A reservoir water gauge recognition method based on polarization imaging and heterogeneous target integration

Abstract

The application belongs to the field of image recognition and relates to a reservoir water gauge recognition method based on polarization imaging and heterogeneous target integration, which comprises the following steps: acquiring a light wave state feature matrix and a polarization feature map of a reservoir water gauge; based on the polarization feature map, performing self-correcting light perception enhancement preprocessing on a to-be-detected image of the reservoir water gauge to obtain a pretreated image; decoupling polarization intensity maps and polarization feature maps under multiple polarization angles through a polarization dual-domain feature decoupling network to obtain a reconstructed image; constructing a heterogeneous target integration optimization framework and processing a transmission component through the heterogeneous target integration optimization framework to obtain an absolute water level measurement value. Through this end-to-end joint optimization, the error accumulation effect caused by the traditional step-by-step processing is effectively avoided, and finally, the accurate, universal and highly robust automatic water level recognition function of different types of water gauges under harsh conditions such as strong reflection and extreme light is realized.

Description

Technical Field

[0001] This invention relates to the field of image recognition, and specifically discloses a method for recognizing reservoir water level gauges based on polarization imaging and heterogeneous target integration. Background Technology

[0002] Reservoir water level gauge identification is a fundamental task in water conservancy projects, hydrological monitoring, and modern reservoir management systems. Accurate and real-time acquisition of water level data is a crucial prerequisite for achieving intelligent flood warning, optimized water resource allocation, and ensuring the safety of water conservancy facilities. Given that traditional manual observation methods are inefficient and susceptible to environmental factors, making it difficult to meet the requirements of high-frequency, high-precision, and all-weather monitoring, automated water level gauge identification technology based on machine vision and deep learning has become a recognized research focus in the industry.

[0003] Existing water level gauge identification methods are mainly based on traditional image processing techniques or early convolutional neural network models. These techniques typically process standard RGB images, and their core methods include edge detection of scale lines, image segmentation of specific regions, and digit extraction based on general optical character recognition (OCR) models. However, in the complex, variable, and uncontrolled environment of outdoor reservoirs, these methods reveal many shortcomings:

[0004] First, as a typical optical mirror, the water surface inevitably produces strong specular reflection (glare) under sunlight, causing the water level gauge information to be severely obscured or distorted. Current reservoir water level gauge recognition technology only includes recognition functions and lacks the ability to deeply perceive light and solve the problem of water surface reflection. It cannot effectively separate the reflected component from the background transmitted component, and the residual reflection artifacts greatly reduce the accuracy of subsequent recognition models.

[0005] Secondly, extreme lighting conditions cause severe distortion of image information. In uncontrolled outdoor environments, the acquisition quality of water level gauge images is highly susceptible to the drastic impact of lighting. Under low-light and underexposed conditions, the dynamic range of image information is compressed, and pixel details become saturated or lost; while under strong-light and overexposed conditions, pixel saturation distortion leads to the loss of scale details and color information. Existing methods lack a unified mechanism that can compensate for the effects of image distortion and achieve adaptive image quality and high-fidelity correction.

[0006] Finally, existing models suffer from severe deficiencies in versatility and generalization ability, and generally employ a multi-stage serial processing architecture. As a non-standardized measurement tool, the scale intervals, numeral styles, and dimensions of water gauges vary significantly across different application scenarios, leading to a sharp decline in the cross-dimensional recognition capabilities of traditional deep learning models. Simultaneously, the pipelined serial structure used in most solutions isolates image preprocessing from the core recognition task, inevitably causing errors from upstream modules to accumulate downstream, and the preprocessing module cannot receive high-order features or error feedback from the core recognition task. This non-joint optimization mechanism significantly limits the performance potential of the entire system across different dimensions and resolutions.

[0007] In view of this, the present invention provides a reservoir water level gauge identification method based on polarization imaging and heterogeneous target integration. By using physical optical perception to deal with reflection interference, a scale-universal feature learning mechanism is established to adapt to changes in dimensions. Finally, through a highly collaborative and jointly optimized architecture, all robust functions are integrated to achieve accurate and universal automated identification of water level gauges in complex environments. Summary of the Invention

[0008] The purpose of this invention is to provide a reservoir water level gauge identification method based on polarization imaging and heterogeneous target integration, addressing the problems of effectively reducing interference from water surface reflection and adapting to extreme lighting conditions such as underexposure in low light or overexposure in strong light, thereby achieving water level gauge identification under different dimensions. The specific solution is as follows:

[0009] The process includes: S1: Obtaining the light wave state feature matrix and polarization feature map of the reservoir water gauge based on the collected polarization intensity maps at various polarization angles; S2: Performing self-correcting illumination sensing enhancement preprocessing on the image to be detected of the reservoir water gauge based on the polarization feature map to obtain a preprocessed image; the preprocessed image includes an enhanced image and a corrected image; S3: Constructing a polarization dual-domain feature decoupling network to decouple the polarization intensity maps and polarization feature maps at various polarization angles to obtain a reconstructed image; the polarization dual-domain feature decoupling network includes a transmission prediction domain and a reflection prediction domain, which mutually guide cross-domain feature decoupling to generate transmission and reflection components respectively; S4: Constructing a heterogeneous target integration optimization framework and processing the transmission component through the heterogeneous target integration optimization framework to obtain the absolute water level measurement value; the heterogeneous target integration optimization framework includes a shared feature encoder, a task-aware decoder, and a water level calculation and output module.

[0010] Further, step S1 includes: S101: acquiring polarization intensity maps under multiple polarization angles using a polarization intensity acquisition device; the multiple polarization angles are greater than or equal to four; S102: determining the light wave state feature matrix based on the sum and difference relationship between the polarization intensity maps under different polarization angles; the light wave state feature matrix includes multiple light wave state feature components; S103: determining the polarization feature map based on the multiple light wave state feature components; the polarization feature map includes the light polarization purity index and the polarization principal axis direction.

[0011] Furthermore, when there are four different polarization angles, the characteristic matrix of the light wave state is:

[0012] ;

[0013] ;

[0014] ;

[0015] ;

[0016] Where M is the light wave state characteristic matrix; , and These represent the characteristic components of the first, second, and third light wave states, respectively; T is the transpose of the matrix. , , and These are the first, second, third, and fourth polarization intensity components, respectively. The polarization angle of the first polarization intensity component is... The polarization angle of the second polarization intensity component is The polarization angle of the third polarization intensity component is The polarization angle of the fourth polarization intensity component is ;

[0017] The light polarization purity index is:

[0018] ;

[0019] The direction of the principal polarization axis is:

[0020] ;

[0021] Where Π is the polarization purity index of the light ray; Φ is the direction of the principal polarization axis; and atan2 is the arctangent function in the four quadrants.

