Physical law and structure perception weakly supervised cloud region label generation method and device, equipment and medium

Abstract

The embodiment of the application provides a cloud region label generation method, device and equipment under physical law and structure perception weak supervision, and a medium, which can be applied to the technical field of remote sensing information, and the method comprises the following steps: extracting activation response of a to-be-detected image from multiple different spatial perspectives to obtain activation response under each perspective; applying consistency constraint to the activation response under each perspective to generate a spatially enhanced activation map; adaptively screening spatial priori according to the spatially enhanced activation map to generate a pseudo label, wherein the spatial priori is constructed by using physical attribute information of a cloud target; applying structure consistency regularization to the spatially enhanced activation map at the spatial sub-region level of the image, and constraining the pseudo label as a supervision signal to obtain an optimized activation map; and outputting the optimized activation map as a cloud region pseudo label under image-level label supervision.

Description

Technical Field

[0001] This invention relates to the field of remote sensing information technology, and more specifically to a method, apparatus, device, and medium for generating cloud area labels under weak supervision of physical laws and structural perception. Background Technology

[0002] Optical remote sensing imagery boasts advantages such as wide coverage and high spatial resolution, making it widely used in resource surveys, meteorological monitoring, environmental governance, and disaster assessment. However, in actual imaging processes, factors such as atmospheric scattering and cloud cover often result in remote sensing images containing cloud regions of varying scales, thicknesses, and shapes. The presence of clouds obstructs surface information and causes radiation interference, significantly reducing the effective utilization rate of remote sensing images. Therefore, cloud detection and cloud mask generation have become key technical steps in the preprocessing workflow of optical remote sensing images.

[0003] Existing cloud detection methods mainly include spectral thresholding-based methods and deep learning-based methods. Spectral thresholding-based methods utilize the high reflectivity of clouds in the visible or near-infrared bands, distinguishing between cloud and non-cloud areas by setting brightness thresholds or band ratios. These methods are simple to implement but are sensitive to imaging conditions and ground background, and have limited generalization ability. Deep learning-based methods achieve cloud region detection through multi-layer feature extraction networks or semantic segmentation models, offering higher detection accuracy but relying on a large number of pixel-level labeled samples, resulting in high labeling costs. Furthermore, their adaptability to different sensors and imaging conditions remains limited.

[0004] To reduce the cost of manual annotation, weakly supervised cloud detection methods based on activation maps have emerged in recent years, generating pixel-level localization results through image-level labels. However, existing methods still have the following technical shortcomings: First, activation maps typically only produce strong responses to the most discriminative local cloud regions, resulting in incomplete activation of cloud regions and severe fragmentation of spatial structure. Second, existing methods mainly rely on data-driven feature responses and do not fully consider the radiation characteristics of clouds during remote sensing imaging, making them prone to false activations in complex backgrounds such as snowfields, bright-colored bare land, and water bodies. Third, clouds exhibit significant multi-scale characteristics in remote sensing images, while existing methods are highly scale-sensitive, resulting in insufficient detection stability at different resolutions or cloud scales. Fourth, existing methods neglect the spatial continuity and regional consistency of clouds, and the generated cloud mask morphology does not conform to the multi-scale structural characteristics of real clouds.

[0005] Therefore, how to generate spatially complete, clearly defined, physically reasonable, and scale-adaptable pseudo-labels for cloud regions under weak supervision using only image-level labels is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0006] In view of the above problems, the present invention provides a method, apparatus, device and medium for generating cloud region labels under weak supervision of physical laws and structural perception.

[0007] According to a first aspect of the present invention, a method for generating cloud region labels under weak supervision based on physical laws and structure awareness is provided, comprising: extracting activation responses from an image to be detected from multiple different spatial perspectives to obtain activation responses under each perspective; applying consistency constraints to the activation responses under each perspective to generate a spatially enhanced activation map; adaptively filtering spatial priors based on the spatially enhanced activation map to generate pseudo-labels, wherein the spatial priors are constructed using physical attribute information of cloud targets; applying structural consistency regularization to the spatially enhanced activation map at the spatial sub-region level of the image, and using the pseudo-labels as supervision signals for constraint to obtain an optimized activation map; and outputting the optimized activation map as a cloud region pseudo-label under image-level label supervision.

[0008] According to an embodiment of the present invention, the step of extracting activation responses from multiple different spatial perspectives of the image to be detected to obtain activation responses under each perspective includes: generating corresponding class activation maps by taking the image to be detected and its horizontally flipped image as input through a global branch; performing an inverse transform on the class activation maps and then fusing them to obtain activation responses under the global perspective; dividing the image to be detected into multiple non-overlapping sub-blocks through a local branch; generating class activation maps for each sub-block and then splicing them according to the original spatial positions of the sub-blocks to obtain activation responses under the local perspective.

[0009] According to an embodiment of the present invention, the step of applying consistency constraints to the activation responses under each viewpoint to generate a spatially enhanced activation map includes: determining the pixel-level difference between the activation responses under the global viewpoint and the activation responses under the local viewpoint; using the difference as a consistency loss function to constrain the activation responses under the global viewpoint and the activation responses under the local viewpoint to maintain consistency, thereby generating a spatially enhanced activation map.

[0010] According to an embodiment of the present invention, the method further includes: determining the dark channel transmittance based on an atmospheric scattering model, wherein the dark channel transmittance is used to highlight thick cloud regions; constructing a fog-optimized transformation index based on a linear combination of blue and red bands, wherein the fog-optimized transformation index is used to enhance the sensitivity of thin cloud regions; and normalizing and fusing the dark channel transmittance and the fog-optimized transformation index to obtain a unified saliency map as the spatial prior.

[0011] According to an embodiment of the present invention, the step of adaptively filtering spatial priors based on the spatially enhanced activation map to generate pseudo-labels includes: performing unsupervised clustering on the saliency map to generate multiple hierarchical candidate pseudo-labels; determining the degree of overlap between each candidate pseudo-label and the binary result of the spatially enhanced activation map; and selecting the candidate pseudo-label with the largest degree of overlap as the pseudo-label.

