Component imaging method and system based on point-surface spectral information fusion

Abstract

The invention discloses a component imaging method and system based on point-surface spectral information fusion. The method comprises the following steps: firstly, acquiring a first image and a first spectral data cube of a sample, and simultaneously acquiring spot scanning spectrums of a plurality of feature point locations and a second image containing the point locations; based on the two images, determining a space corresponding relation between the spot scanning spectrum and the surface scanning spectrum. After target spectral features are extracted from the spot scanning spectrum, a feature prediction model is trained by taking the surface scanning spectrum as an input variable and the corresponding target spectral features as target variables. And finally, a component distribution image can be generated through model prediction only by obtaining the spectral data cube of the sample to be detected. According to the method, the advantages of high speed and high spatial resolution of a surface scanning technology and the high-precision analysis capability of a point scanning technology are effectively fused through spatial correlation and feature migration of point-surface spectral data, and effective technical support is provided for surface component analysis.

Description

Technical Field

[0001] This application relates to the fields of component analysis and spectral imaging technology, and in particular to a component imaging method and system based on point-area spectral information fusion. Background Technology

[0002] Rapid, high-resolution imaging analysis of the chemical composition distribution on sample surfaces is of significant value in numerous fields, including materials science, biomedicine, and environmental monitoring. Achieving spatial visualization of component content provides crucial information for quality assessment and scientific research.

[0003] Currently, component analysis mainly relies on two types of technologies: high-precision point scanning spectroscopy can provide accurate component information, but its point-by-point scanning method results in extremely slow imaging speed, making it difficult to meet the needs of rapid detection; imaging spectroscopy has the advantages of rapid imaging and high spatial resolution, but its spectral information quality is limited and its analytical accuracy is insufficient, making it difficult to achieve reliable quantitative analysis.

[0004] This demonstrates that existing technologies inherently contradict each other, with high precision often resulting in low efficiency, and vice versa, failing to simultaneously meet the demands for rapid, high-resolution, and high-precision component analysis. This technological bottleneck severely restricts the detection capabilities and application effects in related fields. Therefore, there is an urgent need for a component imaging method that can simultaneously satisfy the requirements of rapid detection, high spatial resolution, and high precision. Summary of the Invention

[0005] To address the aforementioned issues, this application provides a component imaging method and system based on point-area spectral information fusion. The aim is to resolve the technical contradiction in existing technologies that cannot simultaneously achieve rapid detection, high spatial resolution, and high-precision component analysis, thereby realizing efficient and accurate visualization of component distribution.

[0006] The embodiments of this application disclose the following technical solutions:

[0007] The first aspect of this application provides a compositional imaging method based on point-area spectral information fusion, the method comprising:

[0008] A first image and a first spectral data cube of the sample are acquired, and point scan spectra of multiple feature points of the sample and a second image containing the feature points are acquired; wherein, the first image and the first spectral data cube are acquired by surface scanning, and the point scan spectra are acquired by point measurement.

[0009] Based on the second image and the first image, the surface scan spectrum corresponding to the point scan spectral space is determined from the first spectral cube;

[0010] Target spectral features are extracted from the point scan spectrum; the target spectral features are spectral features used to characterize the target components.

[0011] Using the surface scan spectrum as the input variable and the target spectral features extracted from the point scan spectrum corresponding to the surface scan spectrum as the target variable, a feature prediction model is trained.

[0012] A second spectral data cube of the sample to be tested is obtained. The predicted spectral features of each pixel in the second spectral data cube are predicted using the feature prediction model. Based on the predicted spectral features of all pixels in the second spectral data cube, a compositional distribution image of the sample to be tested is generated.

[0013] In an optional implementation, determining the surface scan spectrum corresponding to the point scan spectral space from the first spectral cube based on the second image and the first image includes:

[0014] Based on the second image, determine the spatial coordinates of the feature points corresponding to the scanning spectrum of each point;

[0015] Spatial registration is performed between the second image and the first image to obtain a coordinate mapping relationship;

[0016] Based on the coordinate mapping relationship, and using the spatial coordinates of the feature points corresponding to each point scan spectrum, the surface scan spectrum corresponding to the point scan spectrum space is determined from the first spectral data cube.