[0022] Further, step S2 includes: S201: determining the image type of the image to be detected; the image type includes low-light underexposed images and strong-light overexposed images; S202: for low-light underexposed images, performing self-supervised low-light enhancement on the low-light underexposed images through the low-light enhancement module to obtain an enhanced image; S203: for strong-light overexposed images, performing saturation correction and detail restoration on the strong-light overexposed images through the strong-light correction module to obtain a corrected image.

[0023] Furthermore, the low-light enhancement module employs a generative adversarial network (GAN). The generator of the low-light enhancement module uses a U-Net structure, integrating non-local attention blocks to predict illumination maps and convert underexposed low-light images into enhanced images. The discriminator of the low-light enhancement module includes a panoramic discriminator and one or more focus discriminators. The panoramic discriminator evaluates the overall brightness distribution and panoramic realism of the enhanced image; the focus discriminator enhances the local texture and contrast of the watermark area. The strong light correction module employs a frequency-domain geometric information heterogeneous decoupling framework, which includes a deep neural network, a joint correction module, and a detail restoration network. The deep neural network decouples the overexposed strong light image into a low-frequency basal layer and a high-frequency detail layer. The joint correction module performs joint correction on the saturation components of the low-frequency basal layer to obtain the corrected basal layer. The detail restoration network fuses the high-frequency detail layer with the corrected basal layer to obtain the corrected image.

[0024] Furthermore, the low-light enhancement loss of the low-light enhancement module is:

[0025] ;

[0026] ;

[0027] ;

[0028] ;

[0029] in, To enhance loss in low light; For mapping weights; For multi-level semantic feature consistency weights; The weights are the light intensity gradient constraint weights. For mapping loss; is the expected value; log is the logarithmic function; The discrimination result of the panoramic discriminator; To enhance the image; To focus on the discrimination results of the discriminator; This represents the multi-level semantic feature consistency loss; i is the feature layer-level index variable. Let be the total number of elements in the feature map of the i-th layer; Here is the feature map extracted by network Φ at layer i; This is an underexposed image in low light. It is an L1 norm; The loss is defined by the light intensity gradient constraint; x and y are the x and y coordinates of the pixel, respectively. and These are the horizontal and vertical gradient operations at pixel (x, y), respectively. This is the enhanced image at pixel (x,y); Let G be the L2 norm; G is the generator network;

[0030] The intensity correction loss of the intensity correction module is:

[0031] ;

[0032] in, To correct for high light loss; This is the first correction weight; To correct the image; The target image is represented by k, which is the scale variable. This is the second correction weight; Features extracted at the k-th scale;

[0033] The preprocessing loss for the low-light enhancement module and the high-light correction module is:

[0034] ;

[0035] ;

[0036] in, To pre-process losses; Enhance weights for low light conditions; Weighting for strong light correction; Preserve weights for geometric contours; The loss is used to preserve the geometric contour; SSIM() is the structural similarity index. P is the gradient map of the preprocessed image; P is the polarization feature map.

[0037] Further, step S3 includes: S301: performing transmission domain prediction on polarization intensity maps and polarization feature maps under multiple polarization angles to obtain transmission components; S302: performing reflection domain prediction on polarization intensity maps and polarization feature maps under multiple polarization angles to obtain reflection components; S303: fusing transmission components and reflection components to obtain reconstructed images.

[0038] Further, step S4 includes: S401: The shared feature encoder uses a deep convolutional network with channel attention and spatial attention to perform convolution processing on the transmission component of the input to obtain a shared feature representation; S402: The shared feature representation is decoupled by the task-aware decoder to obtain super-resolution features and semantic sequence features; the semantic sequence features include sequence semantics, position and scale context information; S403: The water level calculation and output module calculates the water level by performing water level calculation on the super-resolution features and semantic sequence features to obtain the absolute water level measurement value.

[0039] Further, step S402 includes: extracting features from the shared feature representation using a high-resolution decoder to obtain high-resolution features; the high-resolution features include high-frequency information, edges, and local textures; processing the shared feature representation and the high-resolution features to obtain a high-definition feature map; extracting features from the shared feature representation using a semantic sequence decoder to obtain semantic sequence features; transforming the semantic sequence features to obtain a query signal; and generating high-resolution features based on the query signal.

[0040] Further, step S403 includes: segmenting the water surface line within the water gauge region in the shared feature representation to obtain the pixel positions of the water mask and the water gauge scale line; identifying the pixel positions of the water gauge scale line to determine the water level number and geometric position of the water gauge scale line; determining the reference scale line based on the geometric position of the water gauge scale line; determining the pixel distance between the pixel positions of the reference scale line and the water gauge scale line; and converting the pixel distance to obtain the absolute water level measurement value.

[0041] The present invention has the following advantages and beneficial effects:

[0042] This invention uses physical priors of polarization imaging and a polarization dual-domain feature decoupling network to efficiently separate the specular reflection and background transmission components of the water surface from a physical perspective, effectively solving the recognition failure problem caused by water surface glare, which is difficult to deal with by traditional methods.

[0043] This invention employs a self-calibrating illumination perception enhancement preprocessing mechanism, utilizing a latent variation mapping-realism calibration framework and a frequency domain geometric information heterogeneous decoupling framework to faithfully correct underexposed images in low light and overexposed images in strong light, thus ensuring the integrity of water level gauge information under extreme lighting conditions.

[0044] This invention establishes a scale normalization and dimensional consistency perception mechanism, thereby achieving cross-scale universality of the model for water gauge dimensions. This significantly enhances the model's generalization ability and practical value in the face of non-standardized water gauges and low-resolution images.

[0045] The heterogeneous target integration optimization framework constructed in this invention avoids the error accumulation of traditional serial architectures and achieves efficient system integration and collaborative optimization by guiding each other through the features of tasks. Attached Figure Description

[0046] Figure 1 An exemplary flowchart of a reservoir water level gauge identification method based on polarization imaging and heterogeneous target integration provided by the present invention;

[0047] Figure 2 This is an exemplary flowchart for obtaining polarization feature maps provided by the present invention;

[0048] Figure 3 An exemplary schematic diagram of the self-supervised mechanism based on partial feature reconstruction provided by the present invention;

[0049] Figure 4 This is the original image to be detected obtained by recognizing actual reservoir water gauge images according to the present invention;

[0050] Figure 5 This is the recognition result image obtained by recognizing actual reservoir water gauge images according to the present invention. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0052] This invention provides a reservoir water level gauge identification method based on polarization imaging and heterogeneous target integration, such as... Figure 1 As shown, it includes the following steps:

[0053] Step S1 involves obtaining the optical wave state feature matrix M and polarization feature map P of the reservoir water gauge based on the collected polarization intensity maps at various polarization angles. Specifically, step S1 includes:

[0054] Step S101: Use a polarization intensity acquisition device to collect polarization intensity maps under multiple polarization angles; there are four or more polarization angles. Figure 2 This invention demonstrates the process of addressing the recognition failure problem caused by water surface glare through polarization intensity map acquisition, light wave state feature matrix calculation, and polarization feature map calculation. By introducing a physical optical sensing mechanism, it aims to fundamentally solve the information loss problem of traditional RGB images in complex optical environments such as water surface glare. Specifically, during water level scene image acquisition, a four-channel polarization camera or rotating polarizer device is used to acquire polarization intensity components at at least four polarization angles under fixed lighting and viewing angle conditions. This is used to fully capture the intensity information of light waves in different polarization directions. When there are four different polarization angles, the following is selected: , , and The measured intensities in the four directions are defined as polarization intensity components, respectively. , , and The multi-angle polarization images form the basis of the network input, providing rich and physically constrained prior information for subsequent reflection separation.