[0012] According to an embodiment of the present invention, the step of applying structural consistency regularization to the spatially enhanced activation map at the spatial sub-region level of the image and constraining it with the pseudo-label as a supervision signal to obtain an optimized activation map includes: dividing the image to be detected into multiple superpixel regions with local color and texture consistency using a superpixel segmentation algorithm; calculating the percentage of pixels belonging to the cloud region for each superpixel region based on the binary result of the spatially enhanced activation map; determining the region-level pseudo-label of the superpixel region as the cloud category when the percentage exceeds a preset threshold, otherwise determining it as the non-cloud category; calculating the average predicted response value of the spatially enhanced activation map in each superpixel region; and introducing a region-level binary cross-entropy loss function with the area of ​​the superpixel region as the weight to constrain the average predicted response of each superpixel region to be consistent with the region-level pseudo-label of the superpixel region.

[0013] According to an embodiment of the present invention, before outputting the optimized activation map as a pseudo-label for the cloud region under image-level label supervision, the method further includes: calculating a classification loss using the image-level label, wherein the classification loss adopts a binary cross-entropy loss function; calculating a pixel-level supervision loss using the pseudo-label, wherein the pixel-level supervision loss is used to constrain the spatially augmented activation map to be consistent with the pseudo-label; and training to generate the optimized activation map by jointly optimizing the classification loss and the pixel-level supervision loss.

[0014] A second aspect of the present invention provides a cloud region tag generation apparatus under weak supervision of physical laws and structure awareness, comprising:

[0015] The extraction module is used to extract activation responses from the image to be detected from multiple different spatial perspectives, and obtain the activation responses under each perspective.

[0016] The graph generation module is used to apply consistency constraints to the activation responses under each viewpoint to generate a spatially enhanced activation graph.

[0017] The label generation module is used to adaptively filter the spatial prior based on the spatially enhanced activation graph and generate pseudo-labels. The spatial prior is constructed using the physical attribute information of the cloud target.

[0018] An optimization module is used to apply structural consistency regularization to the spatially enhanced activation map at the spatial sub-region level of the image, and to constrain it using the pseudo-label as a supervision signal to obtain an optimized activation map.

[0019] The output module is used to output the optimized activation map as a pseudo-label for the cloud region under image-level label supervision.

[0020] According to a third aspect of the present invention, an electronic device is provided, comprising: one or more computing units; and a memory for storing one or more computer programs, wherein the one or more computing units execute the one or more computer programs to implement the steps of the method according to the first aspect.

[0021] According to a fourth aspect of the present invention, a computer-readable storage medium is provided having a computer program or instructions stored thereon, which, when executed by a computing unit, implement the steps of the method according to the first aspect.

[0022] According to embodiments of the present invention, by extracting activation responses from multiple different spatial perspectives and applying consistency constraints, the model can establish stable response correlations between global and local scales. This effectively overcomes the problems of incomplete cloud region activation, severe spatial structure fragmentation, and sensitivity to scale changes caused by traditional activation maps that only focus on local discriminative regions. The generated activation map is more continuous and complete in spatial distribution and can adapt to cloud structures of different scales and forms. On this basis, by constructing spatial priors using the physical attribute information of cloud targets and performing adaptive filtering, the physical rationality constraints of cloud radiation characteristics are introduced into the purely data-driven activation response, effectively suppressing false activation and missed detection problems in complex backgrounds such as snow, bright bare land, and water bodies, and significantly improving the semantic reliability of cloud region pseudo-labels. Furthermore, by applying structural consistency regularization at the spatial sub-region level of the image, pixel-level prediction is elevated to region-level decision-making, effectively reducing the fragmentation of activation results, noise interference, and internal holes, making the generated cloud mask more consistent with the multi-scale morphological characteristics of real clouds in terms of boundary integrity and spatial consistency. The three techniques described above work together under extremely weak supervision using only image-level labels to achieve high-precision and robust cloud detection output, significantly reducing the cost of manual pixel-level annotation. At the same time, they enhance the generalization ability of the method across sensors and imaging scenarios, providing a complete, reliable, and economical solution for solving the cloud detection problem in optical remote sensing image preprocessing. Attached Figure Description

[0023] The above-described features, other objects, and advantages of the present invention will become clearer from the following description of embodiments of the invention with reference to the accompanying drawings, in which:

[0024] Figure 1 The flowchart illustrates a cloud region label generation method under weak supervision of physical laws and structure perception according to an embodiment of the present invention.

[0025] Figure 2 The diagram illustrates a cloud region label generation method under weak supervision of physical laws and structure perception according to an embodiment of the present invention.

[0026] Figure 3 This schematic diagram illustrates the structural block diagram of a cloud region tag generation device under weak supervision of physical laws and structural perception according to an embodiment of the present invention.

[0027] Figure 4 The diagram illustrates an electronic device suitable for implementing a cloud region tag generation method under weak supervision of physical laws and structure perception according to an embodiment of the present invention. Detailed Implementation

[0028] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the invention. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of the invention for ease of explanation. However, it will be apparent that one or more embodiments may be practiced without these specific details. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0029] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0030] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0031] When using expressions such as "at least one of A, B and C", they should generally be interpreted in accordance with the meaning that is commonly understood by those skilled in the art (e.g., "a system having at least one of A, B and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B and C, etc.).

[0032] Figure 1The flowchart illustrates a cloud region label generation method under weak supervision of physical laws and structure perception according to an embodiment of the present invention.

[0033] like Figure 1 As shown, the cloud region label generation method under weak supervision of physical laws and structure perception in this embodiment includes operations S110 to S150.

[0034] In operation S110, activation response extraction is performed on the image to be detected from multiple different spatial perspectives to obtain the activation response under each perspective.

[0035] According to an embodiment of the present invention, the activation response refers to the distribution of location response intensity corresponding to the cloud category in the feature map output by a specific layer of a deep neural network. The higher the response intensity, the greater the probability that the location belongs to a cloud region. The activation response can be represented as a Class Activation Mapping (CAM), which is obtained by linearly weighting the convolutional features of the last layer of the classification network with class weights and then processing them with an activation function.

[0036] According to embodiments of the present invention, multiple different spatial perspectives refer to multiple observation angles formed by using different spatial sampling or transformation methods on the image to be detected. For example, one perspective may be a global view at the original resolution, while another perspective may be a magnified view of a local area after block processing; or, for example, one perspective may be the original image orientation, while another perspective may be the orientation of the horizontally flipped image. Spatial perspectives can be used to observe the same image content from different scales and different spatial ranges to obtain complementary cloud structure information.