[0017] In an optional implementation, the target spectral features include the characteristic peak intensity, peak area, and spectral line ratio of the point scan spectrum, or a feature vector obtained after preprocessing and feature extraction of the point scan spectrum.

[0018] In an optional implementation, extracting the target spectral features from the point scan spectrum includes:

[0019] Based on prior knowledge, target spectral features related to the target component are extracted from the point scan spectrum;

[0020] Alternatively, if the point scan spectrum is a mixed spectrum containing both the target component and interfering components, a spectral demixing method is used to extract the target spectral features related to the target component from the mixed spectrum, and the content distribution of the target component at the sample points corresponding to the mixed spectrum is calculated.

[0021] In optional implementations, the spectral unmixing method includes multivariate curve-resolved alternating least squares method or nonnegative matrix factorization method.

[0022] In an optional implementation, the training to obtain a feature prediction model includes:

[0023] A teacher model is trained based on the point scan spectrum and the target spectral features extracted from the point scan spectrum.

[0024] Using the surface scan spectrum as input and the output of the teacher model as the target to be imitated, a student model is trained; wherein, the student model is the feature prediction model.

[0025] In an optional implementation, the area scanning is achieved using a high spatial resolution imaging device, which includes at least one of a hyperspectral imaging device, a multispectral imaging device, or a short video imaging device.

[0026] In an optional implementation, multiple light sources are used for illumination when performing surface scanning using the high spatial resolution imaging device.

[0027] In an optional implementation, the point measurement is achieved using a high-precision spectroscopic device, which includes at least one of laser-induced breakdown spectroscopy, Raman spectroscopy, visible / near-infrared spectroscopy, X-ray fluorescence spectroscopy, or inductively coupled plasma mass spectrometry.

[0028] A second aspect of this application provides a component imaging system based on point-area spectral information fusion, the system comprising:

[0029] The acquisition module is used to acquire a first image and a first spectral data cube of the sample, and to acquire point scan spectra of multiple feature points of the sample and a second image containing the feature points; wherein, the first image and the first spectral data cube are acquired by surface scanning, and the point scan spectra are acquired by point measurement.

[0030] The corresponding module is used to determine, based on the second image and the first image, the surface scan spectrum corresponding to the point scan spectral space from the first spectral cube;

[0031] An extraction module is used to extract target spectral features from the point scan spectrum; the target spectral features are spectral features used to characterize the target components;

[0032] The training module is used to train a feature prediction model by using the surface scan spectrum as an input variable and the target spectral features extracted from the point scan spectrum corresponding to the surface scan spectrum as the target variable.

[0033] An imaging module is used to acquire a second spectral data cube of the sample to be tested, use the feature prediction model to predict the predicted spectral features of each pixel in the second spectral data cube, and generate a compositional distribution image of the sample to be tested based on the predicted spectral features of all pixels in the second spectral data cube.

[0034] Compared with the prior art, this application has the following beneficial effects:

[0035] In this application's technical solution, a first image and a first spectral data cube of the sample are first acquired, along with point scan spectra of multiple feature points and a second image containing those feature points. Then, based on the second and first images, a surface scan spectrum corresponding to the point scan spectral space is determined from the first spectral cube. Next, target spectral features for characterizing the target component are extracted from the point scan spectrum. Then, using the surface scan spectrum as input variables and the target spectral features extracted from the point scan spectrum corresponding to the surface scan spectrum as target variables, a feature prediction model is trained. Finally, a second spectral data cube of the sample to be tested is acquired, and the feature prediction model is used to predict the predicted spectral features of each pixel in the second spectral data cube. Based on the predicted spectral features of all pixels in the second spectral data cube, a component distribution image of the sample to be tested is generated. It is evident that this application's technical solution deeply integrates the high spatial resolution and rapid scanning advantages of surface scanning technology with the high-precision spectral characterization capabilities of point scan spectroscopy technology. It establishes a precise correlation between the two types of data through spatial registration and then achieves the mapping and transfer from surface scan spectra to high-precision target spectral features through model training. This solution requires only a small amount of point scanning measurements in the initial modeling stage, while in actual detection applications, only rapid surface scanning is needed to obtain high-precision composition distribution results. It fundamentally solves the technical problem of balancing detection speed, spatial resolution, and analytical accuracy between point scanning technology and imaging spectroscopy technology, and provides an effective technical means for rapid and accurate analysis of surface composition in various fields. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 A schematic diagram illustrating the principle of component imaging based on point-area spectral information fusion provided in this application embodiment;

[0038] Figure 2 A flowchart of a component imaging method based on point-area spectral information fusion is provided for embodiments of this application;

[0039] Figure 3 This is a schematic diagram illustrating the training of a feature prediction model based on standard availability, provided in an embodiment of this application.