[0055] Step S102: Based on the sum and difference relationship between polarization intensity maps under different polarization angles, determine the light wave state feature matrix; the light wave state feature matrix includes multiple light wave state feature components. Taking four polarization angles as an example, the calculation formula for the light wave state feature matrix is ​​as follows:

[0056] ;

[0057] This matrix is ​​used to fully characterize the total intensity and polarization state of light, and is the core foundation of polarization imaging analysis.

[0058] Three light wave state characteristic components , and The calculation formula is:

[0059] ;

[0060] ;

[0061] ;

[0062] Where M is the light wave state characteristic matrix; , and These are the first, second, and third light wave state characteristic components, which represent different polarization projections of light intensity. The first light wave state characteristic component can be the total light intensity component, the second light wave state characteristic component can be the horizontal linear polarization characteristic component, and the third light wave state characteristic component can be the diagonal linear polarization characteristic component; T is the transpose of the matrix. , , and These are the first, second, third, and fourth polarization intensity components, respectively. The polarization angle of the first polarization intensity component is... The polarization angle of the second polarization intensity component is The polarization angle of the third polarization intensity component is The polarization angle of the fourth polarization intensity component is .

[0063] Step S103: Based on multiple light wave state feature components, determine the polarization feature map; the polarization feature map includes the light polarization purity index Π and the polarization principal axis direction Φ. To further extract features strongly correlated with object surface characteristics, especially the specular reflection characteristics of water surfaces, the light polarization purity index Π and the polarization principal axis direction Φ can be derived from the light wave state feature matrix M, and both are collectively referred to as the polarization feature map P. Π reflects the proportion of polarized light components to the total light intensity, and its high correlation with the specular reflection components of water surfaces makes it a powerful indicator of separated reflection. Φ describes the vibration direction of polarized light and carries the geometric structural information of the water level gauge. The polarization feature map P will serve as the basis for subsequent networks (including the polarization reflection separation network). and light enhancement network The multidimensional input or structural priors of optics enable the efficient embedding of prior knowledge of physical optics into deep learning models. The formula for calculating the polarization purity index of light rays is:

[0064] ;

[0065] The formula for calculating the principal polarization axis direction is:

[0066] ;

[0067] Where Π is the polarization purity index of the light ray; Φ is the direction of the principal polarization axis; and atan2 is the arctangent function in the four quadrants.

[0068] Step S2: The self-calibrating illumination sensing enhancement processing module performs self-calibrating illumination sensing enhancement preprocessing on the image to be detected of the reservoir water level gauge based on the polarization feature map, obtaining a preprocessed image; the preprocessed image includes an enhanced image and a calibrated image. To address the image information distortion problem caused by uncontrolled outdoor strong light overexposure and weak light underexposure in reservoir scenes, a highly robust self-calibrating illumination sensing enhancement preprocessing mechanism is designed. This mechanism first adaptively splits the original image based on brightness to obtain the type of the original image (the image type includes weak light image and strong light image), and then the weak light enhancement module... and strong light correction module Integrated within a unified deep learning framework, a low-light enhancement module and a high-light correction module are used to process the corresponding types of original images, respectively. The low-light enhancement module employs a generative adversarial model, while the high-light correction module uses a multi-scale decomposition structure. Specifically, step S2 includes:

[0069] Step S201: Determine the image type of the image to be detected; image types include underexposed images with low light and overexposed images with strong light. This can be determined through brightness thresholding or histogram analysis. (Low light) or (Strong light) is directed to the corresponding processing module for processing, ultimately outputting a pre-processed image with uniform illumination and clear details. This enables adaptive, high-fidelity correction of water level gauge image quality.

[0070] Step S202: To address the dynamic range compression and detail loss defects in underexposed images caused by low-light underexposure, a low-light enhancement module performs self-supervised low-light enhancement on the underexposed images to obtain an enhanced image. Low-light enhancement module To achieve self-supervised low-light enhancement, this module is built upon a latent variation mapping-realism calibration framework. The low-light enhancement module employs a generative adversarial network (GAN). The core of this module is the generator G, which predicts the illumination map. The generator G (i.e., the generative network) uses a U-Net-based structure, integrating non-local attention blocks. This allows it to utilize global contextual information to predict the illumination map, overcoming the limitations of traditional local convolutions for large-scale illumination unevenness. This enables refined prediction of the illumination distribution in the watermark area, transforming underexposed low-light images... Convert to enhanced image .

[0071] To ensure that the enhanced image possesses high-quality overall illumination distribution and detailed local texture, a panoramic-focus dual-modal collaborative discriminator mechanism is introduced. The discriminator in the low-light enhancement module includes a panoramic discriminator. and one or more focus discriminators Panoramic discriminator The overall brightness distribution and panoramic realism of the enhanced image are evaluated, while the focus discriminator... It focuses on enhancing the local texture and contrast of the watermark scale area. This mechanism, through joint adversarial learning, ensures that the optimization of the generator G takes into account both the overall lighting rationality and the realism of the watermark details.

[0072] In some embodiments, to ensure that the semantic content and high-frequency texture details of the watermark remain high-fidelity while increasing brightness, a multi-level semantic feature consistency loss and low-light enhancement loss are constructed. Used to describe the optimization effect of generator G. The weak light enhancement loss is derived from the mapping loss. Multi-level semantic feature consistency loss Light intensity gradient constraint loss Jointly driven. Mapping loss. The aim is to minimize the enhancement of the image by the panoramic discriminator. and focus discriminator The probability of identifying an image as fake encourages the enhancement of images to achieve a higher level of realism. The mapping loss is:

[0073] ;

[0074] in, For mapping loss; is the expected value, usually calculated from the mean of the batch samples; log is the logarithmic function; The discrimination result of the panoramic discriminator; The generator outputs the image enhancement; To focus on the discrimination results of the discriminator;

[0075] Multi-level semantic feature consistency loss The aim is to ensure that the enhanced image maintains high consistency with the low-light underexposed image in terms of high-level semantic features, preventing geometric distortion. This loss function extracts feature maps through a pre-trained deep network Φ, minimizing the difference in feature space between the low-light input (low-light underexposed image) and the enhanced output (enhanced image), ensuring consistency in content structure. The multi-level semantic feature consistency loss is as follows:

[0076] ;

[0077] in, This represents the semantic feature consistency loss across multiple levels; i is the total number of elements in the layer feature map. Let be the total number of elements in the feature map of the i-th layer; Here is the feature map extracted by network Φ at layer i; This is an underexposed image in low light. It is an L1 norm;

[0078] Light intensity gradient constrained loss The aim is to constrain the smoothness of gradient changes in the enhanced image in space. This loss typically employs total variational loss or a variant thereof, applied to the illumination map R or the enhanced image itself, to eliminate blockiness or abrupt illumination artifacts.

