Sewage component identification analysis method and system based on spectrum and image recognition

Abstract

The present application relates to the field of environmental monitoring, and specifically discloses a sewage component identification analysis method and system based on spectrum and image recognition, which comprises the following steps: collecting a multi-polarization angle image sequence, identifying a surface tension disturbance area, marking an oil pollution priori area, fusing and dividing a region of interest, and adaptively selecting a hyperspectral band subset for collection; outputting a region of interest sequence image and an adaptive band spectral data body; constructing a ternary coupling mapping recognition characteristic overlapping micro area, implementing cross-modal separation, and outputting a pollution component fingerprint feature set; inputting the fingerprint feature set into a hierarchical recognition framework for macro-class discrimination and subdivision identification, generating a dynamic confidence weight, and outputting a recognition result with a confidence index; and the system comprises functionally matched modules. The present application significantly improves the recognition accuracy and robustness of multi-component pollution components in complex sewage environments through polarization image recognition of water surface physical property evolution, interference texture priori guidance, and physical constraint decoupling.

Description

Technical Field

[0001] This invention relates to the field of environmental monitoring technology, and in particular to a wastewater component identification and analysis method and system based on spectral and image recognition. Background Technology

[0002] Current wastewater component detection technologies mainly involve chemical titration, single-spectral analysis, and colorimetric identification based on visible light images. Chemical titration offers high accuracy but suffers from poor timeliness, making it unsuitable for online monitoring. Single-spectral analysis typically involves indiscriminate acquisition across the entire field of view, failing to consider the spatiotemporal heterogeneity of surface contaminant distribution in wastewater, resulting in a large amount of invalid background spectral data and a low signal-to-noise ratio. Visible light image-based identification methods utilize apparent features such as color and turbidity for discrimination, but cannot acquire deep chemical information and lack effective utilization of complex optical phenomena such as oil film interference colors and microbubble scattering.

[0003] In recent years, spectral and image fusion recognition technology has been gradually applied to the field of wastewater monitoring. However, existing fusion methods generally suffer from the following deep-seated technical defects. First, existing technologies mostly use inter-frame difference or background subtraction for moving target detection, failing to utilize polarized image sequences to capture changes in surface tension caused by pollution diffusion. These surface tension changes significantly alter the polarization characteristics of capillary waves on the water surface, a physical mechanism that has not been explored in existing image recognition methods. Second, thin films formed by oil pollutants on the wastewater surface produce interference color fringes, the spatial frequency distribution of which is directly related to the oil film thickness and refractive index. However, existing image recognition technologies typically treat interference colors as noise and filter them out, failing to transform them into prior identification markers for oil pollution. Third, multiple pollutants overlap to varying degrees in both the spectral feature space and the polarization texture feature space. Existing technologies mostly use simple splicing or weighted fusion methods for feature combination, lacking a feature decoupling mechanism based on fluid dynamics physical constraints. Fourth, existing technologies lack quantitative assessments of the confidence level of the fusion's physical consistency in the recognition results. When wastewater is in a low-concentration or strongly mixed state, misidentification results cannot be effectively identified and corrected.

[0004] Therefore, there is an urgent need for a wastewater component identification and analysis method that can identify surface tension disturbances on the water surface using polarization image sequences, guide adaptive spectral acquisition using interference texture priors, achieve cross-modal feature decoupling based on surface tension gradient constraints, and possess dual-path feedback optimization capabilities. Summary of the Invention

[0005] The purpose of this invention is to address the problems in existing wastewater monitoring technologies, such as the inability to use polarization images to identify surface tension disturbances and oil film interference prior information, the difficulty in physically constraining and decoupling the overlap of mixed pollutants in the spectral and image feature spaces, and the insufficient identification accuracy and robustness caused by the redundancy of full-field spectral acquisition data. The invention proposes a wastewater component identification and analysis method and system based on spectral and image recognition.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] The wastewater component identification and analysis method based on spectral and image recognition includes the following specific steps:

[0008] S1. Acquire multi-polarization angle image sequences, calculate the polarization degree time-series evolution matrix to identify surface tension disturbance regions, extract the spatial frequency markers of oil film interference fringes to indicate the prior regions of oil; fuse and divide the region of interest, adaptively select a subset of hyperspectral bands for acquisition, and output the sequence images of the region of interest and the adaptive band spectral data volume;

[0009] S2. Divide the sequence image of the region of interest and the adaptive band spectral data volume into micro-regions, and extract polarization time evolution, interference texture and spectral absorption feature vectors; construct a ternary coupling mapping to identify overlapping micro-regions of features, implement cross-modal separation based on surface tension gradient constraints, and output a fingerprint feature set of contaminant components;

[0010] S3. Input the pollutant component fingerprint feature set into the hierarchical identification framework, perform major category discrimination and sub-category identification, and generate dynamic confidence weights; when the confidence is low, trigger the first feedback path to adjust the surface tension gradient constraint and re-decouple, and trigger the second feedback path to optimize the polarization angle sequence and band selection strategy and re-execute, and output the wastewater component identification result with confidence index.

[0011] As a further technical solution of the present invention, in S1, acquiring multi-polarization angle image sequences and calculating the polarization degree temporal evolution matrix to identify surface tension disturbance regions specifically includes:

[0012] S111. Place the wastewater sample to be tested in a transparent sample container and position it at the polarization image acquisition station. Control the polarizer to rotate step by step according to a preset angle sequence. At each polarization angle, continuously acquire multiple frames of images at a preset time interval. Perform geometric correction, illumination normalization and noise filtering preprocessing on the time sequence images to obtain a multi-polarization angle image sequence.

[0013] S112. Perform pixel-level spatial registration on the images with different polarization angles at the same acquisition time node in the multi-polarization angle image sequence so that the pixels of each polarization angle image correspond precisely in spatial position. Extract the light intensity response value of each pixel at different polarization angles, construct the Stokes vector parameters of the pixel and calculate the polarization degree value of the pixel at the current acquisition time node.

[0014] S113. Arrange the polarization degree values ​​of each pixel within the preset acquisition time window in time sequence to form a polarization degree time sequence evolution curve, construct a polarization degree time sequence evolution matrix according to spatial location, perform time sequence difference analysis on each time sequence evolution curve in the matrix, and calculate the change amount and rate of change characteristics of the polarization degree value of each pixel between adjacent acquisition time nodes.

[0015] S114. Pixels whose rate of change exceeds a preset rate of change threshold are identified as pixels affected by surface tension disturbance. The continuous spatial region occupied by the affected pixels is marked as the surface tension disturbance region, and its spatial boundary position information and the distribution information of the polarization degree change rate of each pixel in the region are recorded.

[0016] As a further technical solution of the present invention, in S1, extracting the spatial frequency marker of the oil film interference fringes to indicate the prior region of the oil specifically includes:

[0017] S121. Select the polarization angle image frame with the highest signal-to-noise ratio from the multi-polarization angle image sequence as the interferometric analysis image, perform grayscale processing on the interferometric analysis image and extract the grayscale value distribution information of each pixel in the image;

[0018] S122. The interferometric analysis image is divided into several local image blocks by sliding window processing. A two-dimensional Fourier transform is performed on each local image block to convert it from the spatial domain to the frequency domain, thereby obtaining a spectral amplitude distribution map. The amplitude value corresponding to each frequency component in the spectral amplitude distribution map represents the intensity contribution of the interference fringes at that spatial frequency. The amplitude peak position is searched in each spectral amplitude distribution map and the corresponding spatial frequency coordinate value is extracted.

[0019] S123. Determine the direction and magnitude of the main frequency of the interference fringes based on the spatial frequency coordinates, and compare them with the standard interference frequency features of each type of oil in the preset oil film interference frequency feature library. When the deviations are all within the preset deviation range, it is determined that there is oil contamination in the spatial region corresponding to the local image block.

[0020] S124. Extract and connect the boundary contours of the spatial regions corresponding to the local image blocks where oil pollution is determined to exist. Mark the continuous regions where the uniformity of the spatial frequency distribution of interference fringes meets the preset uniformity conditions as the prior regions of oil pollution, and record their boundary position information, main frequency direction information and main frequency magnitude information.

[0021] As a further technical solution of the present invention, in S1, the process of fusing and dividing the region of interest, adaptively selecting a subset of hyperspectral bands for acquisition, and outputting a sequence image of the region of interest and an adaptive band spectral data volume specifically includes:

[0022] S131. Map the boundary position information of the surface tension disturbance region and the boundary position information of the oil contamination prior region to the same spatial coordinate system, calculate the ratio of the intersection area to the union area of ​​the two regions as the spatial overlap, and calculate the minimum distance between the boundaries of the two regions as the spatial proximity.

[0023] S132. When the spatial overlap exceeds a preset overlap threshold or the spatial proximity is less than a preset proximity threshold, the two regions are merged into a fused region of interest and the minimum convex hull polygon of the two regions is used as the boundary; otherwise, the surface tension disturbance region and the oil contamination prior region are respectively used as independent regions of interest.

[0024] S133. For each region of interest, determine the spatial sampling density or scanning step size of hyperspectral acquisition based on its spatial area size, and determine the oil characteristic bands and interface characteristic bands based on their degree of overlap with the prior region of oil contamination and the surface tension disturbance region, respectively. After merging and deduplication, these are used as the adaptive hyperspectral band subset of the region of interest.

[0025] S134. Collect hyperspectral image data according to the spatial sampling density or scanning step size and adaptive hyperspectral band subset corresponding to each region of interest. Package the sequence image data and hyperspectral image data of each region of interest according to the identifier, and output the sequence image of the region of interest marked with location information and identifier information and the corresponding adaptive band spectral data volume.

[0026] As a further technical solution of the present invention, in step S2, the sequence image of the region of interest and the adaptive band spectral data volume are divided into micro-regions, and polarization time evolution, interference texture and spectral absorption feature vectors are extracted, specifically including:

[0027] S211. Register the sequence image of the region of interest and the adaptive band spectral data volume at the pixel level according to the spatial position correspondence. Take each pixel in each region of interest as a micro-region unit, and align the polarization time sequence image data and spectral data of each micro-region unit to complete the cross-modal data association of each micro-region.

[0028] S212. Perform time-series feature extraction on the polarization time-series image data of each micro-region unit to obtain the polarization time-series evolution feature vector. Perform interference texture analysis on the sequence image of the region of interest to extract the spatial frequency and direction features of the interference fringes and form the interference texture feature vector. Perform spectral feature extraction on the adaptive band spectral data to obtain the spectral absorption feature vector.