[0040] Figure 4 A comparison chart of quantitative analysis results of nickel oxide content provided for embodiments of this application;

[0041] Figure 5 This application provides a comparison image of non-uniform samples imaged using different measurement methods in its embodiments.

[0042] Figure 6 A schematic diagram showing the predicted and measured values ​​of the intensity of all characteristic spectral lines of a test sample, provided for an embodiment of this application;

[0043] Figure 7 A comparison image of imaging results of spectral line intensities in a test sample provided for an embodiment of this application;

[0044] Figure 8 This is a schematic diagram of the structure of a component imaging system based on point-area spectral information fusion, provided as an embodiment of this application. Detailed Implementation

[0045] As described earlier, rapid, high-resolution imaging analysis of the chemical composition or content distribution on sample surfaces is of great value in fields such as materials science, biomedicine, environmental monitoring, agricultural product testing, and industrial process control. For example, in applications such as material surface defect detection, biological tissue pathology analysis, soil pollutant distribution assessment, and non-destructive testing of agricultural product quality, compositional imaging can intuitively reflect the spatial distribution characteristics of target analytes, providing crucial information for quality assessment, process optimization, and scientific research.

[0046] High-precision point-scanning spectroscopic instruments, such as Laser-Induced Breakdown Spectroscopy (LIBS), Raman Spectroscopy (Raman), and X-ray Fluorescence (XRF), acquire local spectral information of samples through point-by-point acquisition. LIBS utilizes laser ablation of the sample to generate plasma, and analyzes its emission spectrum to achieve qualitative and quantitative elemental identification; Raman spectroscopy uses inelastic scattered light to reflect molecular vibrational information, suitable for structural identification of organic and inorganic substances; XRF uses X-rays to excite the sample to generate characteristic X-rays for elemental composition analysis. These instruments share the advantages of rich spectral information, high measurement accuracy, and the ability to provide accurate spectral fingerprints. However, they typically employ point-by-point scanning, requiring significant time to achieve full sample coverage, and their spatial resolution is limited by spot size or scan step size. Furthermore, benchtop equipment is bulky and expensive, making them unsuitable for on-site or in-situ applications.

[0047] In contrast, imaging spectroscopy techniques such as hyperspectral imaging (HSI) and multispectral imaging (MSI) employ line or area scanning methods to acquire image data of a sample across multiple spectral bands in a single scan, forming a spectral data cube with high spatial resolution. Hyperspectral imaging obtains nearly continuous spectral curves through continuous narrow-band acquisition, while multispectral imaging acquires images in a few specific bands. The advantages of these techniques lie in their fast imaging speed and high spatial resolution, enabling rapid scanning of large-area samples. However, core performance parameters such as spectral resolution and signal-to-noise ratio (SNR) are typically limited in spectral dimensions, resulting in low spectral resolution, narrow band range, low SNR, or insufficient effective spectral feature information. This leads to poor accuracy in complex component analysis and makes them unsuitable for direct use in high-reliability quantitative analysis.

[0048] Therefore, a single technology cannot simultaneously meet the needs of rapid, high spatial resolution, and high-precision component analysis.

[0049] To address the aforementioned problems, the inventors have proposed a component imaging method and system based on the fusion of point and surface spectral information.