[0079] ;

[0080] in, The loss is defined by the light intensity gradient constraint; x and y are the x and y coordinates of the pixel, respectively. and These are the horizontal and vertical gradient operations at pixel (x, y), respectively. This is the enhanced image at pixel (x,y); The L2 norm is used; this loss ensures a natural lighting transition in the water gauge scale area.

[0081] The low-light enhancement loss of the low-light enhancement module can be expressed as:

[0082] ;

[0083] in, To enhance loss in low light; For mapping weights; For multi-level semantic feature consistency weights; The weights are the light intensity gradient constraint weights.

[0084] Step S203: For overexposed images due to strong light, a strong light correction module performs saturation correction and detail restoration on the overexposed images to obtain a corrected image. To address the pixel saturation distortion and scale detail loss caused by strong light overexposure, a heterogeneous decoupling framework based on frequency domain geometric information is used for saturation correction and detail restoration. The strong light correction module employs a heterogeneous decoupling framework based on frequency domain geometric information, which includes a deep neural network, a joint correction module, and a detail restoration network. The deep neural network of the frequency domain geometric information heterogeneous decoupling framework first processes the overexposed image... The image is converted to the HSV color space, and only the saturation (S) component is processed. Subsequently, a series of multi-resolution processing modules decouple the overexposed image into a low-frequency base layer containing global color and brightness information. and a series of high-frequency detail layers containing information such as edges and scale textures. The decomposition process can be expressed as:

[0085] ;

[0086] in, This is an overexposed image due to strong light. For low-frequency base layer; k is the scale variable; N is the number of scale levels in the decomposition; This represents the high-frequency detail layer at the k-th scale. The decomposition mechanism effectively decouples lighting information from structural texture information through internal downsampling, upsampling, and differential operations.

[0087] After decomposition, the saturation component of the low-frequency substrate is analyzed using a joint correction module. By performing joint correction of the nonlinear mapping function f() and polarization guidance, the corrected substrate layer is obtained. Simultaneously, the polarization feature map P calculated in step S1 is used as a structure-guided prior, and the optimization of the network is guided by loss function constraints to ensure that the corrected saturation change is consistent with the physical edge and structural information of the water gauge.

[0088] To address the loss of high-frequency information caused by overexposure, a detailed network recovery system was designed. Using the high-frequency detail layer as input, it is fused with the corrected basal layer to perform the final depth reconstruction operation, generating a high-fidelity corrected image. This enables the prediction and recovery of lost scale edges and texture information. The network training employs a frequency-domain geometric information heterogeneous decoupling loss as a constraint. The strong light correction loss for the strong light correction module is:

[0089] ;

[0090] in, To correct for high light loss; This is the first correction weight; To correct the image; The target image is represented by k, which is the scale variable. This is the second correction weight; The feature extracted at the k-th scale; this loss ensures that the corrected image has high fidelity in terms of frequency domain geometric information features, effectively restoring the details of the water gauge scale damaged by strong light.

[0091] To ensure the preprocessed image To better preserve the geometry of the water gauge, a geometry profile preservation loss based on polarization feature P was designed. The loss function employs either the structural similarity index (SSIM) or gradient domain loss, and minimizes... Based on the difference between the structural features or gradient domain and the polarization feature map P, the structural information of the preprocessed image is forced to remain consistent with P under physical constraints:

[0092] ;

[0093] in, The geometric contour preservation loss is used; SSIM() is a structural similarity index used to measure the brightness, contrast and structural similarity between the two. For preprocessing images The gradient map is denoted by P; P is the polarization feature map. This loss forces the edges of the enhanced image to be precisely aligned to the physical edges reflected by the polarization data.

[0094] We weight and integrate the low-light enhancement loss, high-light correction loss, and geometric contour preservation loss to form the preprocessing loss of the illumination perception enhancement module. And as a heterogeneous target integration optimization framework One of the auxiliary loss terms in the model is used to achieve adaptive, high-fidelity correction of water level gauge images under extreme lighting conditions.

[0095] ;

[0096] in, To pre-process losses; Enhance weights for low light conditions; Weighting for strong light correction; Preserve weights for the geometric contour.

[0097] In step S2, all image enhancement and image correction steps use the polarization feature map P calculated in step S1 as the structural prior input. The optimization of the network is guided by customized loss function constraints to ensure that the geometric structure and information integrity of the water scale are accurately preserved while correcting brightness and color.

[0098] Step S3 involves constructing a polarization dual-domain feature decoupling network. This network decouples the polarization intensity map and polarization feature map under various polarization angles to obtain the reconstructed image. The polarization dual-domain feature decoupling network includes a transmission prediction domain and a reflection prediction domain, which are mutually guided across the feature decoupling domains. Specifically, the water level background information is recovered through transmission prediction; the specular reflection of the water surface is predicted through reflection prediction; and the Feature Attention Guidance Module (AGM) processes the water level background information and specular reflection of the water surface based on the polarization feature map to obtain the transmission component.

[0099] To efficiently address the severe occlusion and distortion of water level gauge readings caused by surface reflection (glare) from the water surface at the physical level, a polarization dual-domain feature decoupling network is established. The polarization dual-domain feature decoupling network takes the multi-angle polarization image or derived polarization feature map obtained in step S1 as input. Utilizing the physical prior of the significant difference in polarization characteristics between reflected and transmitted light, it treats the input image I as an additive mixture of the transmission component T of the water scale background and the reflection component R of the water surface reflection: I = T + R. The polarization dual-domain feature decoupling network employs a dual-domain decoupling structure, consisting of a transmission prediction domain and a transmission prediction domain. and reflection prediction domain The goal is to achieve simultaneous prediction and accurate separation of T and R, effectively decoupling background information and reflection artifacts at the model level. The mapping relationship of the polarization dual-domain feature decoupling network can be expressed as:

[0100] The mapping relationship of the polarization dual-domain feature decoupling network is as follows:

[0101] ;

[0102] Where T is the transmission component of the water level background; R is the reflection component of the water surface reflection. This is a polarization dual-domain feature decoupling network; I is the input image; P is the polarization feature map. Specifically, step S3 includes:

[0103] Step S301: Perform transmission domain prediction on the polarization intensity map and polarization feature map under various polarization angles to obtain the predicted transmission image. Transmission prediction domain. The core objective is to accurately recover the water gauge background and scale information obscured by the reflected component. This prediction stream employs a residual-connected enhanced cascaded feature extraction backbone network, leveraging the properties of residual connections to ensure lossless propagation of information during deep feature extraction. Since the transmitted component carries the water gauge's inherent geometric and color information, therefore... The feature extraction process incorporates a large amount of prior structural information from the polarization feature map, and enhances the sensitivity of the dual-domain network to structural features such as water level scale edges and geometric shapes by introducing polarization-guided gating units after the convolutional layer.