[0029] As a further technical solution of the present invention, in step S2, constructing a ternary coupling mapping to identify overlapping micro-regions of features specifically includes:

[0030] S221. For each micro-region unit, the polarization time-series evolution feature vector, interference texture feature vector, and spectral absorption feature vector are concatenated according to a unified feature dimension index to form the initial ternary coupling feature vector of each micro-region unit;

[0031] S222. Calculate the first, second, and third correlation coefficients between the polarization time-series evolution feature vector and the spectral absorption feature vector, the interference texture feature vector and the spectral absorption feature vector, and the polarization time-series evolution feature vector and the interference texture feature vector of each micro-region unit, and record the values ​​of the three correlation coefficients of the micro-region unit.

[0032] S223. When the absolute values ​​of at least two of the three correlation coefficients exceed the preset correlation threshold at the same time, it is determined that the micro-region unit has multimodal feature coupling interference caused by mixed pollution, and it is marked as a feature overlapping micro-region and its spatial location coordinates are recorded.

[0033] As a further technical solution of the present invention, in S2, cross-modal separation is performed based on surface tension gradient constraints to output a fingerprint feature set of contaminant components, specifically including:

[0034] S231. For each feature-overlapping micro-region, the numerical distribution information of the polarization degree change rate is used as a proxy index of the surface tension disturbance intensity. The surface tension disturbance intensity values ​​of the micro-region and its adjacent micro-regions at each acquisition time node are extracted. The surface tension gradient components of the micro-region in two orthogonal directions are calculated based on its spatial position change rate and combined to form a surface tension gradient vector.

[0035] S232. Using the surface tension gradient vector as a physical constraint term, a joint separation objective function for the spectral-polarization-interference ternary coupling characteristics is established, including separation sub-terms for spectral absorption characteristics, polarization temporal evolution characteristics, and interference texture characteristics, as well as a surface tension gradient constraint regularization term. The weighted summation of each separation sub-term and the linear combination of the regularization term constitute the joint separation objective function.

[0036] S233. Set the contribution components of each separation sub-item in the joint separation objective function as variables to be solved. By minimizing and optimizing the joint separation objective function, obtain the values ​​of the contribution components of each pollutant in the spectral absorption feature vector, polarization time evolution feature vector, and interference texture feature vector.

[0037] S234. The spectral absorption contribution component, polarization time evolution contribution component, and interference texture contribution component corresponding to each pollutant component obtained by the solution are classified and combined according to the component type to form the independent fingerprint features of each pollutant component. The summaries are output as the decoupled fingerprint feature set of pollutant components.

[0038] As a further technical solution of the present invention, in step S3, the fingerprint feature set of the pollutant components is input into the hierarchical recognition framework for major category discrimination and sub-category recognition to generate dynamic confidence weights, specifically including:

[0039] S311. The spectral absorption contribution components corresponding to each pollutant component in each micro-region of the pollutant component fingerprint feature set are arranged according to the component type to form a spectral absorption feature vector. The spectral absorption feature vector is then input into the upper classifier of the hierarchical recognition framework. The upper classifier performs a major classification of the pollutant type based on the spectral absorption features and outputs the major pollution category label corresponding to each micro-region.

[0040] S312. Input the polarization time-series evolution feature vector and interference texture feature vector corresponding to each micro-region belonging to the same pollution category label into the lower-level classifier of the hierarchical recognition framework. The lower-level classifier subdivides and identifies specific pollution types under the same pollution category based on the polarization time-series evolution features and interference texture features, and outputs the specific pollution type label of each micro-region.

[0041] S313. Calculate the spectral signal-to-noise ratio index based on the spectral absorption feature vector of each micro-region, calculate the cross-modal consistency index based on the correlation between the polarization time evolution feature vector, the interference texture feature vector and the spectral absorption feature vector, calculate the surface tension consistency index based on the matching degree between the surface tension gradient vector and the preset standard gradient mode, and generate a dynamic confidence weight by weighting the three together.

[0042] As a further technical solution of the present invention, in S3, when the confidence level is low, the first feedback path is triggered to adjust the surface tension gradient constraint and re-decouple, while the second feedback path is triggered to optimize the polarization angle sequence and band selection strategy and re-execute, outputting the wastewater component identification result with confidence index, specifically including:

[0043] S321. When the dynamic confidence weight of a certain micro-region is lower than the preset confidence threshold, the micro-region is marked as a low confidence micro-region, and the specific values ​​of the spectral signal-to-noise ratio index, cross-modal consistency index and surface tension consistency index of the low confidence micro-region are extracted.

[0044] S322. Based on the distribution of the spectral signal-to-noise ratio index and the cross-modal consistency index, the local maxima in the neighborhood of the band below the preset signal-to-noise ratio threshold are used as the positions of the supplementary bands. The angle positions to be encrypted are determined according to the deviation direction of the cross-modal consistency index and the surface tension consistency index. Based on this, the polarization angle sequence and the hyperspectral band selection strategy are updated, and the acquisition, feature decoupling and recognition steps are re-executed.

[0045] S323. Based on the degree of deviation between the cross-modal consistency index and the preset standard value, determine the adjustment direction and adjustment range of the weight coefficient of the surface tension gradient constraint term in the joint separation objective function, and re-execute the cross-modal feature separation and recognition steps based on the adjusted weight coefficient to obtain the corrected fingerprint feature set of contaminant components;

[0046] S324. Summarize the pollution category labels, specific pollution type labels, and corresponding dynamic confidence weights of all micro-areas, generate wastewater component identification results with confidence index, and output them.

[0047] A wastewater component identification and analysis system based on spectral and image recognition is used to implement a wastewater component identification and analysis method based on spectral and image recognition, including:

[0048] The polarization-spectral adaptive acquisition module is used to acquire multi-polarization angle image sequences, calculate the polarization degree time-series evolution matrix to identify surface tension disturbance regions, extract the spatial frequency markers of oil film interference fringes to mark the prior regions of oil, fuse and divide regions of interest, adaptively select hyperspectral band subsets for acquisition, and output the sequence images of regions of interest and adaptive band spectral data volumes.

[0049] The micro-region coupling and decoupling module is used to divide the sequence image of the region of interest and the adaptive band spectral data volume into micro-regions, extract polarization time evolution, interference texture and spectral absorption feature vectors, construct a ternary coupling mapping to identify overlapping micro-regions, implement cross-modal separation based on surface tension gradient constraints, and output a fingerprint feature set of contaminant components.

[0050] The hierarchical identification and dual-path feedback module is used to input the fingerprint feature set of the pollutant components into the hierarchical identification framework for major category discrimination and sub-category identification, generate dynamic confidence weights, and trigger the first feedback path to adjust the surface tension gradient constraint in the micro-region coupling and decoupling module and re-decouple when the confidence is low. At the same time, it triggers the second feedback path to optimize the polarization angle sequence and band selection strategy in the polarization-spectrum adaptive acquisition module and re-execute it, outputting the wastewater component identification result with confidence index.

[0051] The beneficial effects of this invention are as follows:

[0052] 1. By identifying surface tension disturbances on the water surface through the polarization degree time-series evolution matrix, the spatial frequency of oil film interference fringes is converted into a priori markers of oil pollution. Based on this, a subset of hyperspectral bands is adaptively selected, avoiding data redundancy from indiscriminate acquisition across the entire field of view, and significantly improving the sensitivity of pollution event detection and the signal-to-noise ratio of spectral data.

[0053] 2. By constructing a micro-region spectral-polarization-interference ternary coupling map and introducing the surface tension gradient vector as a physical constraint term to implement cross-modal feature separation, the overlapping information of mixed pollutants in the heterogeneous feature space is decoupled, realizing the independent extraction of pollutant component fingerprints at the micro-region scale and improving the accuracy of mixed pollution identification.

[0054] 3. By fusing dynamic confidence weights of spectral signal-to-noise ratio, cross-modal consistency, and surface tension consistency, dual-path feedback optimization is triggered for low-confidence results. The decoupling constraint strength and front-end acquisition strategy are adjusted respectively to form a closed-loop adaptive evolution mechanism, which significantly reduces the false recognition rate in low-concentration and strongly mixed scenarios. Attached Figure Description

[0055] Figure 1 This is a flowchart of the method proposed in Embodiment 1 of the present invention;

[0056] Figure 2 This is a framework diagram of the system proposed in Embodiment 2 of the present invention. Detailed Implementation

[0057] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0058] Example 1

[0059] Please see the appendix Figure 1 The wastewater component identification and analysis method based on spectral and image recognition specifically includes the following steps:

[0060] S1. Acquire multi-polarization angle image sequences, calculate the polarization degree temporal evolution matrix to identify surface tension disturbance regions, and extract the spatial frequency markers of oil film interference fringes to indicate the prior regions of oil; fuse and divide the region of interest, adaptively select a subset of hyperspectral bands for acquisition, and output the sequence images of the region of interest and the adaptive band spectral data volume; specifically including:

[0061] S11. Construction of the polarization degree temporal evolution matrix and identification of surface tension perturbation regions, specifically including:

[0062] S111. Inject the wastewater sample to be tested into a transparent sample container with a light transmittance of not less than 90%. The liquid level should be controlled within 1 / 2 to 3 / 4 of the container depth to avoid interference between the liquid surface and the reflected light from the bottom of the container. Position the container containing the sample at the polarization image acquisition station, which consists of an area array industrial camera, a motorized rotating polarizer support, and a uniformly diffused light source. The incident direction of the light source is at an angle of 30° to 60° with the optical axis of the camera. Control the polarizer to rotate stepwise according to a preset angle sequence. The basic step interval is between 10° and 30°. In addition to this basic step sequence, four additional angle positions of 0°, 45°, 90°, and 135° must be added to ensure the completeness of the Stokes vector construction. The preferred implementation is a basic step interval of 30°, with additional acquisitions at 45° and 135° positions, i.e., 0°, 30°, 45°, 60°, 90°, 120°, 135°, and 150°. At each polarization angle, multiple frames of images are continuously acquired at a preset time interval, the preset time interval ranging from 0.5s to 5s, with a preferred value of 1.0s.

[0063] After acquisition, the time-series images are preprocessed sequentially by geometric correction, illumination normalization, and noise filtering. Geometric correction adopts a joint correction model for radial and tangential distortion based on checkerboard calibration. Illumination normalization uses homomorphic filtering or the Retinex algorithm to suppress non-uniform illumination. Noise filtering uses Gaussian filtering or nonlocal mean denoising algorithm. The above preprocessing algorithms are all conventional existing techniques in the field of image processing. Finally, a multi-polarization angle image sequence is obtained.