[0037] In one embodiment of the present invention, the above operations are implemented through a dual-branch network: the global branch takes the image to be detected and its horizontally flipped image as input to generate a global-scale class activation map; the local branch divides the image to be detected into four non-overlapping 224×224 pixel sub-blocks, generates a class activation map corresponding to each sub-block, and then stitches them together according to their original spatial positions to form a full-size activation response map. Through this operation, the global branch can capture the overall distribution of large-scale clouds, while the local branch can retain the detailed information of small-scale cloud clusters and cloud edges, forming a complementary relationship.

[0038] In operation S120, a consistency constraint is applied to the activation response under each viewpoint to generate a spatially enhanced activation map.

[0039] According to embodiments of the present invention, a consistency constraint is a mathematical constraint that requires activation responses from different spatial perspectives to remain similar at the pixel level. The essence of this constraint is to force the model to produce consistent activation responses to the same cloud region under different observation conditions, thereby suppressing response fluctuations introduced by perspective changes or data augmentation and improving the stability of the activation map.

[0040] According to embodiments of the present invention, applying consistency constraints can be achieved by constructing a loss function. For example, the pixel-level L1 or L2 distance between the global view activation response and the local view activation response can be calculated, and this distance can be added as a consistency loss term to the overall training objective. By minimizing this loss, the model gradually makes the activation responses of different views tend to be consistent during training.

[0041] According to embodiments of the present invention, the spatially augmented activation map can refer to a novel activation map obtained after optimization with multi-view consistency constraints. Compared with the original CAM, the spatially augmented activation map has the following characteristics: First, it is insensitive to scale changes and can produce stable responses to both large-scale stratus clouds and small-scale cumulus clouds; second, it has more complete spatial coverage, avoiding the defect of traditional CAM that only produces responses in high-discriminative regions; third, it is robust to different data augmentation methods (such as flipping, rotating, and cropping).

[0042] In one embodiment of the present invention, the pixel difference between the global augmented response generated by the global branch and the whole image response generated by the local branch is used as the consistency loss function. The network parameters are updated through backpropagation. After several rounds of iterative training, the outputs of the two branches gradually converge to the same level. The activation map obtained at this time is the spatially augmented activation map.

[0043] In operation S130, spatial priors are adaptively filtered based on the spatially enhanced activation graph to generate pseudo-labels. The spatial priors are constructed using the physical attribute information of cloud targets.

[0044] According to embodiments of the present invention, spatial priors can refer to pixel-level confidence distributions pre-established based on the physical laws governing clouds in optical remote sensing imaging. This prior does not rely on data-driven learning but is calculated based on an interpretable physical model, providing stable spatial guidance in the early stages of model training.

[0045] According to embodiments of the present invention, the physical attribute information of cloud targets includes, but is not limited to, the following two categories: First, thick clouds exhibit high-intensity reflection in the visible light band and have an obstructing effect on surface signals, resulting in extremely low dark channel values ​​for objects below them; second, thin clouds exhibit specific reflectivity differences between the blue and red bands, with the blue band showing significantly higher reflectivity than the red band. Based on the above physical attributes, a spatial prior can be constructed in the following manner: the dark channel transmittance is derived based on an atmospheric scattering model, with lower transmittance indicating thicker and denser clouds; simultaneously, a Haze Optimized Transform (HOT) index is constructed, using the blue band minus half of the red band (i.e., HOT = Band_blue - 0.5 × Band_red), which is sensitive to thin clouds. After normalizing the dark channel transmittance and the HOT index respectively, they are added together to obtain a unified saliency map. High-value areas in this saliency map correspond to areas with a higher probability of cloud presence, thus constituting a spatial prior.

[0046] According to an embodiment of the present invention, adaptive screening refers to the process of automatically selecting the most suitable signal as a pseudo-label from multiple candidate supervision signals based on the dynamic changes of the current model state (i.e., the currently generated spatial augmentation activation map). This screening process is adaptive: in the early stages of training, when the model's discriminative ability is weak, it tends to select conservative pseudo-labels with higher confidence; as training progresses and the model's discriminative ability improves, it gradually selects pseudo-labels with more complete coverage.

[0047] According to an embodiment of the present invention, pseudo-labels refer to pixel-level classification labels automatically generated through the above-described filtering process in the absence of real pixel-level annotations, used to replace manual annotations and provide supervisory signals for the segmentation network. Each pixel of the pseudo-label is marked as either "cloud" or "non-cloud".

[0048] In one embodiment of the present invention, K-means unsupervised clustering can be performed on the saliency map first, dividing it into 5 saliency levels, sorted by the level center value from smallest to largest, and then accumulating to generate 5 candidate pseudo-labels: the first level contains only the cloud core region with the highest saliency, the fifth level contains all possible extended regions that are clouds. Then, the intersection-over-union (IoU) ratio between each candidate pseudo-label and the binary result of the current spatial augmented activation map is calculated, and the candidate pseudo-label with the largest IoU is selected as the final pseudo-label. For example, if the candidate pseudo-label of the third level has the highest IoU with the current activation map, then this level is selected as the pseudo-label for the current training round.

[0049] In operation S140, structural consistency regularization is applied to the spatially enhanced activation map at the spatial sub-region level of the image, and pseudo-labels are used as supervision signals for constraint to obtain the optimized activation map.

[0050] According to embodiments of the present invention, a spatial sub-region can refer to a number of local image patches with internal consistency into which the image to be detected is divided. Unlike fixed grid division, spatial sub-regions aim for semantic-level region division, that is, the pixels in each sub-region are similar in color, texture or spatial location, and usually correspond to a semantically continuous region, such as part of a continuous cloud or a complete ground feature.

[0051] According to embodiments of the present invention, structural consistency regularization can refer to constraints applied at the level of the aforementioned spatial sub-regions, requiring pixels within the same sub-region to have the same or similar cloud category prediction results, while not imposing mandatory constraints on prediction results between different sub-regions. This regularization can refer to elevating pixel-level independent prediction to a region-level overall decision, thereby utilizing the spatial structural information of the image itself to suppress local isolated noise and fragmented activation.