[0104] Step S302 involves performing reflection domain prediction on polarization intensity maps and polarization feature maps under various polarization angles to obtain the predicted reflection image. The core objective of reflection prediction is to predict the specular reflection component of the water surface with high fidelity. This stream adopts a similar architecture to transmission prediction, but it tends to capture bright, smooth, and large-area reflection features in feature extraction and activation function design. This stream complements the transmission prediction domain by suppressing high-frequency detail information such as water level scales through a reflection-sensitive attention mechanism, focusing on extracting the inherent reflection characteristics of water in the image.

[0105] In some embodiments, to maximize the independence and complementarity of features in the transmission prediction domain and the reflection prediction domain, this invention embeds an Attention-Guided Module (AGM) at different levels of the dual-domain network. The AGM generates a cross-stream attention mask by calculating the feature correlation between the two flows, which is used to adaptively learn and reweight important information in the feature maps. Specifically, this module ensures, through a complementary feature guidance mechanism, that reflection-related features are accurately separated and fed into the reflection prediction domain for prediction, while background structural information such as water level markings is accurately fed into the transmission prediction domain for prediction, thereby achieving efficient decoupling of transmission and reflection features. The function of the AGM can be expressed as:

[0106] ;

[0107] in, The feature map is extracted from the transmission prediction and mainly contains structural information such as the water level gauge background and scale; AGM is the feature attention guidance module. The feature map extracted from the reflection prediction stream mainly contains information on bright, smooth regions such as water surface mirror reflection; This is the optimized or reweighted output feature map after processing by the AGM module. AGM generates an attention mask by calculating the correlation between the transmission component of the water gauge background and the reflection component of the water surface reflection, thereby more accurately separating or enhancing the features of a specific flow and achieving feature decoupling.

[0108] Since water glare, as a form of specular reflection, is often a large-area, continuous artifact with long-distance dependencies, this invention integrates a self-attention mechanism into the intermediate layer of the network's feature extraction process to effectively capture and separate such non-locally correlated large-area reflections. This mechanism can calculate the correlation between any two pixels in an image, allowing the network to capture the association between distant pixels, significantly enhancing the network's ability to identify and model large-scale water surface reflection areas.

[0109] Step S303: Fuse the predicted transmission image and the predicted reflection image to obtain the reconstructed image. To achieve both physical accuracy and perceptual quality in reflection separation, a feature decoupling loss for the polarization dual-domain feature decoupling network is defined. To rebuild the losses and local structure texture consistency loss The weighted sum. The feature decoupling loss of the polarization dual-domain feature decoupling network is:

[0110] ;

[0111] ;

[0112] in, For feature decoupling loss; For reconstruction losses; Weights for local structural texture consistency; This represents a loss of local structural texture consistency. For structural weights; For structural loss; denoted as L1 norm; I is the input image; T is the predicted transmission component; and R is the predicted reflection component.

[0113] The reconstruction loss employs pixel-level L1 or L2 norms to minimize the reconstruction error between the input image I and the additive mixture of the predicted transmission and reflection components. The reconstruction loss forces the model to adhere to the physically additive reconstruction relation I=T+R, ensuring that the transmission and reflection components can accurately reconstruct the original image from an energy conservation perspective. The local structure-texture consistency loss is a key component of the network's robustness; it utilizes a local resolver focused on high-frequency details. The predicted transmission component is supervised for local regions. This loss function aims to enhance the structural consistency, sharpness, and local realism of the transmission component within the water gauge scale region, effectively removing artifacts caused by water surface glare residue in the transmission component and ensuring high-resolution output of the water gauge scale information. The structural loss is constrained using either the gradient domain information of the transmission component or the gradient domain information of the SSIM and polarization feature map. Since the polarization feature map is structurally robust, the structural loss ensures that the separated transmission component is geometrically highly consistent with the polarization physics prior, preventing distortion of the water gauge scale geometry during reflection separation.

[0114] Step S4 involves constructing a global-local collaborative heterogeneous target integration optimization framework. This framework is then used to process the transmission component to obtain the absolute water level measurement. The heterogeneous target integration optimization framework integrates multiple functional modules into an end-to-end whole, including a shared feature encoder, a task-aware decoder, and a water level calculation and output module. A drawback of traditional water level gauge identification systems is the accumulation and amplification of errors. Errors in upstream tasks (steps S1-S3) propagate and cause a decrease in the accuracy of downstream tasks (steps S402 and S403). To overcome this limitation and achieve the highest overall system robustness, this invention constructs a global-local collaborative heterogeneous target integration optimization framework. This framework aims to integrate all core functional modules, such as illumination enhancement, reflection separation, scale normalization, and water level identification, into an end-to-end whole, achieving efficient system integration and deep collaborative optimization at the feature level. The main task module for water level gauge recognition (including the localization and segmentation of water level gauge scale lines, sequential text recognition of water level gauge numbers, and accurate detection of water surface lines) is jointly optimized end-to-end with auxiliary task modules such as illumination enhancement correction, polarization reflection separation, and scale super-resolution. By defining a joint loss function, the overall objective function of the heterogeneous objective integration optimization framework, which achieves collaborative enhancement of parameters across all task modules, is composed of a weighted sum of the loss functions of all core functional modules based on their contribution to the overall task. This joint optimization mechanism ensures that the training objectives of each module compete and cooperate within the same gradient space, effectively improving the system's generalization ability. Its mathematical expression is:

[0115] ;

[0116] in, For the joint loss function; This represents the total loss of the polarization reflection separation network. The total loss of the illumination perception enhancement module, For the loss of the scale-normalized super-resolution module, Self-supervised loss for dimensional consistency awareness. This represents a loss in the final water level identification task; These are the weighting coefficients for the total loss of the polarization reflection separation network; The weighting coefficients represent the total loss of the illumination perception enhancement module; These are the weighting coefficients for the scale-normalized super-resolution module loss; The weighting coefficients of the self-supervised loss for the dimensional consistency perception loss. The weighting coefficients for the task loss in determining the final water level are determined. These coefficients are dynamically adjusted using task importance weighting or adaptive loss balancing techniques to ensure that the gradient contributions of all tasks remain balanced during the joint optimization process, avoiding the dominant or inhibitory effect of any single task on the overall optimization.

[0117] Specifically, step S4 includes:

[0118] Step S401: The shared feature encoder uses a deep convolutional network with channel attention and spatial attention to convolve the input transmission component, obtaining a shared feature representation. To promote efficient knowledge transfer and feature collaboration among all downstream tasks, the heterogeneous target integration optimization framework employs a unified and high-performance shared feature encoder. This encoder receives the high-fidelity transmission component T, which is free from reflection and illumination distortion, from the output of step S3 as input. Through a multi-layered deep convolutional network integrating channel attention and spatial attention, it extracts a high-dimensional and robust shared feature representation. This shared feature representation serves as the starting point for both the scale-normalized super-resolution in step S402 and the water level recognition and output in step S403. Through this design, the present invention achieves centralized management and global collaboration of features, enabling the advantageous features of image preprocessing to be directly and losslessly fed into high-level recognition tasks, significantly improving the feature robustness of the entire system under complex environmental changes.