[0064] S112. Pixel-level spatial registration is performed on images with different polarization angles at the same acquisition time node in the multi-polarization angle image sequence. The 0° polarization angle image is selected as the reference frame. For the remaining polarization angle images, a scale-invariant feature transform algorithm is used to extract feature points. This algorithm constructs a Gaussian difference scale space to detect local extrema, removes low-contrast and edge response points, assigns principal directions to stable feature points, and generates 128-dimensional feature descriptors. The feature descriptors of the reference frame and the remaining images are matched using Euclidean distance, and the ratio of the nearest neighbor distance to the second nearest neighbor distance is calculated, i.e.: In the formula: This is the ratio of the Euclidean distance between the nearest neighbor and the second nearest neighbor feature descriptors. The nearest neighbor feature descriptor is the Euclidean distance. The Euclidean distance is the feature descriptor for the second nearest neighbor; when the ratio Points less than a preset ratio threshold are retained as valid matching points. The preset ratio threshold ranges from 0.5 to 0.8, with a preferred value of 0.6.

[0065] False matches are eliminated using a random sampling consensus algorithm. Specifically, four sets of corresponding points are randomly selected from the set of matching points to calculate the homography transformation matrix, i.e.: In the formula: The homogeneous coordinates of the feature points in the reference frame. These are the homogeneous coordinates of the corresponding feature points in the image to be registered after transformation. for Homography transformation matrix; substitute all matching points into this matrix to calculate the projection error, i.e.: In the formula: For projection error, The homogeneous coordinates of the corresponding feature points actually detected in the image to be registered; statistical projection error. Matching points smaller than a preset error threshold are used as inliers. After repeating the iteration a preset number of times, the model with the most inliers is selected as the optimal transformation model. The preset number of iterations ranges from 500 to 5000, with a preferred value of 2000. The preset projection error threshold ranges from 1 to 5 pixels, with a preferred value of 3 pixels. The remaining images are mapped to the reference frame coordinate system using the optimal homography transformation matrix, ensuring accurate correspondence between image pixels at each polarization angle, with a registration error of less than 0.5 pixels.

[0066] After registration, the light intensity response value of each pixel at different polarization angles is extracted, i.e., the grayscale value of that pixel in the image at each polarization angle. Based on the extracted light intensity response values, the Stokes vector parameters of the pixel are constructed. The Stokes vector contains three components: total light intensity, the difference between horizontal and vertical polarization light intensity, and the difference between 45° and 135° polarization light intensity. , , In the formula: This represents the light intensity response value of a pixel in the image at the corresponding polarization angle. These are the three components of the Stokes vector. Based on the numerical relationships of the components in the Stokes vector parameters, the polarization degree of the pixel at the current acquisition time node is calculated, i.e.: In the formula: This is the polarization degree value of the pixel at the current acquisition time point. This polarization degree value represents the degree of polarization of the light wave at the pixel point, and the value ranges from 0 to 1.

[0067] S113. Arrange the polarization degree values ​​of each pixel within a preset acquisition time window in chronological order to form a polarization degree temporal evolution curve. The preset acquisition time window ranges from 10s to 300s, with a preferred value of 60s, covering sixty acquisition time nodes. Organize the polarization degree temporal evolution curves of each pixel into a two-dimensional matrix according to their spatial location. The rows and columns of the matrix correspond to the pixel coordinates in the image space, and the elements of the matrix are the polarization degree temporal evolution curves of the corresponding pixels, thereby constructing a polarization degree temporal evolution matrix.

[0068] Temporal difference analysis is performed on each temporal evolution curve in the matrix to calculate the change in polarization degree value of each pixel between adjacent acquisition time nodes, that is, the difference between the polarization degree value at the later time and the polarization degree value at the previous time: In the formula: For the first The change in polarization degree between adjacent acquisition time points For pixels at the acquisition time node The degree of polarization at that location, , The start time of data collection. This represents the actual time interval between adjacent data collection points. Further calculation of the rate of change characteristic is needed, which is the ratio of the aforementioned change to the actual time interval between adjacent data collection points: In the formula: The polarization degree change rate characteristic is obtained through this time-series differential analysis. The dynamic change amount and rate of change of polarization degree of each pixel within the acquisition time window are obtained, which is used to characterize the spatial distribution characteristics of the surface tension of the liquid surface over time.

[0069] S114. Pixels whose rate of change exceeds a preset rate of change threshold are identified as pixels affected by surface tension disturbance. The preset rate of change threshold ranges from 0.05 to 0.5, with a preferred value of 0.1. Eight-neighbor connected component analysis is performed on all pixels identified as affected by surface tension disturbance to extract the continuous spatial region they occupy. This continuous spatial region is then marked as the surface tension disturbance region.

[0070] A contour tracking algorithm is used to extract boundary position information of the surface tension disturbance region. Specifically, starting from the top-left corner pixel of the region, the first pixel belonging to the region within its eight-neighborhood is searched clockwise as the next contour point. This process is iteratively connected to form a closed boundary chain code, and the set of boundary pixel coordinates is recorded. The minimum bounding rectangle algorithm is used to calculate the extreme values ​​of the boundary pixel coordinate set along the horizontal and vertical axes. The minimum horizontal coordinate, minimum vertical coordinate, maximum horizontal coordinate, and maximum vertical coordinate determine the coordinates of the four vertices of the rectangle, forming the minimum bounding rectangle enclosing the region. The spatial boundary position information of the region is recorded, along with the maximum, minimum, mean, and standard deviation statistics of the polarization degree change rate of each pixel within the region, forming a complete quantitative description of the surface tension disturbance region. , In the formula: This represents the average rate of change of polarization degree of each pixel within the surface tension disturbance region. The standard deviation of the rate of change of polarization degree of each pixel within the surface tension disturbance region. This represents the total number of pixels within the surface tension disturbance region. For the first in the region The polarization degree change rate feature of each pixel.

[0071] S12. Extraction of spatial frequencies of oil film interference fringes and marking of prior regions of oil contamination, specifically including:

[0072] S121. Select the image frame with the highest signal-to-noise ratio from the multi-polarization angle image sequence as the interferometric analysis image. In specific implementation, calculate the signal-to-noise ratio evaluation value for each image corresponding to each polarization angle in the sequence. This signal-to-noise ratio evaluation value is quantified by the ratio of the image gray-level mean to the gray-level standard deviation of the background region. The background region is selected as a rectangular area of ​​10% of the area at each of the four corners of the image, i.e.: In the formula: For the first The signal-to-noise ratio evaluation value of images at each polarization angle. For the first The average gray level of an image at each polarization angle The standard deviation of gray levels in the background regions at the four corners of the image is used. The signal-to-noise ratio (SNR) evaluation values ​​for images at different polarization angles are compared. The image frame corresponding to the maximum value is selected as the interferometric analysis image.

[0073] The selected interferometric analysis image is converted to grayscale using a weighted average method, i.e.: In the formula: For interferometric analysis of the image in pixel coordinates grayscale value at that location , , For pixels The light intensity response values ​​in the red, green, and blue channels, These are the grayscale weighting coefficients. The value ranges from 0.2 to 0.4, with a preferred value of 0.3; The value ranges from 0.3 to 0.6, with a preferred value of 0.59; The value ranges from 0.1 to 0.3, with a preferred value of 0.11. The above weighting coefficients satisfy... If the sum of the individually selected coefficients is not 1, normalization is performed according to the proportion of each coefficient to the sum of the three to ensure brightness consistency during grayscale conversion. After grayscale processing, the grayscale value of each pixel in the image is extracted pixel by pixel to generate a grayscale value distribution matrix. The rows and columns of this matrix correspond to the pixel coordinates in the image space, and the matrix elements are the grayscale values ​​of the corresponding pixels, thereby obtaining the grayscale value distribution information of the interferometric analysis image.

[0074] S122. Perform sliding window segmentation on the interferometric image. Set the sliding window to a square window with a side length ranging from 16 to 128 pixels, preferably 64 pixels; the sliding step size ranges from 1 / 4 to 1 / 2 of the window side length, preferably half the window side length. Traverse the interferometric image sequentially according to the sliding window and sliding step size to obtain several local image blocks. For parts of the image boundary that are smaller than the window size, use zero padding or mirror continuation.

[0075] A two-dimensional discrete Fourier transform is performed on each local image patch to transform it from the spatial domain to the frequency domain. The magnitude of the transformed complex result is calculated and logarithmic enhancement is applied to obtain the spectral amplitude distribution map, i.e.: In the formula: The frequency domain coordinates of a local image patch Complex spectral values ​​at that location, For local image patches in pixel coordinates grayscale value at that location The value represents the side length of the sliding window. In the spectral amplitude distribution diagram, the horizontal and vertical axes correspond to the spatial frequency components in the horizontal and vertical directions, respectively. The amplitude value corresponding to each frequency component represents the intensity contribution of the interference fringes at that spatial frequency. The peak amplitude location is searched in each spectral amplitude distribution diagram using a local maximum search method. Centered on the maximum spectral amplitude value, the uniqueness of the peak is confirmed within its 8-neighborhood or 24-neighborhood. The spatial frequency coordinates corresponding to this peak are extracted, including both the horizontal and vertical frequency components.

[0076] S123. Calculate the dominant frequency direction and magnitude of the interference fringes based on the spatial frequency coordinates corresponding to each local image patch: The dominant frequency direction is calculated using the arctangent function, i.e., the arctangent of the ratio of the vertical frequency component to the horizontal frequency component is used as the direction angle, with the direction angle ranging from 0° to 180°. In the formula: The dominant frequency direction of interference fringes in a local image patch, in degrees. These are the horizontal and vertical spatial frequency coordinates corresponding to the peak amplitude of the spectrum; when When, add it Mapped to to Range, to ensure azimuth coverage to The full range. The magnitude of the dominant frequency is calculated by taking the square root of the sum of the squares of the horizontal and vertical frequency components, i.e.: In the formula: The magnitude of the dominant frequency of the interference fringes in the local image patch.

[0077] The calculated principal frequency direction is compared with the principal frequency magnitude and the standard interference frequency characteristics of each oil type in the preset oil film interference frequency characteristic library, that is: , In the formula: The directional deviation between the main frequency direction and the standard interference frequency characteristic The center value of the standard principal frequency direction corresponding to the oil type in the preset oil film interference frequency feature library, The relative magnitude deviation between the dominant frequency and the standard interference frequency characteristic. The standard principal frequency magnitude center value for the corresponding oil type is defined in a preset oil film interference frequency feature library. This library records the standard principal frequency direction range and magnitude range of oil film interference fringes formed by various oil types under standard experimental conditions. The deviation between the principal frequency direction and the center value of the corresponding oil type's standard principal frequency direction is considered. Within the preset directional deviation range, and the deviation between the magnitude of the main frequency and the center value of the corresponding standard main frequency of the oil product. When the size deviation is within a preset range, it is determined that there is oil contamination in the spatial region corresponding to the local image block; wherein the preset directional deviation range is 0° to 15°, with a preferred value of 10°; and the preset size deviation range is 5% to 20%, with a preferred value of 10%.