[0052] According to embodiments of the present invention, superpixels can be a typical technique for achieving the aforementioned spatial sub-region partitioning. A superpixel refers to a pixel block composed of pixels that are geographically adjacent and have similar characteristics such as color and texture, exhibiting good edge-preserving properties. Commonly used superpixel segmentation algorithms include the Simple Linear Iterative Clustering (SLIC) algorithm, which generates uniformly sized, edge-fitting superpixel segmentation results by performing local clustering in a five-dimensional space (LAB color coordinates and pixel coordinates). Furthermore, GraphCut is a graph-based image segmentation method. Its core idea is to represent the image as a weighted graph structure, where pixels or regions serve as graph nodes, and adjacent nodes are connected by edges, with edge weights reflecting the similarity between pixels. Subsequently, the optimal partitioning of image regions is achieved by solving the minimum cut (Min-Cut) or maximum flow (Max-Flow) problem. This method can simultaneously consider region consistency and boundary continuity, thus possessing good edge-preserving and global optimization capabilities. Because it can effectively refine target boundaries and suppress noise, GraphCut is widely used in tasks such as superpixel generation, weakly supervised semantic segmentation, remote sensing target extraction, and CAM post-processing.

[0053] In one embodiment of the present invention, the GraphCut algorithm can be used to divide the image to be detected into N superpixel regions. Then, based on the binary result of the spatially augmented activation map, the proportion of pixels judged as clouds in each superpixel region is calculated: if the proportion exceeds 0.5 (i.e., most pixels are judged as clouds), the region-level pseudo-label of the superpixel is marked as "cloud"; otherwise, it is marked as "non-cloud". Next, the average predicted response value of the spatially augmented activation map in each superpixel region is calculated, and a region-level binary cross-entropy loss is constructed using the proportion of the area of ​​each superpixel region (i.e., the number of pixels it contains) to the total number of pixels in the entire image as the weight. After adding the above loss to the training objective, the model is forced to output consistent prediction results within the same superpixel, thereby significantly reducing isolated noise points and holes inside the cloud, and using pseudo-labels as supervision signals for constraint, resulting in an optimized activation map. For example, a cloud activation region that originally had multiple holes inside is filled with holes after this regularization constraint, and the cloud as a whole presents a continuous and smooth shape.

[0054] During operation of S150, under image-level label supervision, the optimized activation map is output as a pseudo-label for the cloud region.

[0055] According to embodiments of the present invention, image-level label supervision can refer to training the model using only binary or multi-class labels indicating the presence or absence of clouds in the entire image, without using pixel-level labels. Image-level label annotation is low-cost; for example, for a remote sensing image, the annotator only needs to determine whether the image contains clouds, without needing to delineate cloud boundaries pixel by pixel. The method of the present invention achieves high-quality cloud detection under such extremely weak supervision conditions.

[0056] According to an embodiment of the present invention, outputting cloud region pseudo-labels through the optimized activation map can mean directly outputting the optimized activation map obtained in operation S140 as the cloud region pseudo-label. The candidate pseudo-labels generated in operation S130 serve as intermediate supervision signals to constrain the learning of the global branch class activation map, employing pixel-level cross-entropy loss. Operation S140 applies structural consistency regularization to the spatially enhanced activation map at the superpixel region level, making the activation map more consistent with the multi-scale morphological characteristics of real clouds in terms of spatial structure and boundary contours. The final output optimized activation map can then be used as the cloud region pseudo-label for downstream tasks.

[0057] According to embodiments of the present invention, by extracting activation responses from multiple different spatial perspectives and applying consistency constraints, the shortcomings of traditional activation maps that only focus on local discriminative regions are effectively overcome, making the generated activation map more continuous and complete in spatial distribution, and able to adapt to the detection needs of clouds of different scales and shapes. By utilizing the physical attribute information of cloud targets to construct spatial priors and perform adaptive filtering, the physical rationality constraints of cloud radiation characteristics are introduced on the basis of pure data-driven approach, significantly suppressing false activation phenomena in complex backgrounds such as snow, bright-colored bare land, and water bodies. By applying structural consistency regularization at the spatial sub-region level of the image, pixel-level prediction is elevated to region-level decision-making, effectively reducing the fragmentation, noise, and hole problems of activation results, making the generated cloud mask more consistent with the multi-scale morphological characteristics of real clouds in terms of boundary integrity and spatial consistency. Under extremely weak supervision using only image-level labels, high-precision and highly robust cloud detection output is achieved through synergistic effects, effectively reducing manual annotation costs and improving the generalization ability across sensors and scenes.

[0058] Figure 2 The diagram illustrates a cloud region label generation method under weak supervision of physical laws and structure perception according to an embodiment of the present invention.

[0059] like Figure 2 As shown, activation response extraction from multiple spatial perspectives of the image to be detected can begin by using the image to be detected and its horizontally flipped image (Flip Image) as input via a global branch to generate corresponding class activation maps. The class activation map is defined as follows. Let the feature map output by the feature extraction network be f(x), where x represents the spatial location, and the classification weight vector corresponding to category k is... Then for the class activation graph of category k Calculated using the following formula:

[0060] According to an embodiment of the present invention, for the binary classification task of cloud detection, the category k takes the value of either cloud or non-cloud. For horizontally flipped images... Similarly, a corresponding class activation graph is generated. Then, it undergoes an inverse spatial transformation (flipped horizontally again) to align its spatial orientation with that of the original image, and is fused with the class activation map M generated from the original image to obtain the global augmentation response. The fusion method can be either taking the maximum value for each pixel or using a weighted average.

[0061] Simultaneously, the image to be detected is divided into multiple non-overlapping sub-blocks (Patch Image) through local branching. For example, a 512×512 pixel image can be divided into four 256×256 pixel sub-blocks. Each sub-block is input into the local branch network to generate a local class activation graph corresponding to each sub-block. Finally, the class activation maps of each sub-block are stitched together according to the spatial position of each sub-block in the original image to restore a complete activation response map with the same size as the original image. This figure shows the activation response from a local perspective.

[0062] By working in concert with global and local branches, this invention can simultaneously capture the overall distribution patterns of large-scale clouds and the fine structural information of small-scale cloud clusters.

[0063] In some embodiments, the specific implementation of applying consistency constraints to activation responses from different perspectives and generating spatially enhanced activation graphs is as follows.