[0119] Step 402: Although shared features achieve global feature sharing, different downstream tasks have drastically different priorities in their feature requirements. To resolve this potential conflict in feature requirements, the heterogeneous target ensemble optimization framework employs a task-aware decoder to locally decouple the shared feature representation, ensuring that each task can extract key information from the shared features as needed. Specifically, for the scale-normalized super-resolution task, a high-frequency detail extraction decoder is designed. This decoder focuses on high-frequency information, edges, and local textures in shared features that are crucial for image detail reconstruction, and outputs high-resolution features. A semantic sequence decoder was designed for the water level recognition task. This decoder focuses on the sequence semantics, positional and scale context information in shared features and outputs semantic sequence features. The semantic sequence decoder adopts an autoregressive or non-autoregressive structure based on Transformer. This task-specific decoder design avoids the feature degradation problem in multi-objective optimization and achieves efficient satisfaction of local feature requirements.

[0120] Specifically, step S402 includes addressing the issue of inconsistent dimensions among different water level gauges in reservoir water level gauge recognition scenarios. This invention proposes a collaborative design of a self-supervised mechanism combining scale-normalized super-resolution (SR) and dimensional consistency awareness. The goal of this mechanism is to achieve universal recognition of water level gauge images of different sizes and scale intervals, and to integrate the SR module into a heterogeneous target ensemble optimization framework as an auxiliary task for end-to-end optimization. To efficiently utilize the guiding role of high-level recognition results in low-level image detail reconstruction, this invention introduces a super-resolution decoder based on a task-guided interactive transformer architecture. The super-resolution decoder, based on the highly robust shared feature representation extracted by the shared encoder and the high-resolution features output by the high-frequency detail extraction decoder, outputs a scale-normalized high-definition feature map. To achieve feedback enhancement of the recognition task through super-resolution, a semantic sequence decoder extracts features from the shared feature representation to obtain semantic sequence features. These semantic sequence features include the sequence text recognition results, i.e., the watermark digits. via encoder This is transformed into a query signal Q. The query signal Q contains the currently identified scale number, position, and contextual information, representing the network's focus on high-frequency details of the water gauge:

[0121] ;

[0122] Where Q is the query signal; L is the sequence length. For feature dimensions; The mathematical properties of the query signal Q are defined, indicating that Q is a matrix in the real number field with dimensions of . The query signal Q interacts with shared features through a cross-modal attention module to generate task-aware super-resolution features:

[0123] ;

[0124] in, Super-resolution features; For cross-modal attention processing.

[0125] This mechanism enables super-resolution decoders Feature extraction can prioritize high-frequency detail regions that are strongly correlated with the current recognition results, thereby reconstructing the edges and textures of the water level gauge lines in a targeted manner and improving the semantic accuracy of super-resolution.

[0126] Define the scale-normalized super-resolution loss as Super-resolution decoder All network parameters The optimization objective is to minimize its corresponding total loss function. Total loss function The loss consists of three weighted components: pixel reconstruction loss, content loss, and dimensional consistency loss, which can be expressed as:

[0127] ;

[0128] in, This is the scale-normalized super-resolution loss; These are the weighting coefficients for pixel reconstruction loss; For pixel reconstruction loss; The weighting coefficient for content loss; This represents content loss. Pixel reconstruction loss employs... or Norm minimization super-resolution output image Compared with real high-resolution target images The pixel-level differences between the images are used; while the content loss is constrained by the VGG perceptual loss feature domain to limit the super-resolution output image. Compared with real high-resolution target images Similarity based on higher-order semantic features ensures high fidelity in the reconstructed image, and the formula is as follows:

[0129] ;

[0130] in, For content loss; For super-resolution output images; is the high-resolution target image; j is the layer index in the VGG network; This represents the feature mapping of the j-th layer of the VGG-19 network. These represent the number of channels, height, and width of the feature map, respectively. It is the square of the L2 norm.

[0131] To address the challenge of directly expressing the dimensional consistency constraints (such as equidistance and continuity) of water level gauge markings using supervised signals, this invention constructs a self-supervised mechanism based on partial feature reconstruction. This mechanism treats the equidistance and continuous geometric features of the water level gauge markings as linguistically perceived semantic structures, which are then used for model pre-training or joint training. A schematic diagram of the mechanism is shown below. Figure 3 As shown, the process includes original feature input, random structure masking, model prediction and reconstruction, and calculation of reconstruction loss. In the original feature input stage, shared features are input; in the random structure masking stage, the shared features are masked to obtain the masked features; in the model prediction and reconstruction stage, the encoder... and decoder The masked features are processed to obtain reconstructed features. During the reconstruction loss calculation stage, the losses of the original features and the masked regions of the reconstructed features are calculated to obtain the reconstruction loss, and the reconstruction constraint is used as a dimensional consistency constraint. Specifically, during training, shared features or non-reflective transmission components are randomly structurally masked, for example, by randomly obscuring a segment of a scale or number on a water level gauge. Encoder and decoder The task is to use the surrounding unmasked scale features to predict and reconstruct the scale details and numerical features of the masked portion. This prediction mechanism based on geometric context forces the network to learn the inherent equidistant and continuous geometric regularities of the water gauge scale, thereby implicitly embedding these dimensional consistency constraints into the model's feature representation.

[0132] The dimensional consistency constraint uses the reconstruction error of the masked region as the loss to minimize the reconstructed scale features. Compared with the original scale features The distance between them. This ensures that regardless of the actual dimensions or resolution of the water gauge, the model can learn and follow the equidistant geometric relationship between the scale lines, significantly improving the model's cross-scale universality and generalization ability for water gauges of different dimensions. The process of constructing the dimension consistency constraint can be expressed as:

[0133] ;

[0134] in, The loss is dimensional consistency; M is the set of all masking regions, and m is one of the masking regions. It is a weight calculated based on the scale space position, used to dynamically adjust the loss contribution of different scale regions; The reconstructed scale features; This represents the original scale features; It is an L1 norm.