[0078] S124. Extract and connect the boundary contours of the spatial regions corresponding to the local image blocks identified as having oil contamination. Specifically, merge the spatial regions corresponding to adjacent local image blocks that are both identified as having oil contamination to form candidate oil contamination connected regions. For each candidate oil contamination connected region, use the eight-neighbor boundary tracking algorithm to extract the outer boundary contour. The specific steps of this algorithm are as follows: Starting from the upper left corner of the image, scan pixels in row-priority order. Record the first pixel identified as having oil contamination as the boundary starting point and mark it as visited. Examine its eight neighboring pixels centered on this starting point, checking the neighboring pixels in the eight directions (up, upper right, right, lower right, lower, lower left, left, upper left) in a clockwise direction. Record the first neighboring pixel identified as having oil contamination as the current boundary point and mark it as visited. Question: Taking the current boundary point as the center and the direction of entry into the current boundary point in the previous step as the reference, backtrack one position counterclockwise as the initial direction for the next check, and then check eight neighboring pixels in a clockwise direction. Record the first neighboring pixel that is determined to have oil contamination and has not been marked as visited as the next boundary point. If no matching pixel exists within the eight neighbors, backtrack to the previous boundary point and continue searching. Iterate until the next boundary point coincides with the boundary starting point, forming a closed boundary chain code, and record the boundary pixel coordinate set. For the extracted continuous region, calculate the uniformity of the distribution of the main frequency direction of each local image patch. Specifically, calculate the absolute value of the difference between the main frequency direction of each local image patch and the average main frequency direction of the region, i.e.: In the formula: It represents the maximum absolute value of the difference between the dominant frequency direction of each local image patch within the prior region of oil pollution and the average dominant frequency direction of the region. This is the set of indices for all local image patches within the candidate oil contamination connected region. For the first The dominant frequency direction of a local image patch The direction of the dominant frequency of the regional weighted average; when the absolute value reaches its maximum value. When the value does not exceed the preset uniformity threshold, the uniformity of the spatial frequency distribution of the interference fringes is determined to meet the preset uniformity condition; wherein the preset uniformity threshold ranges from 5° to 20°, with a preferred value of 10°.

[0079] Continuous regions meeting preset uniformity conditions are marked as prior regions of oil contamination, and their boundary location information is recorded. This boundary location information is represented by the set of boundary pixel coordinates or the coordinates of the vertices of the minimum bounding rectangle. Simultaneously, the dominant frequency direction information of this region is recorded, i.e., the weighted average of the dominant frequency directions of each local image patch within the region. In the formula: For the first The peak spectral amplitude of each local image patch; and the magnitude of the dominant frequency, i.e., the weighted average of the dominant frequencies of each local image patch within the region, with the weights being the magnitude of the peak spectral amplitude of each local image patch. In the formula: The weighted average dominant frequency of the prior region of oil pollution. For the first The main frequency size of a local image patch.

[0080] S13. Region of Interest Fusion and Adaptive Hyperspectral Band Acquisition, specifically including:

[0081] S131. Map the boundary location information of the surface tension disturbance region and the boundary location information of the prior oil contamination region to the same spatial coordinate system. Specifically, the pixel coordinate system of the zero-degree polarization angle image in the multi-polarization angle image sequence is used as the unified spatial coordinate system. The set of boundary pixel coordinates or the vertex coordinates of the minimum bounding rectangle recorded during the extraction process of each region are transformed to this unified coordinate system. Calculate the spatial intersection area of ​​the two regions, i.e., the total number of pixels in the spatially overlapping part enclosed by the boundaries of the two regions. Calculate the spatial union area of ​​the two regions, i.e., the total number of pixels in the spatially jointly covered part enclosed by the boundaries of the two regions.

[0082] The ratio of the intersection area to the union area is used as the spatial overlap, i.e.: In the formula: The area represents the intersection of the surface tension disturbance region and the prior oil contamination region, expressed in pixels. The area represents the union of the surface tension disturbance region and the prior oil contamination region, expressed in pixels. The spatial overlap is defined as 0 to 1. The minimum distance between the boundaries of two regions is calculated by iterating through the Euclidean distances between pixels on the boundary of the surface tension disturbance region and pixels on the boundary of the oil contamination prior region, and taking the minimum value as the spatial proximity. In the formula: Spatial proximity, in pixels; This is the set of pixels at the boundary of the surface tension perturbation region. This represents the set of pixels at the boundary of the prior region of oil pollution. For boundary pixels coordinates For boundary pixels The coordinates.

[0083] S132. When the spatial overlap exceeds a preset overlap threshold, or the spatial proximity is less than a preset proximity threshold, the surface tension disturbance region and the prior oil contamination region are merged into a single region of interest. The preset overlap threshold ranges from 0.3 to 0.8, with a preferred value of 0.5; the preset proximity threshold ranges from 5 pixels to 50 pixels, with a preferred value of 20 pixels. During merging, all pixels in the boundary pixel coordinate sets of the two regions are treated as a single point set. The minimum convex hull polygon of this point set is calculated using a convex hull algorithm, and the boundary of this convex hull polygon is used as the boundary of the merged region of interest.

[0084] When the spatial overlap does not exceed a preset overlap threshold and the spatial proximity is not less than a preset proximity threshold, the surface tension disturbance region and the prior oil contamination region are treated as independent regions of interest, each retaining its original boundary location information. A unique region identifier is assigned to each of the merged regions of interest or independent regions of interest for subsequent region differentiation and source tracing during hyperspectral acquisition and data packaging.

[0085] S133. For each region of interest, determine the spatial sampling density or scanning step size for hyperspectral acquisition based on its spatial area. The spatial area is quantified by the total number of pixels within the region, i.e.: In the formula: The region of interest is the area of ​​the region, which is the total number of pixels within the region. This represents the set of pixel coordinates for the region of interest. When the area is larger than a preset large area threshold, a lower sampling density or a larger scan step size is used; when the area is smaller than a preset small area threshold, a higher sampling density or a smaller scan step size is used; when the area is between the preset small area threshold and the preset large area threshold, linear interpolation is used to determine a medium sampling density or a medium scan step size. The preset large area threshold ranges from 5000 to 50000 pixels, with a preferred value of 10000 pixels; the preset small area threshold ranges from 500 to 4999 pixels, with a preferred value of 1000 pixels; the spatial sampling density ranges from one spectral point per 10 pixels to one spectral point per 200 pixels, with a preferred value of one spectral point per 50 pixels; the scan step size ranges from 2 pixels to 20 pixels, with a preferred value of 10 pixels. When using a pushbroom hyperspectral imager, the scan step size controls the spatial sampling position; when using a filter-type hyperspectral camera, the spatial sampling density controls the spectral point acquisition interval.

[0086] The characteristic bands are determined based on the degree of overlap between the region of interest and the prior region of oil contamination, as well as the surface tension perturbation region. The ratio of the area of ​​the intersection between the region of interest and the prior region of oil contamination to the area of ​​the region of interest is calculated as the oil feature overlap degree, i.e.: In the formula: The degree of overlap in oil characteristics, The area of ​​the intersection between the region of interest and the prior region of oil contamination is expressed in pixels. The ratio of the area of ​​the intersection with the surface tension perturbation region to the area of ​​the region of interest is calculated as the interface feature overlap, i.e.: In the formula: For interface feature overlap, The area represents the intersection of the region of interest (ROI) and the surface tension perturbation region, expressed in pixels. Oil feature bands are selected when the overlap of oil features exceeds a preset band selection threshold, and interface feature bands are selected when the overlap of interface features exceeds a preset band selection threshold. The preset band selection threshold ranges from 0.2 to 0.5, with a preferred value of 0.3. The selected oil and interface feature bands are merged and deduplicated. The oil feature bands are defined as 400 nm to 500 nm and 600 nm to 700 nm, while the interface feature bands are defined as 500 nm to 600 nm and 900 nm to 1000 nm. These band ranges are pre-defined based on standard spectral feature libraries for various pollutants. The merged and deduplicated bands form the adaptive hyperspectral band subset for the ROI.

[0087] S134. Hyperspectral image data is acquired for each region of interest according to the spatial sampling density or scanning step size corresponding to each region of interest, as well as the adaptive hyperspectral band subset. In specific implementation, a pushbroom hyperspectral imager or a filter-type hyperspectral camera is used. The spatial sampling position of the imager is controlled according to the scanning step size, and the wavelength acquisition range or filter switching sequence of the imager is controlled according to the adaptive hyperspectral band subset.

[0088] After acquisition, the sequence image data of each region of interest (ROI), i.e., the polarization angle images of the corresponding region positions in the multi-polarization angle image sequence, are associated and packaged with the hyperspectral image data according to the region identifier number. Each data packet contains the location information of the region, i.e., the set of pixel coordinates of the region boundary or the vertex coordinates of the minimum bounding rectangle, and the region identifier number. The final output is a sequence image of the ROI marked with location and identifier information, and the corresponding adaptive band spectral data volume. This adaptive band spectral data volume is a three-dimensional data cube, with two spatial dimensions and one spectral dimension of the adaptive hyperspectral band subset.

[0089] S2. Divide the region of interest sequence image and adaptive band spectral data volume into micro-regions, and extract polarization temporal evolution, interference texture, and spectral absorption feature vectors; construct a ternary coupling map to identify overlapping micro-regions, implement cross-modal separation based on surface tension gradient constraints, and output a fingerprint feature set of contaminant components; specifically including:

[0090] S21. Micro-region segmentation and cross-modal feature vector extraction, specifically including:

[0091] S211. Perform pixel-level registration between the region of interest sequence image and the adaptive band spectral data volume. In practice, compare the spatial resolutions of the two. When the spatial resolutions are inconsistent, use the pixel coordinate system of the zero-degree polarization angle image in the multi-polarization angle image sequence as the reference spatial grid, and perform spatial resampling on the adaptive band spectral data volume using bilinear interpolation or bicubic interpolation. This ensures that the spatial dimension of the resampled spectral data volume corresponds one-to-one with the pixel coordinate system of the reference image. The spatial registration error after resampling is controlled within 0.5 pixels.