[0050] First, a first image and a first spectral data cube of the sample are acquired, along with point scan spectra of multiple feature points and a second image containing those feature points. Then, based on the second and first images, a surface scan spectrum corresponding to the point scan spectral space is determined from the first spectral cube. Next, target spectral features for characterizing the target component are extracted from the point scan spectrum. Then, using the surface scan spectrum as input variables and the target spectral features extracted from the point scan spectrum corresponding to that surface scan spectrum as target variables, a feature prediction model is trained. Finally, a second spectral data cube of the sample is acquired, and the feature prediction model is used to predict the predicted spectral features of each pixel in the second spectral data cube. Based on the predicted spectral features of all pixels in the second spectral data cube, a component distribution image of the sample is generated. It is evident that the technical solution of this application deeply integrates the high spatial resolution and rapid scanning advantages of surface scanning technology with the high-precision spectral characterization capability of point scan spectroscopy technology. It establishes a precise correlation between the two types of data through spatial registration, and then achieves the mapping and transfer from surface scan spectra to high-precision target spectral features through model training. This solution requires only a small number of point scan measurements in the initial modeling stage, while in actual detection applications, only rapid surface scanning is needed to obtain high-precision composition distribution results. It fundamentally solves the technical problem of existing technologies that make it difficult to balance detection speed, spatial resolution and analysis accuracy, and provides effective technical support for rapid and accurate analysis of surface composition in various fields.

[0051] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0052] Figure 1 This is a schematic diagram illustrating the principle of component imaging based on point-area spectral information fusion, provided as an embodiment of this application. Figure 1 As shown, existing high-precision point scanning technologies, such as laser-induced breakdown spectroscopy, can provide rich spectral fingerprint information, but their spatial resolution is low and the time consumption is long; while imaging spectroscopy technologies, such as hyperspectral imaging, have the advantages of high spatial resolution and fast scanning, but their spectral information accuracy is insufficient and it is difficult to use them directly for high-reliability quantitative analysis.

[0053] The core idea of ​​this application lies in resolving this contradiction through the fusion of point and area spectral information. Specifically, during the training phase, a feature transformation model is trained using high-precision point scan data as ground truth. This model can learn and predict high-precision spectral features from rapid area scan data. During the testing / application phase, for the sample to be tested, only a rapid area scan is required. The trained model can then be used to transform the sample into a compositional distribution image that combines the advantages of high information content, high spatial resolution, and speed, thus achieving a breakthrough in technical performance.

[0054] Based on the above principles, see Figure 2 The figure shows a flowchart of a component imaging method based on point-area spectral information fusion provided in an embodiment of this application. The method can be executed by a dedicated spectral imaging system, which typically integrates high spatial resolution imaging equipment and high-precision spectroscopic equipment, with process control and data processing handled by a computing unit (such as a computer). Figure 2 As shown, the method includes the following steps:

[0055] S201. Acquire the first image and the first spectral data cube of the sample, and acquire the point scan spectrum of multiple feature points of the sample and the second image containing the feature points.

[0056] In this embodiment, the first image is a digital image covering the entire sample acquired through surface scanning, which can intuitively present the spatial distribution morphology of the sample surface. The first spectral data cube is a three-dimensional data structure (height × width × number of bands) generated synchronously during the imaging process. Each pixel contains a curve reflecting the spectral response of the material at that location, thereby achieving a unified global characterization of the image and spectrum.

[0057] Point scanning spectroscopy is spectral data obtained using high-precision spectroscopic equipment at multiple pre-selected feature points on the sample surface through point measurement. Its characteristics include rich spectral information and strong fingerprinting, making it a precise benchmark for component analysis. The second image is a digital image containing spatial information of all feature points, captured simultaneously during point scanning spectroscopy acquisition. Its core function is to provide coordinate references for subsequent spatial registration, ensuring a one-to-one spatial correspondence between point measurement data and area scanning data.

[0058] In one alternative implementation, area scanning is achieved using a high spatial resolution imaging device. High spatial resolution imaging devices include, but are not limited to, at least one of hyperspectral imaging devices, multispectral imaging devices, or short-video imaging devices. The advantage of such devices lies in their extremely high imaging speed and spatial resolution, enabling rapid global scanning of large-sized samples. However, the spectral information directly acquired by these devices typically has inherent limitations in dimensionality and accuracy. For example, it may exhibit one or more of the following: low spectral resolution, narrow spectral range, low signal-to-noise ratio, or insufficient effective information content. This limits its ability to be directly used for high-reliability quantitative analysis.