[0064] According to an embodiment of the present invention, a global enhanced response is obtained. Local whole-image response Next, a pixel-level consistency loss function is introduced:

[0065]

[0066] Where ||·||_1 represents the L1 norm, which is the sum of absolute differences calculated pixel by pixel. This loss function measures the degree of inconsistency at the pixel level between the class activation maps generated by the two branches.

[0067] The consistency loss function described above is incorporated into the overall training objective of the entire model. The model can also perform image-level classification tasks, with the classification loss function being:

[0068]

[0069] in The category score is obtained by applying global average pooling to the class activation graph. σ is the image-level label (with a value of 0 or 1, indicating whether the image contains clouds), and σ(·) is the Sigmoid activation function.

[0070] The overall classification loss function is = + λ· λ is the weighting coefficient balancing the two types of losses, typically set to 1. By minimizing the consistency loss, the activation responses of the global and local branches are forced to gradually converge in spatial distribution, ultimately generating a spatially enhanced activation graph. Compared to the original single-branch activation graph, this activation graph significantly improves both the integrity of cloud region coverage and cross-scale stability.

[0071] According to embodiments of the present invention, the operation of constructing spatial priors may also be included, which may be performed independently before operation S130 or calculated in real time during operation S130.

[0072] Specifically, the dark channel transmittance is determined based on an atmospheric scattering model. Dark channel transmittance. Calculated using the following formula:

[0073]

[0074] Where x represents the pixel position, Ω(x) is the local neighborhood window centered at x, and c iterates through the red, green, and blue color channels. Let be the pixel value of pixel x in channel c. The formula means: within a local window Ω(x) centered at x, first take the minimum value of each pixel across the three color channels, then take the minimum value of all pixels within the window, and finally subtract this minimum value from 1 to obtain the transmittance estimate. Lower The value indicates that the light transmission ability of this local area is weak, corresponding to a thicker cloud layer.

[0075] Meanwhile, a fog optimization transformation index (HOT index) is constructed based on a linear combination of the blue and red bands, and its calculation formula is as follows:

[0076]

[0077] in This represents the reflectance value in the blue band. This represents the reflectance value in the red band. This index is based on the physical property that the reflectance of thin clouds in the visible light band decreases with increasing wavelength: the reflectance in the blue band is significantly higher than that in the red band. Therefore, the HOT index shows a high positive value in cloudless areas, while in cloudless surface areas (such as vegetation and bare land), the value is close to zero or even negative.

[0078] Then, the dark channel transmittance and the haze optimization transformation index are normalized and fused to obtain a unified significance map S:

[0079]

[0080] Norm(·) represents the normalization operation, typically using min-max normalization to map the numerical range to the [0,1] interval. This fusion operation simultaneously enhances the contrast between thick and thin clouds relative to the background and smooths the numerical differences between the two types of clouds, thereby generating a continuous and stable saliency representation. This saliency map S serves as a spatial prior and is used in subsequent operations to generate candidate pseudo-labels.

[0081] In some embodiments, the specific implementation of adaptive filtering of spatial priors and generation of candidate pseudo-labels based on spatially enhanced activation graphs is as follows.

[0082] According to an embodiment of the present invention, unsupervised clustering is first performed on the saliency map S to generate multiple hierarchical candidate pseudo-labels. Let c(x) be the cluster affiliation of each pixel x after clustering, and let c(x) be the corresponding cluster center. The cluster centers were then sorted in ascending order. Based on these clustering results, a set of candidate pseudo-labels was constructed, with the salience levels progressively increasing. :

[0083]

[0084] The first-layer pseudo-label contains only the core region of the high-confidence cloud (the category with the smallest cluster center), the second-layer pseudo-label contains both the first and second layers, and so on, with the Kth-layer pseudo-label containing all K layers. This construction method allows the candidate pseudo-labels to form a progressively expanding sequence from "conservative" to "complete," exhibiting good adaptability to different scenarios and cloud morphologies.

[0085] Then, determine each candidate pseudo-label. The degree of overlap between the binary result M̂ and the spatially augmented activation map. The binary result M̂ is obtained by comparing the spatially augmented activation map. Applying a threshold (typically 0.5) yields: M̂ = { x | > 0.5}

[0086] The degree of overlap is measured by the intersection-union ratio (IoU), and the calculation formula is as follows:

[0087]

[0088] Where |·| represents the number of pixels in the set. The IoU value is between 0 and 1, and the larger the value, the more similar the two binary masks are.

[0089] Finally, the candidate pseudo-label with the highest degree of overlap is selected as the pseudo-label P*:

[0090]

[0091] This adaptive selection mechanism dynamically couples the pseudo-label selection process with the model's current cognitive state: in the early stages of training, when the model's accuracy is low, it tends to select conservative candidate pseudo-labels (smaller k values); as training progresses, the model's activation map gradually becomes more complete and accurate, automatically migrating to more expanded candidate pseudo-labels (larger k values), thus providing the model with progressively enhanced supervision signals. The finally selected pseudo-label P* is used to constrain the learning of the global branch class activation map, employing pixel-level cross-entropy loss.

[0092]

[0093] Where BCE(·,·) represents the binary cross-entropy loss.

[0094] In some embodiments, the specific implementation of applying structural consistency regularization to the activation map of spatial enhancement at the spatial sub-region level of the image is as follows.

[0095] According to an embodiment of the present invention, a superpixel segmentation algorithm (such as Simple Linear Iterative Clustering, SLIC) is first used to divide the image to be detected into multiple superpixel regions with local color and texture consistency, forming a superpixel set. , where N is the total number of superpixels.

[0096] Then, the binary result based on the spatially augmented activation map. Generate region-level pseudo-labels for each superpixel region. The binary result Ĉ is obtained through the analysis of... The maximum value is taken along the channel dimension; for binary classification tasks, this is the thresholded cloud / non-cloud determination. The generation rule for region-level pseudo-labels is:

[0097]

[0098] Where |·| represents the number of pixels within the region, and τ is a preset region consistency threshold, which is 0.5 in this embodiment. The rule means that if more than half of the pixels within a superpixel region are identified as clouds by the current activation map, then the entire superpixel is marked as a cloud; otherwise, it is marked as non-cloud. This operation elevates pixel-level discrete prediction to region-level decision-making, effectively suppressing interference from isolated noise pixels.