[0092] Each registered pixel within a region of interest (ROI) is considered a micro-unit, with each micro-unit possessing a unique spatial coordinate index. A one-to-one mapping relationship is established between the polarization time-series image data corresponding to each micro-unit (i.e., the sequence of light intensity response values ​​or polarization degree values ​​for that pixel at various polarization angles and acquisition time points) and the spectral data at the corresponding spatial coordinate location (i.e., the sequence of reflectance or radiance values ​​for that pixel in each band of the adaptive hyperspectral band subset), forming cross-modal data pairs. For micro-units at the ROI boundary where data is missing or abnormal due to resampling, the values ​​of neighboring valid micro-units are used to fill in the gaps or mark them as invalid micro-units. Units marked as invalid micro-units do not participate in the subsequent feature coupling, decoupling, and recognition operations in S221 to S234. When generating the final recognition result map, they are filled with null values ​​or interpolated results from neighboring valid results, completing the cross-modal data association for each micro-unit.

[0093] S212. Extract time-series features from the polarization time-series image data of each micro-unit to obtain a polarization time-series evolution feature vector. Specifically, calculate statistical features for the polarization degree time-series evolution curve of the micro-unit, including the mean polarization degree, standard deviation, range, mean first derivative, mean second derivative, skewness, kurtosis, and entropy, to construct the polarization time-series evolution feature vector. The dimension value ranges from 6 to 10, with an optimal value of 8, i.e.: Wherein: mean degree of polarization Standard deviation of polarization Extremely poor polarization First derivative mean Second derivative mean Polarization degree skewness Polarization degree kurtosis Polarization degree entropy In the formula: This represents the total number of data collection time nodes within the preset data collection time window. For micro-units at the acquisition time node The degree of polarization at that location, Let be the maximum and minimum polarization values ​​of the micro-unit within the acquisition time window, respectively. This represents the actual time interval between adjacent data collection points. The number of bins for the degree of polarization is specified, ranging from 8 to 32, with a preferred value of 16. The degree of polarization value falls into the first The frequency within each sub-box.

[0094] Interference texture analysis is performed on the sequence of images of the region of interest. A local texture analysis window is selected with the micro-unit as the center. The side length of the window ranges from 3 to 15 pixels, with a preferred value of 5 pixels. A two-dimensional Fourier transform is performed on the image within the window to extract the direction, magnitude, and peak amplitude of the dominant frequency of the interference fringes, thus constructing the interference texture feature vector. The dimension value ranges from 3 to 5, with a preferred value of 4, i.e.: Among them: micro-region window spectral amplitude Spectral amplitude peak Micro-region dominant frequency direction ,when Add it at that time Mapped to to Interval; Micro-region main frequency magnitude Spectral concentration In the formula: This represents the side length of the local texture analysis window, ranging from 3 to 15 pixels, with a preferred value of 5 pixels. pixels within the window grayscale value, for Take the maximum value Frequency domain coordinates at time.

[0095] Spectral features are extracted from the spectral reflectance curves of each micro-region in the adaptive band spectral data. The absorption peak depth, half-width at half-maximum, number of extreme points of the first derivative of the spectrum, and normalized spectral index are calculated to construct the spectral absorption feature vector. The dimension value ranges from 4 to 6, with a preferred value of 4, i.e.: Where: absorption peak depth Half-peak width Number of extreme points of the first derivative of the spectrum Normalized spectral index In the formula: To absorb the baseline reflectance at the shoulder of the peak, To absorb the reflectivity at the peak and valley bottom, These are the left and right wavelength values ​​corresponding to half the depth of the absorption peak, respectively, in nanometers; To adapt to the total number of bands in the hyperspectral band sub-set, This is an indicator function that takes the value 1 when the condition is true and 0 otherwise. wavelength The first derivative of the spectral reflectance at that point, The characteristic wavelengths are respectively and The reflectivity at that location.

[0096] The elements in each feature vector are all subjected to min-max normalization to unify the numerical range of each vector element to the interval between 0 and 1, that is: In the formula: These are the normalized eigenvector element values. This is the minimum value of the element across all microcells. This is the maximum value of the element across all microcells.

[0097] S22. Ternary coupling mapping and feature overlap micro-region identification, specifically including:

[0098] S221. For each micro-region unit, the polarization time-series evolution feature vector, interference texture feature vector, and spectral absorption feature vector are concatenated according to a unified feature dimension index. In specific implementation, the dimension of the three feature vectors is first determined. When the dimensions are inconsistent, principal component analysis is used to unify the feature dimensions: a sample matrix is ​​constructed using the feature vectors to be unified, and principal component analysis is performed. If the original feature dimension is higher than the preset unified dimension, the first preset unified dimension of principal components is retained to achieve dimensionality reduction. If the original feature dimension is lower than the preset unified dimension, the remaining dimension is filled with zero vectors to unify all feature vectors to the same dimension. The value range of the preset unified dimension is 5 to 20, with a preferred value of 10.

[0099] The polarization temporal evolution feature vector, interference texture feature vector, and spectral absorption feature vector, after unifying the dimensions, are connected end to end in sequence to form the initial ternary coupled feature vector of each micro-unit. The total dimension of the initial ternary coupled feature vector is three times that of the unified feature dimension. Its first half of the elements characterizes the dynamic characteristics of polarization temporal evolution, the middle half of the elements characterizes the spatial characteristics of interference texture, and the second half of the elements characterizes the material characteristics of spectral absorption. This achieves a high-dimensional joint characterization of the three modal characteristics within the same micro-unit.

[0100] S222. For each micro-region unit, calculate the first correlation coefficient between the polarization time-series evolution eigenvector and the spectral absorption eigenvector, calculate the second correlation coefficient between the interference texture eigenvector and the spectral absorption eigenvector, and calculate the third correlation coefficient between the polarization time-series evolution eigenvector and the interference texture eigenvector. Specifically, the Pearson correlation coefficient calculation method is used, which is the ratio of the sum of the products of the differences between corresponding dimensional elements of the two eigenvectors to the product of their respective standard deviations.

[0101]

[0102]

[0103]

[0104] in: The first correlation coefficient, namely the Pearson correlation coefficient between the polarization time-series evolution feature vector and the spectral absorption feature vector, characterizes the degree of linear correlation between the dynamic evolution of surface tension and the spectral absorption of matter. The second correlation coefficient, namely the Pearson correlation coefficient between the interference texture feature vector and the spectral absorption feature vector, characterizes the degree of linear correlation between the interference texture morphology and the spectral absorption of the material. The third correlation coefficient, namely the Pearson correlation coefficient between the polarization time-series evolution feature vector and the interference texture feature vector, characterizes the degree of linear correlation between the dynamic evolution of surface tension and the morphology of interference texture. Let be the first feature vectors of polarization temporal evolution, interference texture, and spectral absorption after unification of dimensions. Dimensional elements; Let be the mean values ​​of the corresponding feature vectors of the same dimension.

[0105] Record the three correlation coefficient values ​​of the micro-region unit, which correspond to the correlation strength between the polarization time-series evolution feature vector and the spectral absorption feature vector, the interference texture feature vector and the spectral absorption feature vector, and the polarization time-series evolution feature vector and the interference texture feature vector, respectively.

[0106] S223. For each micro-region, read the values ​​of the three correlation coefficients of the micro-region, calculate the absolute value of each correlation coefficient, and when the absolute value of at least two of the three correlation coefficients exceeds the preset correlation threshold at the same time, it is determined that the micro-region has multimodal feature coupling interference caused by mixed contamination, and it is marked as a feature overlapping micro-region and its spatial coordinates are recorded; when only one or zero correlation coefficients exceed the preset correlation threshold, the micro-region is determined to be a feature clear micro-region and no overlap marking is performed; the preset correlation threshold ranges from 0.3 to 0.9, with a preferred value of 0.7.

[0107] After determining all micro-region units, the spatial coordinates of all micro-regions marked as feature-overlapping regions are summarized to form a feature-overlapping micro-region distribution map, which is used to determine the spatial range of physical constraint decoupling in the subsequent cross-modal separation step.

[0108] S23. Surface tension gradient-constrained cross-modal separation and fingerprint feature output, specifically including:

[0109] S231. For each overlapping micro-region, the polarization degree change rate distribution information of the corresponding pixel points recorded in S114 is used as a proxy index for the surface tension disturbance intensity. In specific implementation, the polarization degree change rate feature value of the micro-region at each acquisition time node is read. When the micro-region is composed of multiple pixels, the average value of the polarization degree change rate feature of all pixels within the micro-region is taken as the surface tension disturbance intensity value of the micro-region at each acquisition time node, that is: In the formula: For the first Each micro-area at the data collection time point The surface tension disturbance intensity value at that location, For the first The total number of pixels contained in each micro-region For the first A set of pixel indices for each micro-region For the first Each pixel at the acquisition time point The polarization degree change rate characteristics at the location are analyzed. The surface tension disturbance intensity values ​​of the micro-region and its neighboring micro-regions are extracted at each acquisition time node. The neighboring micro-regions are defined as all micro-region units whose spatial coordinate distance from the micro-region is less than a preset neighborhood distance threshold. The preset neighborhood distance threshold ranges from 2 to 10 pixels, with a preferred value of 5 pixels.

[0110] Based on the spatial positional differences of each adjacent micro-region relative to the current micro-region, the spatial variation rate of the surface tension perturbation intensity value in two orthogonal directions is calculated. This is calculated by dividing the difference in surface tension perturbation intensity value between the current micro-region and its adjacent micro-region along the horizontal direction by the horizontal pixel distance, and by dividing the difference in surface tension perturbation intensity value between the current micro-region and its adjacent micro-region along the vertical direction by the vertical pixel distance. The spatial variation rate components in the two orthogonal directions are combined to form the surface tension gradient vector of the micro-region, i.e.: In the formula: For micro-regions Surface tension gradient vector, These represent the surface tension disturbance intensity values ​​of adjacent micro-regions in the horizontal and vertical directions at the corresponding acquisition time points, respectively. Let be the spatial coordinates of adjacent micro-regions in the horizontal and vertical directions, respectively. Let m be the spatial coordinates of the micro-region. The value represents the surface tension disturbance intensity of micro-region m at the corresponding acquisition time node.