[0059] Optionally, to compensate for the inherent limitations of such devices in terms of spectral dimension, resolution, or signal-to-noise ratio, multiple light source illumination can be used when performing area scanning with a high spatial resolution imaging device. For example, by sequentially illuminating the sample with light sources of different wavelengths or band combinations, richer and more discriminative optical response signals can be generated, thereby improving the quality of the original spectral data and laying a more reliable foundation for subsequent model training.

[0060] In one alternative implementation, point measurements are performed using high-precision spectroscopic equipment. Optionally, high-precision spectroscopic equipment includes, but is not limited to, at least one of laser-induced breakdown spectroscopy, Raman spectroscopy, visible / near-infrared spectroscopy, X-ray fluorescence spectroscopy, or inductively coupled plasma mass spectrometry. The core advantage of this type of equipment is its ability to provide extremely accurate and rich spectral fingerprint information, enabling accurate qualitative or quantitative analysis of target components. However, its main drawback is that it typically employs a point-by-point measurement method, which can lead to extremely long measurement times when pursuing high spatial resolution imaging, and the spatial resolution is limited by the equipment's spot size or sampling step size.

[0061] In one example implementation, a high spatial resolution imaging device (such as short video imaging) is used to scan the sample surface. First, a first digital image of the sample is acquired. Its spatial resolution is Pixels. Then, the device performs line or area scanning to acquire a cube of spectral data corresponding to the image. Each pixel contains spectral information across p bands. A high-precision point-scanning spectral device (such as laser-induced breakdown spectroscopy) is used to select m feature points on the sample surface for spectral measurements, obtaining a point-scanning spectral data matrix. d represents the number of spectral bands. Simultaneously, a second digital image containing these points is acquired using the device's built-in positioning and imaging unit. Its resolution is Pixel.

[0062] This step, through the aforementioned collaborative acquisition scheme, provides both high-spatial-coverage area scan data and high-precision point scan reference data for subsequent processing, forming the physical basis for point-area information fusion.

[0063] S202. Based on the second image and the first image, determine the surface scan spectrum corresponding to the point scan spectral space from the first spectral cube.

[0064] In the embodiments of this application, spatial correspondence refers to aligning point scan points with area scan pixels in space using image registration technology, thereby extracting spectral data at the corresponding locations.

[0065] The surface scan spectrum is the average spectral curve of a pixel or region that is extracted from the first spectral data cube and matched with the point scan point, representing the surface optical properties at that location.

[0066] In one alternative implementation, the specific steps for determining the surface scan spectrum corresponding to the point scan spectral space from the first spectral cube based on the second image and the first image include:

[0067] S2021. Based on the second image, determine the spatial coordinates of the feature points corresponding to the scan spectrum of each point.

[0068] In one example implementation, feature points can be automatically identified using image processing algorithms. For instance, for ablation points left by laser-induced breakdown spectroscopy or focal points marked during X-ray fluorescence spectroscopy, circle detection algorithms or spot analysis algorithms can be used to locate the center pixel coordinates of these traces, forming a coordinate set. , where m is the number of feature points.

[0069] In another example implementation, if the high-precision spectral equipment and positioning imaging device are rigorously calibrated and have a high-precision displacement platform and preset scanning path, the precise coordinates of the feature point corresponding to each point scan in the second image can be directly calculated based on the calibration relationship between the preset scanning path and the imaging optical path of the equipment.

[0070] This step digitizes the physical location of the point scan, providing an indispensable spatial reference for subsequent image registration and data extraction.

[0071] S2022. Spatial registration is performed between the second image and the first image to obtain the coordinate mapping relationship.

[0072] In this embodiment of the application, since the first image and the second image may be acquired by different devices, at different viewpoints or resolutions, and there are spatial differences such as translation, rotation, and scaling, a strict coordinate mapping relationship between the two must be established through a registration algorithm.

[0073] In one example implementation, a feature-based registration algorithm can be used. First, stable feature points are extracted from both the second and first images, for example using Scale-Invariant Feature Transform (SIFT), Oriented FAST and Rotated BRIEF (ORB), or Speeded-Up Robust Features (SURF) algorithms. Then, matching is performed using feature descriptors, and mismatched point pairs are eliminated using the Random Sample Consensus (RANSAC) algorithm. Finally, an optimal spatial transformation model (such as an affine transformation or projection transformation matrix) is estimated, which defines the coordinate mapping relationship R from the first image coordinate system to the second image coordinate system.