[0099] Next, the activation responses corresponding to cloud categories in the computationally augmented activation graph (denoted as...) In each superpixel region Average response value within :

[0100]

[0101] in (x) represents the cloud category activation response value (a continuous value between 0 and 1) at pixel x.

[0102] Finally, a region-level binary cross-entropy loss function is introduced, using the superpixel region area as the weight. :

[0103]

[0104] Where |I| is the total number of pixels in the image, used to normalize the area weight to the [0,1] interval; This is the binary cross-entropy loss function. In this function, larger superpixel regions (corresponding to the main body of large-scale cloud clusters) receive higher weights because their structural integrity has a more significant impact on the final cloud mask quality.

[0105] By minimizing the shape fitting loss, the model is forced to output consistent predictions within the superpixel region, thereby significantly suppressing fragmented activation, filling holes inside the cloud, and smoothing irregular boundaries. Finally, the pseudo-label is used as a supervision signal for constraint to obtain the optimized activation map.

[0106] In some embodiments, before outputting the optimized activation map as a pseudo-label for the cloud region under image-level label supervision, the model also needs to train to generate the optimized activation map by jointly optimizing the classification loss and the pixel-level supervision loss.

[0107] According to an embodiment of the present invention, the classification loss is first calculated using image-level labels. The formula for calculating the classification loss L_cls is:

[0108]

[0109] This classification loss enables the model to learn the ability to distinguish whether the entire image contains clouds.

[0110] Simultaneously, pixel-level supervised loss is calculated using pseudo-labels. The final selected pseudo-label P is used to constrain the learning of the global branch class activation map, employing pixel-level cross-entropy loss. Let the spatially augmented activation map generated by the global branch be... If the pseudo-label is P, then the pixel-level supervision loss is... The calculation formula is:

[0111]

[0112] Where BCE(·,·) represents the binary cross-entropy loss function, which constrains the activation map of the space-enhanced system to maintain consistency with the high-quality pseudo-labels obtained by adaptive screening at the pixel level.

[0113] Finally, an optimized activation map is generated by jointly optimizing the classification loss, consistency loss, pixel-level supervision loss, and shape fitting loss. Specifically, the classification loss, consistency loss, pixel-level supervision loss, and shape fitting loss are weighted and summed as the overall loss function: = + .in and To balance the weighting coefficients of the two types of losses, they are typically set to 1 in practical applications. The backpropagation algorithm is used to simultaneously optimize each loss, enabling the model to learn both global semantic discrimination from image-level labels and pixel-level spatial localization from pseudo-labels. After sufficient training, the activation map generated by the model achieves optimal performance in terms of cloud region coverage integrity, boundary clarity, and background suppression. This activation map is the optimized activation map and can be directly output as a pseudo-label for the cloud region.

[0114] Based on the aforementioned cloud region label generation method under weak supervision of physical laws and structure perception, this invention also provides a device for cloud region label generation under weak supervision of physical laws and structure perception. The following will be combined with... Figure 4 This paper provides a detailed description of the cloud region label generation method and apparatus under weak supervision of physical laws and structural perception.

[0115] Figure 3 The diagram illustrates the structure of a cloud region label generation method apparatus under weak supervision of physical laws and structure perception according to an embodiment of the present invention.

[0116] like Figure 3 As shown, the cloud region label generation method device under weak supervision of physical laws and structure perception in this embodiment includes an extraction module 310, a graph generation module 320, a label generation module 330, an optimization module 340, and an output module 350.

[0117] The extraction module 310 is used to extract activation responses from the image to be detected from multiple different spatial perspectives, obtaining the activation responses under each perspective. In one embodiment, the extraction module 310 can be used to perform the operation S110 described above, which will not be repeated here.

[0118] The graph generation module 320 is used to apply consistency constraints to the activation responses from different viewpoints to generate a spatially enhanced activation graph. In one embodiment, the graph generation module 320 can be used to perform the operation S120 described above, which will not be repeated here.

[0119] The label generation module 330 is used to adaptively filter spatial priors based on the spatially enhanced activation graph to generate pseudo-labels. The spatial priors are constructed using the physical attribute information of the cloud targets. In one embodiment, the label generation module 330 can be used to perform the operation S130 described above, which will not be repeated here.

[0120] The optimization module 340 applies structural consistency regularization to the spatially enhanced activation map at the spatial sub-region level of the image, and uses the pseudo-label as a supervision signal for constraint, to obtain the optimized activation map. In one embodiment, the optimization module 340 can be used to perform the operation S140 described above, which will not be repeated here.

[0121] The output module 350 is used to output cloud region pseudo-labels using the optimized activation map and pseudo-labels under image-level label supervision. In one embodiment, the output module 350 can be used to perform the operation S150 described above, which will not be repeated here.

[0122] According to an embodiment of the present invention, the activation response extraction of the image to be detected from multiple different spatial perspectives to obtain the activation response under each perspective includes: generating a corresponding class activation map by taking the image to be detected and its horizontally flipped image as input through a global branch; performing an inverse transform on the class activation maps and then fusing them to obtain the activation response under the global perspective; dividing the image to be detected into multiple non-overlapping sub-blocks through a local branch; generating the class activation map of each sub-block and then splicing them according to the original spatial position of the sub-blocks to obtain the activation response under the local perspective.

[0123] According to an embodiment of the present invention, applying consistency constraints to activation responses under each viewpoint to generate a spatially enhanced activation map includes: determining the pixel-level difference between activation responses under a global viewpoint and activation responses under a local viewpoint; using the difference as a consistency loss function to constrain the activation responses under a global viewpoint to remain consistent with the activation responses under a local viewpoint, thereby generating a spatially enhanced activation map.

[0124] According to an embodiment of the present invention, the device further includes: a priori module, used to determine the dark channel transmittance based on an atmospheric scattering model, wherein the dark channel transmittance is used to highlight thick cloud regions; to construct a fog optimization transformation index based on a linear combination of blue and red bands, wherein the fog optimization transformation index is used to enhance the sensitivity of thin cloud regions; and to fuse the dark channel transmittance and the fog optimization transformation index after normalization to obtain a unified saliency map as a spatial prior.