[0111] S232. Using the surface tension gradient vector as a physical constraint, a joint separation objective function is established that couples spectral absorption features, polarization temporal evolution features, and interference texture features. This joint separation objective function consists of a linearly weighted combination of four sub-terms: a spectral absorption feature separation sub-term, a polarization temporal evolution feature separation sub-term, an interference texture feature separation sub-term, and a surface tension gradient constraint regularization term, i.e.:

[0112] In the formula: To jointly separate the objective functions, Separate sub-terms for spectral absorption characteristics. To separate sub-items based on polarization temporal evolution characteristics, To separate sub-items for interference texture features, For surface tension gradient constraint regularization, The weighting coefficients for each separated sub-item range from 0.1 to 2.0, with a preferred value of 1.0. is the weighting coefficient for the surface tension gradient constraint regularization term, with a value ranging from 0.01 to 1.0, and a preferred value of 0.5.

[0113] Each separated sub-item represents the reconstruction error after decomposing the corresponding modal feature vector into contribution components of different pollutants. Specifically, it is measured by the squared Euclidean distance between the weighted reconstructed value of each pollutant contribution component and the original feature vector, i.e.:

[0114]

[0115]

[0116]

[0117] in: The total number of overlapping micro-regions. The preset number of pollutant types ranges from 2 to 5, with a preferred value of 3. Micro-regions Unified dimensional spectral absorption, polarization temporal evolution, and interference texture feature vectors; Pollutant components In micro-region The spectral absorption, polarization time-series evolution, and interference texture contribution vector.

[0118] The surface tension gradient constraint regularization term characterizes the continuity constraint of the contribution components of each pollutant after separation in the spatial domain. The gradient magnitude is used as the weight for the spatial continuity penalty. The physical basis for this is that regions with significant surface tension gradients correspond to the physical boundaries of pollutant concentration changes. If there are drastic jumps in the contribution components of each decoupled pollutant at these boundaries, it is more likely due to cross-modal decoupling artifacts rather than genuine concentration jumps. Therefore, by weighting the contribution components of adjacent micro-regions at the boundary with the surface tension gradient magnitude, the consistency constraint is strengthened, suppressing decoupling artifacts caused by modal aliasing. Specifically, this is measured by the inner product of the surface tension gradient vector and the spatial differences in the contribution components of corresponding pollutants in adjacent micro-regions, i.e.: In the formula: Pollutant components In micro-region The concatenated contribution vector, , For micro-regions The set of adjacent micro-region indices, surface tension gradient magnitude As a spatial continuity penalty weight, it is multiplied by the square of the L2 norm of the difference in contribution components of corresponding pollutants in adjacent micro-regions, and then summed over all adjacent micro-region pairs, all feature-overlapping micro-regions, and all pollutants.

[0119] S233. The contribution components of each separation sub-item in the joint separation objective function, belonging to different pollutant components, are set as variables to be solved. It is assumed that there is a preset number of pollutant component types, which ranges from 2 to 5, with a preferred value of 3. For each pollutant component, its contribution component vector in the spectral absorption feature vector, polarization time-series evolution feature vector, and interference texture feature vector is defined, and the dimension of each contribution component vector is consistent with the unified feature dimension.

[0120] The alternating direction multiplier method is used for minimization optimization. The specific steps are as follows: Initialize each contribution component vector and auxiliary variable vector as a zero vector, initialize the Lagrange multiplier vector, and set an initial value for the penalty parameter, which ranges from 0.01 to 10, with a preferred value of 1.0. The spectral absorption contribution component vector group, polarization temporal evolution contribution component vector group, and interference texture contribution component vector group in the joint objective function are treated as three variable blocks. An auxiliary variable is introduced to transform the coupling relationship of the three contribution components in the surface tension gradient constraint regularization term into an equality constraint, constructing an augmented Lagrange function. Two variable blocks are then fixed sequentially. The proximal gradient descent method is used to solve the sub-optimization problem for the current variable block, applying a non-negativity constraint so that all elements are greater than or equal to 0, and updating the corresponding Lagrange multiplier terms and penalty parameter. Subsequently, the auxiliary variable vector and Lagrange multiplier vector are updated, and the penalty parameter is increased by a preset factor, ranging from 1.1 to 2, with a preferred value of 1.5. The original residual and the dual residual are calculated. The iteration terminates when both are less than a preset convergence threshold, which has a range of 10. −6 Up to 10 −3 The preferred value is 10. −4 Alternatively, the iteration may terminate when the preset maximum number of iterations is reached, with the maximum number ranging from 500 to 5000, and a preferred value of 2000. After the iteration terminates, the contribution values ​​of each pollutant component in the three feature vectors are obtained.

[0121] S234. The spectral absorption contribution, polarization temporal evolution contribution, and interference texture contribution components corresponding to each pollutant component are categorized and combined according to the pollutant component type. For each pollutant component, its spectral absorption contribution, polarization temporal evolution contribution, and interference texture contribution components are concatenated sequentially to form an independent fingerprint feature vector for that pollutant component in that micro-region. All overlapping micro-regions are traversed, and the independent fingerprint feature vectors of each pollutant component within each micro-region are organized according to their spatial coordinates to form a spatial distribution fingerprint spectrum of each pollutant component within the overlapping micro-regions. The spatial distribution fingerprint spectra of each pollutant component are summarized and output as a decoupled fingerprint feature set for the pollutant components.

[0122] S3. Input the pollutant component fingerprint feature set into the hierarchical recognition framework for broad category discrimination and detailed identification, generating dynamic confidence weights; when the confidence is low, the first feedback path is triggered to adjust the surface tension gradient constraint and re-decouple, while the second feedback path is triggered to optimize the polarization angle sequence and band selection strategy and re-execute, outputting the wastewater component identification result with a confidence index; specifically including:

[0123] S31. Hierarchical classification and dynamic confidence weight generation, specifically including:

[0124] S311. Extract the spectral absorption contribution components corresponding to each pollutant component in each micro-region from the decoupled fingerprint feature set of pollutant components. Arrange the spectral absorption contribution components of each pollutant component in the same micro-region according to the pollutant component type to form the spectral absorption feature vector of the micro-region. Input the spectral absorption feature vector into the upper classifier of the hierarchical recognition framework.

[0125] The upper-level classifier is constructed using a random forest structure, consisting of a preset number of decision trees operating in parallel. This preset number ranges from 50 to 500, with an optimal value of 200. Each decision tree is trained independently based on a random subset of spectral absorption feature vectors. The final discrimination result is output through a majority voting mechanism. Each decision tree uses a bootstrap sampling method to draw the same number of samples with replacement from the training samples as from the original training set to construct a training subset. The feature random subset sampling ratio ranges from [value missing]. to ,in Let be the dimension of the spectral absorption eigenvector, with the preferred value being... The upper-level classifier takes as input the spectral absorption feature vector of each micro-region, with its feature dimension consistent with the unified feature dimension, and outputs the probability distribution of each pollution category. The pollution categories include preset categories such as oil pollution, organic pollution, inorganic pollution and suspended particulate pollution, with the number of preset categories ranging from 2 to 6, and the preferred value being 4 categories.

[0126] Each micro-region's spectral absorption feature vector is input into the upper-level classifier, which outputs the probability value of belonging to each pollution category. The pollution category corresponding to the maximum probability is taken as the pollution category label of the micro-region. When the maximum probability is lower than the preset category discrimination threshold, it is marked as an undetermined category. The threshold value ranges from 0.5 to 0.9, with a preferred value of 0.7.

[0127] S312. For all micro-regions under the same pollution category label output by the upper-level classifier, extract the polarization temporal evolution contribution component and interference texture contribution component of each pollution component in the corresponding micro-region from the pollution component decoupled fingerprint feature set, and construct the polarization temporal evolution feature vector and interference texture feature vector respectively. Concatenate and fuse the two feature vectors into a joint subdivision feature vector, and input it into the lower-level classifier of the hierarchical recognition framework.

[0128] The lower-level classifier sets up an independent sub-classifier for each pollution category. Each sub-classifier has the same structure but independent training data. It adopts a neural network structure with two hidden layers. The number of neurons in each layer ranges from 32 to 256, with an optimal value of 128. The activation function is a modified linear unit or a hyperbolic tangent function. The number of neurons in the output layer is consistent with the number of specific pollution types in the pollution category, and the number of neurons in the input layer is consistent with the total dimension of the joint subdivision feature vector. The output layer uses the Softmax activation function, the loss function is cross-entropy loss, and the optimizer is the Adam optimizer. The initial learning rate ranges from 0.001 to 0.01, with an optimal value of 0.001.

[0129] The lower-level classifier takes as input a joint subdivision feature vector, which is the concatenation of the polarization temporal evolution feature vector and the interference texture feature vector. The total dimension is twice the dimension of the unified feature vector. The output is the probability distribution of each specific pollution type within the pollution category. The joint subdivision feature vector of each micro-region is input into the lower-level classifier of the corresponding pollution category. The specific pollution type corresponding to the highest probability is taken as the label of the specific pollution type in that micro-region. When the highest probability is lower than a preset subdivision discrimination threshold, it is marked as an undetermined subdivision type. This threshold ranges from 0.4 to 0.8, with a preferred value of 0.6.

[0130] S313. Calculate three indices for each micro-region and generate dynamic confidence weights by weighted combination: spectral signal-to-noise ratio index. This is the ratio of the mean to the standard deviation of each element in the spectral absorption characteristic vector of this micro-region, i.e.: In the formula: micro-region The mean value of each element in the spectral absorption eigenvector; micro-region Standard deviation of each element in the spectral absorption eigenvector; For micro-regions Spectral absorption eigenvector Each element is composed of the spectral absorption contribution components of each pollutant arranged according to component type, i.e., the spectral absorption characteristic vector described in S212. After dimensional unification, the first Dimensional components; This represents the dimension of the spectral absorption eigenvector. When this ratio... When the signal-to-noise ratio is below a preset low signal-to-noise ratio threshold, the value of which ranges from 2 to 10, with a preferred value of 5.

[0131] Cross-modal consistency index The value of the mean absolute value of the correlation coefficients among the three features of the micro-region: polarization time-series evolution feature vector and spectral absorption feature vector, interference texture feature vector and spectral absorption feature vector, and polarization time-series evolution feature vector and interference texture feature vector. In the formula: For micro-regions The Pearson correlation coefficient between the polarization time-series evolution eigenvector and the spectral absorption eigenvector. For micro-regions The Pearson correlation coefficient between the interference texture eigenvector and the spectral absorption eigenvector. For micro-regions Pearson correlation coefficient between polarization time-series evolution eigenvectors and interference texture eigenvectors.

[0132] Surface tension uniformity index The value of the cosine similarity between the surface tension gradient vector of this micro-region and each standard gradient pattern in the preset standard gradient pattern library is the maximum value. This preset standard gradient pattern library contains typical surface tension gradient vectors of various contaminants measured under standard experimental conditions, namely: In the formula: For micro-regions The surface tension gradient vector has been defined in S231. For the first standard gradient pattern in the preset standard gradient pattern library A standard gradient pattern vector, This represents the total number of standard gradient patterns in the preset standard gradient pattern library.