[0135] Step 403: Calculate and output the water level based on super-resolution features and semantic sequence features to obtain the absolute water level measurement. Step S403 includes:

[0136] This step aims to efficiently utilize the main task module within the heterogeneous target ensemble optimization framework. The precise output, especially from water surface line precision detection tasks. waterline position And sequence text recognition tasks from water level gauge numbers Scale number sequence First, the water level detection module. Based on shared feature representation, a high-precision segmentation network is used to output an accurate mask or contour of the water surface line pixels within the water gauge region. The pixel positions of the water mask and the water level gauge line are obtained. Meanwhile, the digital recognition module The pixel positions of the water level gauge markings are identified, and a series of water level numbers corresponding to the markings and their geometric positions in the image are output. These number sequences A reference standard for water level measurement is provided. (Obtain) and Then, the system first utilizes Provided geometric position Determine the pixel position of the water level gauge scale line The most recent, identified, and most reliable reference scale. Next, based on the dimensional consistency constraint... This constraint ensures that the model has the ability to perceive the geometric spacing of different water level gauge markings, and the system calculates the pixel positions of the water level gauge marking lines. Pixel position and baseline pixel position The precise pixel distance ΔP between the water surface and the known absolute water level reference scale. This pixel distance ΔP represents the relative position of the water surface with respect to the known absolute water level reference scale and is crucial for achieving accurate measurement. Due to its characteristics... After scale-normalized super-resolution processing, the calculation of ΔP has high accuracy and cross-scale robustness.

[0137] Finally, the calculated pixel distance ΔP is converted into an absolute water level measurement. The conversion process requires the use of the known geometric parameters of the water gauge, namely the actual physical spacing of the water gauge's graduation lines. (e.g., 1 cm or 1 decimeter) and the current pixel-to-physical scale ratio The scale ratio It can be done The isometric prior learned during training, or the pixel distance between two adjacent identified tick marks, and The ratio is dynamically estimated. The absolute water level measurement of the water surface line. The calculation formula is as follows:

[0138] ;

[0139] in, It is the baseline scale line. The corresponding absolute water level value. (From...) The sequence is provided directly. The sign depends on the waterline. Relative to the reference scale line The high and low positions. This calculation method effectively utilizes high-level semantics (digit recognition). ) and underlying geometric constraints (ΔP and This enables high-precision, multidimensional automated water level measurement in complex environments, and ultimately outputs the current water level measurement results.

[0140] To address the challenges of gradient conflicts and inconsistent task learning rates in heterogeneous objective ensemble optimization, this invention introduces an Adaptive Gradient Routing (AGR) mechanism. The core of the AGR mechanism lies in dynamically monitoring the losses of different tasks. Relative to shared parameters gradient direction In each training iteration, the gradient directions of the two tasks conflict, which can be represented as:

[0141] ;

[0142] in, and Loss functions for different tasks; It is a small positive number. When the AGR detects a collision, it can correct the conflicting gradients by projecting the gradient of one task onto the orthogonal space of the gradients of another task, thereby ensuring the shared encoder. The optimized update direction will not lead to any performance degradation in any task. This mechanism ensures the stability and convergence of the system training process, and solves the problems of unstable training and easy getting trapped in local optima in traditional heterogeneous target ensemble optimization.

[0143] By constructing the aforementioned heterogeneous target integration optimization framework, this invention achieves closed-loop optimization from low-level image preprocessing to high-level semantic recognition. Specifically, this is reflected in: the non-reflective transmission component provided by S3. Serving as high-quality input to S401 directly improves the upper limit of super-resolution sharpness. Scale-normalized super-resolution features significantly reduce the difficulty and uncertainty of the recognition task. Simultaneously, the final recognition feedback can serve as a high-level semantic constraint, guiding super-resolution detail reconstruction and S3 reflection separation through backpropagation. This mechanism enables high-level semantic guidance for low-level feature correction, exhibiting higher robustness compared to traditional serial architectures.

[0144] Figure 4 and Figure 5 The results of recognizing actual reservoir water level images using the method of the present invention are demonstrated. Among them, Figure 4 The original image to be detected. Figure 5 This is the recognition result image of the method of the present invention. It can be seen that... Figure 4 The lighting in the Central Plains image was weak, and the water surface reflected images, which interfered with the recognition. Meanwhile... Figure 5Although a small amount of residual reflection remains on the water surface, the method of this invention can accurately eliminate reflection interference, effectively weaken the impact of water surface reflection, and significantly enhance the light intensity on the basis of the original image, ultimately accurately identifying the dimensions of the reservoir water gauge and the water level measurement results.

[0145] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for reservoir water level identification based on polarization imaging and heterogeneous target integration, characterized in that, include: S1: Based on the collected polarization intensity maps at various polarization angles, obtain the optical wave state feature matrix and polarization feature map of the reservoir water gauge; S2: Based on the polarization feature map, the image to be detected of the reservoir water gauge is subjected to self-correcting illumination sensing enhancement preprocessing to obtain an enhanced image and a corrected image. A preprocessing loss function is then constructed based on the enhanced and corrected images. The preprocessing loss function is: ； in, To pre-process losses; Enhance weighting for low light conditions; To enhance loss in low light; Weighting for strong light correction; To correct for loss due to strong light; Preserve weights for geometric contours; Loss is preserved for geometric contours; The low-light enhancement loss of the low-light enhancement module is: ； in, To enhance loss in low light; For mapping weights; For mapping loss; For multi-level semantic feature consistency weights; For multi-level semantic feature consistency loss; The weights are the light intensity gradient constraint weights. Loss is due to light intensity gradient constraints. The intensity correction loss of the intensity correction module is: ； in, To correct for loss due to strong light; This is the first correction weight; To correct the image; The target image is represented by k, which is the scale variable. It is an L1 norm; This is the second correction weight; Features extracted at the k-th scale; The geometric profile preservation loss is: ； in, The loss is used to preserve the geometric contour; SSIM() is the structural similarity index. The image is a preprocessed image, including an enhanced image and a corrected image; P is a polarization feature map; S3: Construct a polarization dual-domain feature decoupling network. This network decouples polarization intensity maps and polarization feature maps at various polarization angles, yielding transmission and reflection components. Based on these components, a feature decoupling loss is constructed. The polarization dual-domain feature decoupling network includes a transmission prediction domain and a reflection prediction domain. These domains guide each other through cross-domain feature decoupling, generating transmission and reflection components respectively. The mapping relationship of the polarization dual-domain feature decoupling network is as follows: ； Where T is the transmission component of the water level background; R is the reflection component of the water surface reflection. This is a polarization dual-domain feature decoupling network; I is the input image; P is the polarization feature map; Cross-domain feature decoupling guidance is achieved by embedding feature attention guidance modules at different levels of the polarization dual-domain feature decoupling network; the function of the feature attention guidance module is as follows: ； in, The feature map is extracted from the transmission prediction; AGM is the feature attention guidance module. Feature maps extracted from the reflection prediction stream; This is the optimized or reweighted output feature map after processing by the AGM module; The feature decoupling loss is: ； ； in, For feature decoupling loss; For reconstruction losses; Weights for local structural texture consistency; This represents a loss of local structural texture consistency. For structural weights; For structural loss; is the L1 norm; I is the input image; T is the predicted transmission component; R is the predicted reflection component; S4: Construct a heterogeneous target integrated optimization framework, and process the transmission component through the framework to obtain the absolute water level measurement value. The heterogeneous target integrated optimization framework includes a shared feature encoder, a task-aware decoder, and a water level calculation and output module. The framework performs end-to-end joint optimization of the main task module and auxiliary task module for water level gauge recognition. The main task module includes the localization and segmentation of water level gauge scale lines, the sequential text recognition of water level gauge numbers, and the detection of water surface lines. The auxiliary task modules include illumination enhancement correction, polarization reflection separation, and scale super-resolution. The mathematical expression of the heterogeneous target integrated optimization framework is: ； in, For the joint loss function; This represents the total loss of the polarization reflection separation network. The total loss of the illumination perception enhancement module, For the loss of the scale-normalized super-resolution module, Self-supervised loss for dimensional consistency awareness. This represents a loss in the final water level identification task; These are the weighting coefficients for the total loss of the polarization reflection separation network; The weighting coefficients represent the total loss of the illumination perception enhancement module; These are the weighting coefficients for the scale-normalized super-resolution module loss; The weighting coefficients of the self-supervised loss for the dimensional consistency perception loss. The weighting coefficients are used to calculate the loss in the final water level identification task.