[0125] According to an embodiment of the present invention, the process of adaptively filtering spatial priors based on spatially enhanced activation maps to generate pseudo-labels includes: performing unsupervised clustering on saliency maps to generate multiple hierarchical candidate pseudo-labels; determining the degree of overlap between each candidate pseudo-label and the binary result of the spatially enhanced activation map; and selecting the candidate pseudo-label with the highest degree of overlap as the pseudo-label.

[0126] According to an embodiment of the present invention, structural consistency regularization is applied to the spatially enhanced activation map at the spatial sub-region level of the image, and the pseudo-label is used as a supervision signal for constraint to obtain an optimized activation map. This includes: dividing the image to be detected into multiple superpixel regions with local color and texture consistency using a superpixel segmentation algorithm; calculating the percentage of pixels belonging to the cloud region for each superpixel region based on the binary result of the spatially enhanced activation map; determining the region-level pseudo-label of the superpixel region as the cloud category when the percentage exceeds a preset threshold, otherwise determining it as the non-cloud category; calculating the average predicted response value of the spatially enhanced activation map in each superpixel region; and introducing a region-level binary cross-entropy loss function with the superpixel region area as the weight to constrain the average predicted response of each superpixel region to be consistent with the region-level pseudo-label of the superpixel region.

[0127] According to an embodiment of the present invention, before outputting the optimized activation map as a pseudo-label for the cloud region under image-level label supervision, the method further includes: calculating a classification loss using the image-level label, wherein the classification loss adopts the binary cross-entropy loss function; calculating a pixel-level supervision loss using the pseudo-label, wherein the pixel-level supervision loss is used to ensure that the activation map of the constraint space enhancement is consistent with the pseudo-label; and training to generate the optimized activation map by jointly optimizing the classification loss and the pixel-level supervision loss.

[0128] According to embodiments of the present invention, any plurality of modules among the extraction module 310, image generation module 320, tag generation module 330, optimization module 340, and output module 350 may be combined into one module, or any one of these modules may be split into multiple modules. Alternatively, at least part of the functionality of one or more of these modules may be combined with at least part of the functionality of other modules and implemented in one module. According to embodiments of the present invention, at least one of the extraction module 310, image generation module 320, tag generation module 330, optimization module 340, and output module 350 may be at least partially implemented as hardware circuitry, such as a field-programmable gate array (FPGA), a programmable logic array (PLA), a system-on-a-chip, a system-on-a-substrate, a system-on-package, an application-specific integrated circuit (ASIC), or any other reasonable means of integrating or packaging the circuitry, or implemented in software, hardware, or firmware, or in any one of the three implementation methods or a suitable combination of any of them. Alternatively, at least one of the extraction module 310, the graph generation module 320, the label generation module 330, the optimization module 340, and the output module 350 may be implemented at least partially as a computer program module, which can perform corresponding functions when the computer program module is run.

[0129] Figure 4The diagram illustrates an electronic device suitable for implementing a cloud region tag generation method under weak supervision of physical laws and structural awareness according to an embodiment of the present invention.

[0130] like Figure 4 As shown, an electronic device according to an embodiment of the present invention includes a computing unit 401. The computing unit 401 can execute various actions and processes in the data transmission method of the present invention according to a program stored in a read-only memory 402 or a program loaded from a storage unit 408 into a random access memory 403. The computing unit 401 may include, for example, a general-purpose microcomputing unit (e.g., a central processing unit), a graphics computing unit, an embedded hardware acceleration unit (e.g., a field-programmable gate array or a high-performance system-on-a-chip), and / or a related chipset. The computing unit 401 may also include onboard memory for caching purposes.

[0131] The random access memory 403 stores various programs and data required for the operation of the electronic device 400, such as the operating system, device drivers, address information of the target storage area, and target data obtained from the storage device. The computing unit 401, the read-only memory 402, and the random access memory 403 are interconnected via a bus 404. The computing unit 401 executes various operations of the method flow according to the embodiments of the present invention by executing programs stored in the read-only memory 402 and / or the random access memory 403. It should be noted that programs may also be stored in one or more memories other than the read-only memory 402 and the random access memory 403. The computing unit 401 may also execute various operations of the method flow according to the embodiments of the present invention by executing programs stored in one or more memories.

[0132] According to an embodiment of the present invention, the electronic device may further include an input / output interface 405, which is also connected to a bus 404. The electronic device may also include one or more of the following components connected to the input / output interface 405: an input unit 406, including a keyboard, mouse, touchscreen, etc., for receiving user input; an output unit 407, including a liquid crystal display, status indicator lights, speakers, etc.; a storage unit 408, including a hard disk or solid-state drive, for storing the OS, applications, and target data; and a communication unit 409, including a network interface card, such as a local area network card, modem, etc. The communication unit 409 can be used to communicate with the BMC or to interact with a remote management terminal. A drive is also connected to the input / output interface 405 as needed. Removable media, such as disks, optical disks, magneto-optical disks, semiconductor memories, etc., are installed on the drive as needed so that computer programs read from them can be installed into the storage unit 408 as needed.

[0133] The present invention also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs, which, when executed, implement the data transmission method according to embodiments of the present invention.

[0134] According to embodiments of the present invention, a computer-readable storage medium may be a non-volatile computer-readable storage medium, such as including, but not limited to: portable computer disks, hard disks, random access memory, read-only memory, erasable programmable read-only memory, portable compact disk read-only memory, optical storage devices, magnetic storage devices, or any suitable combination thereof. In the present invention, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, according to embodiments of the present invention, a computer-readable storage medium may include the read-only memory 402 and / or random access memory 403 described above and / or one or more memories other than read-only memory 402 and random access memory 403.

[0135] Embodiments of the present invention also include a computer program product comprising a computer program containing program code for executing the method flows shown in the embodiments of the present invention. When the computer program product is run on a computer system, the program code is used to enable the computer system to implement the data transmission method provided in the embodiments of the present invention.

[0136] When the computer program is executed by the computing unit 401, it performs the functions defined in the system / apparatus of this embodiment of the invention. According to embodiments of the invention, the systems, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0137] In one embodiment, the computer program may rely on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may also be transmitted and distributed in the form of signals over a network medium, and downloaded and installed via communication unit 409, and / or installed from a removable medium. The program code contained in the computer program can be transmitted using any suitable network medium, including but not limited to: wireless, wired, etc., or any suitable combination thereof.