[0133] The three indices are weighted and summed, and the sum is normalized to the interval between 0 and 1. This sum is the dynamic confidence weight for that micro-region, i.e.: In the formula: For dynamic confidence weights, micro-region Raw weighted confidence scores; This represents the minimum weighted confidence level among all microregions. The maximum value of the original weighted confidence score across all microregions; These are the weighting coefficients for the three indices, ranging from 0.1 to 1.0, with preferred values ​​of 0.33, 0.33, and 0.34; if ,in The preset minimum value has a range of values. to The preferred value is Then, the dynamic confidence weight of all microregions is uniformly set to 1.0.

[0134] S32. Adaptive correction of low-confidence dual-feedback path and output of confidence results, specifically including:

[0135] S321. When the dynamic confidence weight of a certain micro-region is lower than the preset confidence threshold, the micro-region is marked as a low-confidence micro-region. The preset confidence threshold ranges from 0.3 to 0.8, with a preferred value of 0.6. The specific values ​​of the spectral signal-to-noise ratio index, cross-modal consistency index, and surface tension consistency index of the low-confidence micro-region are extracted. In practice, the corresponding values ​​are retrieved from the three index results recorded during the dynamic confidence weight calculation process, according to the spatial coordinate index of the micro-region. If the micro-region is marked as an undetermined major category by the upper-level classifier or as an undetermined sub-category by the lower-level classifier, it is preferentially included in the low-confidence micro-region set. For each micro-region in the low-confidence micro-region set, a correction record entry is established, containing its spatial coordinate location, the specific values ​​of the three indices, the current pollution major category label, and the specific pollution type label, for subsequent targeted correction in the dual feedback path.

[0136] S322. For low-confidence micro-regions, trigger the second feedback path, optimize the polarization angle sequence and band selection strategy, and re-execute acquisition and identification. Specifically, analyze the spectral signal-to-noise ratio (SNR) index of the low-confidence micro-region and identify band positions below the preset low SNR threshold in S313. Within the neighborhood of this low SNR band, use the local maximum search method to find local maxima points on the spectral reflectance curve, and use their wavelength positions as supplementary band positions. When the number of supplementary bands exceeds a preset upper limit, select the first preset upper limit of supplementary bands according to the magnitude of the local maxima from high to low; the preset upper limit ranges from 2 to 10, with a preferred value of 5. The degree of deviation of the cross-modal consistency index and the surface tension consistency index from their respective preset thresholds determines the insufficient polarization information acquisition, and the angle positions that need to be encrypted are determined accordingly. The specific determination method is as follows: when the cross-modal consistency index is lower than the preset cross-modal threshold and the surface tension consistency index is lower than the preset surface tension threshold, the local correlation between the polarization time-series evolution feature vector and the spectral absorption feature vector of the low confidence micro-region at each polarization angle is analyzed, and the local Pearson correlation coefficient between the polarization time-series evolution feature vector and the spectral absorption feature vector at each polarization angle is calculated. The polarization angles with cross-correlation coefficients lower than the preset local correlation threshold are marked as angle positions with insufficient information. The values ​​of the two thresholds are both in the range of 0.3 to 0.8, and the preferred value is 0.5.

[0137] The location with insufficient information is designated as the specific angle location requiring encrypted acquisition. An encrypted acquisition angle is added between each of the two adjacent original polarization angles. This newly added angle is located at the inner dividing point of the original angle step interval, with an inner dividing ratio ranging from 1:1 to 1:2, preferably 1:2 (i.e., the newly added angle is located at 1 / 3 of the original interval from the left end). The corresponding encrypted angle interval is 1 / 3 to 1 / 2 of the original angle step interval, preferably 1 / 3 of the original interval. The preset local correlation threshold ranges from 0.3 to 0.7, preferably 0.5. Based on this, the polarization angle sequence and hyperspectral band selection strategy are updated, and the acquisition, feature decoupling, and recognition steps are re-executed.

[0138] S323. For low-confidence micro-regions, trigger the first feedback path, adjust the surface tension gradient constraint, and re-decouple. Specifically, calculate the deviation of the cross-modal consistency index of the low-confidence micro-region from a preset standard value. The preset standard value ranges from 0.5 to 0.9, with a preferred value of 0.7. When the cross-modal consistency index is lower than the preset standard value, it indicates that the current surface tension gradient constraint is too strong, leading to misjudgment of feature coupling. In this case, reduce the weight coefficient of the surface tension gradient constraint regularization term in the joint separation objective function. When the cross-modal consistency index is higher than the preset standard value, it indicates that the current surface tension gradient constraint is insufficient, leading to incomplete decoupling. In this case, increase the weight coefficient. The adjustment range of the weight coefficient is calculated proportionally based on the deviation, ranging from 10% to 50%, with a preferred value of 20%. If the adjusted weight coefficient is greater than 1.0, truncate it to 1.0; if it is less than 0.01, truncate it to 0.01. Based on the adjusted weight coefficients, the joint separation objective function is reconstructed, and cross-modal feature separation and recognition steps are performed to obtain the corrected contaminant fingerprint feature set. The corrected contaminant fingerprint feature set is then re-input into the hierarchical recognition framework for further class discrimination and sub-classification to generate new dynamic confidence weights.

[0139] S324. Summarize the pollution category labels, specific pollution type labels, and corresponding dynamic confidence weights for all micro-regions. In practice, the new identification results obtained after re-execution via the second feedback path, and the new identification results obtained after correction via the first feedback path, are fused with the original identification results. The fusion rule is: when the dynamic confidence weight after re-identification is higher than the original dynamic confidence weight, the re-identification result replaces the original result; otherwise, the original result is retained. For each micro-region finally determined, record its pollution category label, specific pollution type label, dynamic confidence weight value, and spatial coordinate position. Organize all micro-region identification results into an identification result map with the same spatial resolution as the original image according to the spatial coordinate position. Each pixel position in the map is labeled with the corresponding pollution category label, specific pollution type label, and dynamic confidence weight. Generate and output the wastewater component identification results with confidence index. The output format includes four-dimensional information: spatial coordinates of each micro-region, pollution category, specific pollution type, and confidence index.

[0140] Example 2

[0141] Please see the appendix Figure 2 A wastewater component identification and analysis system based on spectral and image recognition is used to implement a wastewater component identification and analysis method based on spectral and image recognition, including:

[0142] The polarization-spectral adaptive acquisition module is used to acquire multi-polarization angle image sequences, calculate the polarization degree temporal evolution matrix to identify surface tension disturbance regions, extract the spatial frequency markers of oil film interference fringes to indicate the prior region of oil, fuse and divide the region of interest, adaptively select a subset of hyperspectral bands for acquisition, and output the sequence image of the region of interest and the adaptive band spectral data volume; including:

[0143] The system includes a polarization image acquisition unit and a hyperspectral imaging unit. The polarization image acquisition unit includes a multi-angle polarizer array and an industrial camera. The hyperspectral imaging unit includes a hyperspectral imager and a band selection controller. The band selection controller outputs a band index based on the degree of overlap between the region of interest and the prior region of oil contamination and the surface tension disturbance region to control the acquisition band range of the hyperspectral imager.

[0144] The micro-region coupling and decoupling module is used to divide the sequence image of the region of interest and the adaptive band spectral data volume into micro-regions, extract polarization time evolution, interference texture and spectral absorption feature vectors, construct a ternary coupling map to identify overlapping micro-regions, implement cross-modal separation based on surface tension gradient constraints, and output a fingerprint feature set of contaminant components.

[0145] The hierarchical identification and dual-path feedback module is used to input the fingerprint feature set of pollutants into the hierarchical identification framework for major category discrimination and sub-category identification, generate dynamic confidence weights, and trigger the first feedback path to adjust the surface tension gradient constraint in the micro-region coupling and decoupling module and re-decouple when the confidence is low. At the same time, it triggers the second feedback path to optimize the polarization angle sequence and band selection strategy in the polarization-spectral adaptive acquisition module and re-execute it, outputting the wastewater component identification result with confidence index.

[0146] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention is limited to these examples; within the framework of the invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of the different aspects of the invention as described above, which are not provided in detail for the sake of brevity.

[0147] This invention is intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this specification. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A wastewater component identification and analysis method based on spectral and image recognition, characterized in that, include: Collect multi-polarization angle image sequences, calculate the polarization degree temporal evolution matrix to identify surface tension disturbance regions, and extract the spatial frequency of oil film interference fringes to mark the prior region of oil. The region of interest is fused and divided, and a subset of hyperspectral bands is adaptively selected for acquisition. The output sequence of images of the region of interest and the adaptive band spectral data volume are then used. The sequence image of the region of interest and the adaptive band spectral data volume are divided into micro-regions, and polarization temporal evolution, interference texture and spectral absorption feature vectors are extracted; A ternary coupling map is constructed to identify overlapping micro-regions of features. Cross-modal separation is performed based on surface tension gradient constraints, and a fingerprint feature set of contaminant components is output. The fingerprint feature set of the pollutant components is input into the hierarchical recognition framework for major category discrimination and sub-category recognition, and dynamic confidence weights are generated. When the confidence level is low, the first feedback path is triggered to adjust the surface tension gradient constraint and re-decouple it. At the same time, the second feedback path is triggered to optimize the polarization angle sequence and band selection strategy and re-execute it, outputting the wastewater component identification result with confidence index.

2. The wastewater component identification and analysis method based on spectral and image recognition according to claim 1, characterized in that, The acquisition of multi-polarization angle image sequences and the calculation of the polarization degree temporal evolution matrix to identify surface tension disturbance regions specifically include: The wastewater sample to be tested is placed in a transparent sample container and positioned at the polarization image acquisition station. The polarizer is controlled to rotate step by step, and multiple frames of images are continuously acquired at each polarization angle. The time-series images are preprocessed to obtain a multi-polarization angle image sequence. Pixel-level spatial registration is performed on images with different polarization angles at the same acquisition time node in the multi-polarization angle image sequence. The light intensity response value of each pixel at different polarization angles is extracted, the Stokes vector parameters of the pixel are constructed, and the polarization degree value of the pixel at the current acquisition time node is calculated. The polarization degree values ​​of each pixel within the preset acquisition time window are arranged in time sequence to form a polarization degree time sequence evolution curve. A polarization degree time sequence evolution matrix is ​​constructed according to spatial location. Time sequence difference analysis is performed on each time sequence evolution curve in the matrix to calculate the change amount and rate of change characteristics of the polarization degree value of each pixel between adjacent acquisition time nodes. Pixels whose rate of change exceeds a preset rate of change threshold are identified as pixels affected by surface tension disturbance. The continuous spatial region occupied by the affected pixels is marked as the surface tension disturbance region, and its spatial boundary position information and the distribution information of the polarization degree change rate of each pixel in the region are recorded.