2. The reservoir water level gauge identification method based on polarization imaging and heterogeneous target integration according to claim 1, characterized in that, Step S1 includes: S101: A polarization intensity acquisition device is used to collect polarization intensity maps under multiple polarization angles; multiple polarization angles equal four types; S102: Determine the optical wave state feature matrix based on the sum and difference relationship between polarization intensity maps under different polarization angles; the optical wave state feature matrix includes multiple optical wave state feature components. S103: Determine the polarization feature map based on multiple light wave state characteristic components; the polarization feature map includes the light polarization purity index and the polarization principal axis direction.

3. The reservoir water level gauge identification method based on polarization imaging and heterogeneous target integration according to claim 2, characterized in that, When there are four different polarization angles, the characteristic matrix of the light wave state is: ； ； ； ； Where M is the light wave state characteristic matrix; , and These represent the characteristic components of the first, second, and third light wave states, respectively; T is the transpose of the matrix. , , and These are the first, second, third, and fourth polarization intensity components, respectively. The polarization angle of the first polarization intensity component is... The polarization angle of the second polarization intensity component is The polarization angle of the third polarization intensity component is The polarization angle of the fourth polarization intensity component is ; The light polarization purity index is: ； The direction of the principal polarization axis is: ； Where Π is the polarization purity index of the light ray; Φ is the direction of the principal polarization axis; and atan2 is the arctangent function in the four quadrants.

4. The reservoir water level gauge identification method based on polarization imaging and heterogeneous target integration according to claim 1, characterized in that, Step S2 includes: S201: Determine the image type of the image to be detected; image types include low-light underexposed images and strong-light overexposed images. S202: For underexposed images in low light, a self-supervised low light enhancement module is used to enhance the underexposed images in low light to obtain an enhanced image. S203: For overexposed images in strong light, the strong light correction module performs saturation correction and detail restoration on the overexposed images in strong light to obtain a corrected image.

5. The reservoir water level gauge identification method based on polarization imaging and heterogeneous target integration according to claim 4, characterized in that, The low-light enhancement module employs a generative adversarial network. The generator of the low-light enhancement module adopts a U-Net structure and integrates non-local attention blocks to predict illumination maps and convert low-light underexposed images into enhanced images. The discriminator of the low-light enhancement module includes a panoramic discriminator and one or more focus discriminators. The panorama discriminator is used to evaluate the overall brightness distribution and panorama realism of the enhanced image; A focus discriminator is used to enhance the local texture and contrast of the water gauge scale area; The strong light correction module adopts a frequency domain geometric information heterogeneous decoupling framework, which includes a deep neural network, a joint correction module, and a detail restoration network. The deep neural network is used to decouple the strong light overexposed image into a low-frequency basal layer and a high-frequency detail layer. The joint correction module is used to jointly correct the saturation component of the low-frequency basal layer to obtain the corrected basal layer. The detail restoration network is used to fuse the high-frequency detail layer with the corrected basal layer to obtain the corrected image.

6. The reservoir water level gauge identification method based on polarization imaging and heterogeneous target integration according to claim 5, characterized in that, The mapping loss, multi-level semantic feature consistency loss, and light intensity gradient constraint loss are respectively: ； ； ； in, For mapping loss; is the expected value; log is the logarithmic function; The discrimination result of the panoramic discriminator; To enhance the image; To focus on the discrimination results of the discriminator; This represents the multi-level semantic feature consistency loss; i is the feature layer-level index variable. Let be the total number of elements in the feature map of the i-th layer; Here is the feature map extracted by network Φ at layer i; This is an underexposed image in low light. It is an L1 norm; The loss is defined by the light intensity gradient constraint; x and y are the x and y coordinates of the pixel, respectively. and These are the horizontal and vertical gradient operations at pixel (x, y), respectively. This is the enhanced image at pixel (x,y); Let G be the L2 norm; G is the generator network.

7. The reservoir water level gauge identification method based on polarization imaging and heterogeneous target integration according to claim 1, characterized in that, Step S3 includes: S301: Perform transmission domain prediction on polarization intensity maps and polarization feature maps under multiple polarization angles to obtain transmission components; S302: Perform reflection domain prediction on polarization intensity maps and polarization feature maps under multiple polarization angles to obtain reflection components; S303: Fuse the transmission and reflection components to obtain the reconstructed image.

8. The reservoir water level gauge identification method based on polarization imaging and heterogeneous target integration according to claim 1, characterized in that, Step S4 includes: S401: The shared feature encoder uses a deep convolutional network with channel attention and spatial attention to convolve the transmission components of the input to obtain a shared feature representation; S402: Decouple the shared feature representation through the task-aware decoder to obtain super-resolution features and semantic sequence features; the semantic sequence features include sequence semantics, position and scale context information; S403: The water level calculation and output module calculates the absolute water level measurement value by performing water level calculation on super-resolution features and semantic sequence features.

9. The reservoir water level gauge identification method based on polarization imaging and heterogeneous target integration according to claim 8, characterized in that, Step S402 includes: High-resolution features are obtained by extracting features from the shared feature representation using a high-resolution decoder; high-resolution features include high-frequency information, edges, and local textures. The shared feature representation and high-resolution features are processed to obtain a high-definition feature map; The semantic sequence features are obtained by extracting features from the shared feature representation using a semantic sequence decoder. The semantic sequence features are transformed to obtain the query signal; Super-resolution features are generated based on the query signal.

10. The reservoir water level gauge identification method based on polarization imaging and heterogeneous target integration according to claim 8, characterized in that, Step S403 includes: The water surface line within the water gauge region in the shared feature representation is segmented to obtain the pixel positions of the water body mask and the water gauge scale line. Identify the pixel positions of the water level gauge markings to determine the numerical and geometric positions of the water level readings. Determine the baseline graduation line based on the geometric position of the water gauge graduation line; Determine the pixel distance between the pixel positions of the baseline scale line and the water gauge scale line; The pixel distance is converted to obtain the absolute water level measurement value.

Images