[0138] In such an embodiment, the computer program can be downloaded and installed from a network via the communication unit 409, and / or installed from a removable medium. When the computer program is executed by the computing unit 401, it performs the functions defined in the system of this embodiment of the invention. According to embodiments of the invention, the systems, devices, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0139] According to embodiments of the present invention, program code for executing the computer programs provided in the embodiments of the present invention can be written in any combination of one or more programming languages. Specifically, these computational programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. Programming languages ​​include, but are not limited to, languages ​​such as Java, C++, Python, "C", or similar programming languages. The program code can be executed entirely on the user's computing device, integrally on the user's device, integrally on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network or a wide area network, or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0140] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0141] Those skilled in the art will understand that the features described in the various embodiments of the present invention can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in the present invention. In particular, the features described in the various embodiments of the present invention can be combined and / or combined in various ways without departing from the spirit and teachings of the present invention. All such combinations and / or combinations fall within the scope of the present invention.

Claims

1. A method for generating cloud region labels under weak supervision based on physical laws and structural perception, characterized in that, include: Activation responses are extracted from the image to be detected from multiple different spatial perspectives to obtain activation responses under each perspective, and the activation responses are class activation maps. A consistency constraint is applied to the activation response under each viewpoint to generate a spatially enhanced activation graph. The application of the consistency constraint is achieved by constructing a loss function. Based on the spatially enhanced activation graph, the spatial prior is adaptively filtered to generate pseudo-labels. The spatial prior is constructed using the physical attribute information of the cloud target. Structural consistency regularization is applied to the spatially enhanced activation map at the spatial sub-region level of the image, and the pseudo-label is used as a supervision signal for constraint to obtain the optimized activation map. Under image-level label supervision, the optimized activation map is output as a pseudo-label for the cloud region; The spatially enhanced activation map is subjected to structural consistency regularization at the spatial sub-region level of the image, and constrained using the pseudo-label as a supervision signal, resulting in an optimized activation map including: The image to be detected is divided into multiple superpixel regions with local color and texture consistency using a superpixel segmentation algorithm; Based on the binary result of the spatially enhanced activation map, the percentage of pixels belonging to the cloud region is calculated for each superpixel region; When the proportion exceeds a preset threshold, the region-level pseudo-label of the superpixel region is determined to be a cloud category; otherwise, it is determined to be a non-cloud category. Calculate the average predicted response value of the spatially enhanced activation map in each superpixel region; Using the area of ​​the superpixel region as the weight, a region-level binary cross-entropy loss function is introduced to constrain the average predicted response of each superpixel region to be consistent with the region-level pseudo-label of the superpixel region.

2. The cloud region label generation method under weak supervision based on physical laws and structure perception as described in claim 1, characterized in that, The process of extracting activation responses from the image to be detected from multiple different spatial perspectives to obtain activation responses under each perspective includes: The corresponding class activation map is generated by taking the image to be detected and its horizontally flipped image as input through a global branch; The activation graphs of the class are aligned and then fused to obtain the activation responses from a global perspective. The image to be detected is divided into multiple non-overlapping sub-blocks by local branching; After generating class activation maps for each sub-block, the maps are spliced ​​together according to the original spatial positions of the sub-blocks to obtain the activation response from a local perspective.

3. The cloud region label generation method under weak supervision based on physical laws and structure perception as described in claim 2, characterized in that, The process of applying consistency constraints to the activation responses under each viewpoint to generate a spatially enhanced activation graph includes: Determine the pixel-level difference between the activation response in the global view and the activation response in the local view; The difference is used as a consistency loss function to constrain the activation response under the global view to be consistent with the activation response under the local view, thereby generating a spatially enhanced activation graph.

4. The cloud region label generation method under weak supervision based on physical laws and structure perception as described in claim 1, characterized in that, The method further includes: Dark channel transmittance is determined based on an atmospheric scattering model, and the dark channel transmittance is used to highlight thick cloud regions. A fog optimization transformation index is constructed based on a linear combination of blue and red bands, which is used to enhance the sensitivity of thin cloud regions. The dark channel transmittance and the fog optimization transformation index are normalized and then fused to obtain a unified saliency map as the spatial prior.

5. The cloud region label generation method under weak supervision based on physical laws and structure perception as described in claim 4, characterized in that, The step of adaptively filtering spatial priors based on the spatially enhanced activation map to generate pseudo-labels includes: Unsupervised clustering is performed on the saliency graph to generate multiple levels of candidate pseudo-labels; Determine the degree of overlap between each candidate pseudo-label and the binary result of the spatially augmented activation map; The candidate pseudo-label with the highest degree of overlap is selected as the pseudo-label.

6. The cloud region label generation method under weak supervision based on physical laws and structure perception according to claim 5, characterized in that, Before outputting the optimized activation map as a pseudo-label for the cloud region under image-level label supervision, the process further includes: The classification loss is calculated using the image-level labels, and the classification loss adopts the binary cross-entropy loss function; The pixel-level supervision loss is calculated using the pseudo-label, and the pixel-level supervision loss is used to constrain the spatially augmented activation map to remain consistent with the pseudo-label; The optimized activation map is generated by jointly optimizing the classification loss and the pixel-level supervision loss.

7. A cloud region tag generation device under weak supervision of physical laws and structural perception, characterized in that, include: The extraction module is used to extract activation responses from the image to be detected from multiple different spatial perspectives, and obtain the activation responses under each perspective. The graph generation module is used to apply consistency constraints to the activation responses under each viewpoint and generate a spatially enhanced activation graph. The tag generation module is used to adaptively filter the spatial prior based on the spatially enhanced activation graph and generate pseudo-tags. The spatial prior is constructed using the physical attribute information of the cloud target. The optimization module is used to apply structural consistency regularization to the spatially enhanced activation map at the spatial sub-region level of the image, and to constrain it using the pseudo-label as a supervision signal to obtain the optimized activation map. The output module is used to output the optimized activation map as a pseudo-label for the cloud region under image-level label supervision.

8. An electronic device, comprising: One or more computing units; Memory, used to store one or more computer programs. The characteristic feature is that the one or more computing units execute the one or more computer programs to implement the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed by the computing unit, they implement the steps of the method according to any one of claims 1 to 6.

Images