3. The wastewater component identification and analysis method based on spectral and image recognition according to claim 2, characterized in that, The extraction of the spatial frequency markers for oil film interference fringes to indicate the prior region of oil specifically includes: The image frame with the highest signal-to-noise ratio from the multi-polarization angle image sequence is selected as the interferometric analysis image. The interferometric analysis image is then converted to grayscale, and the grayscale value distribution information of each pixel in the image is extracted. The interferometric analysis image is divided into several local image blocks by a sliding window. A two-dimensional Fourier transform is performed on each local image block to obtain a spectral amplitude distribution map. The amplitude peak position is searched in each spectral amplitude distribution map and the corresponding spatial frequency coordinate value is extracted. The direction and magnitude of the main frequency of the interference fringes are determined based on the spatial frequency coordinates. They are then compared with the standard interference frequency features of each type of oil in the preset oil film interference frequency feature library. When the deviations are all within the preset deviation range, it is determined that there is oil contamination in the spatial region corresponding to the local image block. The spatial regions corresponding to the local image blocks where oil pollution is determined to exist are extracted and connected. The continuous regions where the spatial frequency distribution uniformity of the interference fringes meets the preset uniformity condition are marked as the prior regions of oil pollution, and their boundary position information, main frequency direction information and main frequency magnitude information are recorded.

4. The wastewater component identification and analysis method based on spectral and image recognition according to claim 3, characterized in that, The process of fusing and dividing the region of interest, adaptively selecting a subset of hyperspectral bands for acquisition, and outputting a sequence of images of the region of interest and an adaptive band spectral data volume specifically includes: The boundary location information of the surface tension disturbance region and the boundary location information of the oil contamination prior region are mapped to the same spatial coordinate system. The ratio of the intersection area to the union area of ​​the two regions is calculated as the spatial overlap, and the minimum distance between the boundaries of the two regions is calculated as the spatial proximity. When the spatial overlap exceeds a preset overlap threshold or the spatial proximity is less than a preset proximity threshold, the two regions are merged into a fused region of interest, with the minimum convex hull polygon of the two regions as the boundary; otherwise, the surface tension disturbance region and the oil contamination prior region are each taken as an independent region of interest. For each region of interest, the spatial sampling density or scanning step size of hyperspectral acquisition is determined according to its spatial area. The oil characteristic bands and interface characteristic bands are determined according to their degree of overlap with the prior oil contamination region and the surface tension disturbance region, respectively. After merging and deduplication, they are used as the adaptive hyperspectral band subset of the region of interest. Hyperspectral image data are collected according to the spatial sampling density or scanning step size and adaptive hyperspectral band subset corresponding to each region of interest. The sequence image data and hyperspectral image data of each region of interest are packaged according to the identifier, and the sequence image of the region of interest and the corresponding adaptive band spectral data volume are output.

5. The wastewater component identification and analysis method based on spectral and image recognition according to claim 1, characterized in that, The region of interest sequence image and adaptive band spectral data volume are divided into micro-regions, and polarization temporal evolution, interference texture, and spectral absorption feature vectors are extracted, specifically including: The image sequence of the region of interest and the adaptive band spectral data are registered at the pixel level according to the spatial correspondence. Each pixel in each region of interest is taken as a micro-unit, and the polarization time sequence image data and spectral data of each micro-unit are associated and aligned. The polarization time-series image data of each micro-region unit is subjected to time-series feature extraction to obtain polarization time-series evolution feature vectors. Interference texture analysis is performed on the sequence images of the region of interest to extract the spatial frequency and direction features of interference fringes and form interference texture feature vectors. Spectral feature extraction is performed on the adaptive band spectral data to obtain spectral absorption feature vectors.

6. The wastewater component identification and analysis method based on spectral and image recognition according to claim 5, characterized in that, The construction of the ternary coupling mapping to identify overlapping micro-regions of features specifically includes: For each micro-region unit, the polarization time-series evolution feature vector, interference texture feature vector, and spectral absorption feature vector are concatenated according to a unified feature dimension index to form the initial ternary coupling feature vector of each micro-region unit. Calculate the first, second, and third correlation coefficients between the polarization time-series evolution feature vector and the spectral absorption feature vector, the interference texture feature vector and the spectral absorption feature vector, and the polarization time-series evolution feature vector and the interference texture feature vector of each micro-region unit, and record the values ​​of the three correlation coefficients for that micro-region unit. When the absolute values ​​of at least two of the three correlation coefficients exceed the preset correlation threshold simultaneously, it is determined that the micro-region unit has multimodal feature coupling interference caused by mixed pollution, and it is marked as a feature overlapping micro-region and its spatial location coordinates are recorded.

7. The wastewater component identification and analysis method based on spectral and image recognition according to claim 6, characterized in that, The method of performing cross-modal separation based on surface tension gradient constraints and outputting a fingerprint feature set of contaminant components specifically includes: For each overlapping micro-region, the surface tension perturbation intensity values ​​of the micro-region and its adjacent micro-regions at each acquisition time node are extracted. The surface tension gradient components of the micro-region in two orthogonal directions are calculated based on the rate of change of its spatial position and combined to form a surface tension gradient vector. Using the surface tension gradient vector as a physical constraint, a joint separation objective function for the spectral-polarization-interference ternary coupling features is established, including separation sub-terms for spectral absorption features, polarization temporal evolution features, and interference texture features, as well as a surface tension gradient constraint regularization term; The contribution components of each separation sub-term in the joint separation objective function, belonging to different pollutant components, are set as variables to be solved. By minimizing and optimizing the joint separation objective function, the contribution component values ​​of each pollutant component in the spectral absorption feature vector, polarization time evolution feature vector, and interference texture feature vector are obtained. The spectral absorption contribution component, polarization time evolution contribution component, and interference texture contribution component corresponding to each pollutant component are classified and combined according to the component type to form the independent fingerprint features of each pollutant component. The summaries are output as the decoupled fingerprint feature set of pollutant components.

8. The wastewater component identification and analysis method based on spectral and image recognition according to claim 7, characterized in that, The fingerprint feature set of the pollutant components is input into a hierarchical recognition framework for major category discrimination and sub-category recognition, generating dynamic confidence weights, specifically including: The spectral absorption feature vectors of each micro-region in the fingerprint feature set of the pollution components are input into the upper classifier of the hierarchical recognition framework. The upper classifier makes a major classification of the pollution type based on the spectral absorption features and outputs the pollution category label corresponding to each micro-region. The polarization time-series evolution feature vector and interference texture feature vector corresponding to each micro-region belonging to the same pollution category label are input into the lower-level classifier of the hierarchical recognition framework. The lower-level classifier subdivides and identifies specific pollution types under the same pollution category based on the polarization time-series evolution features and interference texture features, and outputs the specific pollution type label of each micro-region. The spectral signal-to-noise ratio index is calculated based on the spectral absorption feature vector of each micro-region. The cross-modal consistency index is calculated based on the correlation between the polarization time evolution feature vector, the interference texture feature vector, and the spectral absorption feature vector. The surface tension consistency index is calculated based on the degree of matching between the surface tension gradient vector and the preset standard gradient mode. The three are then weighted together to generate a dynamic confidence weight.

9. The wastewater component identification and analysis method based on spectral and image recognition according to claim 8, characterized in that, When the confidence level is low, the first feedback path is triggered to adjust the surface tension gradient constraint and re-decouple it. Simultaneously, the second feedback path is triggered to optimize the polarization angle sequence and band selection strategy and re-execute, outputting wastewater component identification results with a confidence index, specifically including: When the dynamic confidence weight of a certain micro-region is lower than the preset confidence threshold, the micro-region is marked as a low confidence micro-region, and the specific values ​​of the spectral signal-to-noise ratio index, cross-modal consistency index and surface tension consistency index of the low confidence micro-region are extracted. Based on the distribution of the spectral signal-to-noise ratio index and the cross-modal consistency index, the local maxima in the neighborhood of the band below the preset signal-to-noise ratio threshold are used as the positions of the supplementary bands. The angle positions that need to be encrypted are determined according to the deviation direction of the cross-modal consistency index and the surface tension consistency index. Based on this, the polarization angle sequence and the hyperspectral band selection strategy are updated, and the acquisition, feature decoupling and recognition steps are re-executed. Based on the degree of deviation between the cross-modal consistency index and the preset standard value, the adjustment direction and adjustment range of the weight coefficient of the surface tension gradient constraint term in the joint separation objective function are determined, and the cross-modal feature separation and recognition steps are re-executed based on the adjusted weight coefficient to obtain the corrected fingerprint feature set of contaminant components; The pollution category labels, specific pollution type labels, and corresponding dynamic confidence weights of all micro-areas are summarized to generate and output wastewater component identification results with confidence index.

10. A wastewater component identification and analysis system based on spectral and image recognition, characterized in that, The wastewater component identification and analysis method based on spectral and image recognition as described in any one of claims 1-9 includes: The polarization-spectral adaptive acquisition module is used to acquire multi-polarization angle image sequences, calculate the polarization degree time-series evolution matrix to identify surface tension disturbance regions, extract the spatial frequency markers of oil film interference fringes to mark the prior regions of oil, fuse and divide regions of interest, adaptively select hyperspectral band subsets for acquisition, and output the sequence images of regions of interest and adaptive band spectral data volumes. The micro-region coupling and decoupling module is used to divide the sequence image of the region of interest and the adaptive band spectral data volume into micro-regions, extract polarization time evolution, interference texture and spectral absorption feature vectors, construct a ternary coupling mapping to identify overlapping micro-regions, implement cross-modal separation based on surface tension gradient constraints, and output a fingerprint feature set of contaminant components. The hierarchical identification and dual-path feedback module is used to input the fingerprint feature set of the pollutant components into the hierarchical identification framework for major category discrimination and sub-category identification, generate dynamic confidence weights, and trigger the first feedback path to adjust the surface tension gradient constraint in the micro-region coupling and decoupling module and re-decouple when the confidence is low. At the same time, it triggers the second feedback path to optimize the polarization angle sequence and band selection strategy in the polarization-spectrum adaptive acquisition module and re-execute it, outputting the wastewater component identification result with confidence index.

Images