptability to complex changes in illumination, thereby effectively improving the accuracy and convergence efficiency of solar azimuth estimation.

Images