[0074] In another implementation, if the two imaging systems are relatively fixed and their approximate positional relationship is known, a template- or region-based registration method can be used to solve for more accurate rigid transformation parameters by optimizing similarity measures such as mutual information or correlation coefficients.

[0075] This step eliminates spatial inconsistencies caused by different imaging systems and establishes a precise and quantifiable coordinate transformation rule.

[0076] S2023. Based on the coordinate mapping relationship and the spatial coordinates of the feature points corresponding to each point's scan spectrum, determine the surface scan spectrum corresponding to the point scan spectrum space from the first spectral data cube.

[0077] In this embodiment of the application, the established mapping relationship is used to accurately determine the spectral data corresponding to each high-precision measurement point from the huge cube of surface scan data, thus completing the precise pairing of point and surface data.

[0078] In one example implementation, the coordinate mapping relationship R obtained in step S2022 is first used to map the first spectral data cube. Perform geometric transformations and resampling, such as using bilinear or cubic convolution interpolation, to generate a result similar to the second image. A new data cube that is perfectly aligned in space Then, based on the set of point scan coordinates obtained in step S2021, the data is directly extracted from the aligned data cube. Extract the p-dimensional spectral vector of the corresponding pixel to form an area scanning spectral matrix. In another implementation, to reduce the impact of random noise and minor registration errors, a small neighborhood (e.g., a 3×3 pixel region) centered on the feature point coordinates can be extracted. The average spectrum of all pixels within this region is calculated, and this average spectrum is used as the area scan spectrum for that point. Finally, the extracted area scan spectrum matrix is ​​compared with the point scan spectrum matrix. Pair the points one by one in order to construct a paired dataset for modeling.

[0079] This step achieves precise spatial alignment of point and area data, ensuring that point scan spectra and area scan spectra are paired at the same physical location, providing a consistent dataset for subsequent feature extraction and model training. This improves the accuracy and reliability of the feature prediction model.

[0080] S203. Extract target spectral features from point scan spectrum.

[0081] In this embodiment, the target spectral features are spectral information extracted from point scan spectra that can specifically characterize the target components. They are core indicators reflecting the types and contents of sample components. Essentially, they transform complex spectral curves into feature parameters with clear physical or chemical meanings, facilitating subsequent model learning and prediction.

[0082] Target components refer to specific chemical components (such as nickel oxide and magnesium elements) or characteristic substances associated with physical properties that are determined based on specific testing scenarios such as materials, geology, biology, environment, and agriculture, possess specific spectral characteristics that can be captured by spectroscopic technology, and whose spatial distribution needs to be obtained through qualitative identification or quantitative analysis.

[0083] In one alternative implementation, characteristic peak intensities, peak areas, or intensity ratios of specific spectral lines related to the target analyte can be directly extracted based on prior knowledge. For example, when analyzing magnesium using LIBS, the peak intensity of the MgI 383.8 nm spectral line can be directly extracted as the target feature.

[0084] In another alternative implementation, the original point scan spectrum can be preprocessed by smoothing, baseline correction, and standardization, and then dimensionality reduction methods such as principal component analysis can be used to extract the main principal component scores as feature vectors to compress the data dimensionality and retain most of the effective information.

[0085] Furthermore, when the point scan spectrum is a signal from a complex mixture, i.e., containing both the target component and interfering components, this step can employ spectral demixing methods to extract purer target features. For example, using multivariate curve-resolved alternating least squares or nonnegative matrix factorization, the endmember spectra representing the pure component and their abundance (i.e., content distribution) at each point can be decomposed from the mixed spectral data matrix at multiple points. In this case, the extracted target spectral features can be the endmember spectral vector of the pure component, or the content value of that component at the corresponding point.

[0086] This step transforms raw spectral data into effective features, filtering out key information directly related to the target components and eliminating irrelevant noise and interference signals. The extracted target spectral features simplify the data dimensions, reduce the complexity of model training, and highlight the specificity of component features, providing high-quality target variables for building accurate prediction models.

[0087] S204. Using the surface scan spectrum as the input variable and the target spectral features extracted from the point scan spectrum corresponding to the surface scan spectrum as the target variable, a feature prediction model is trained.

[0088] In this embodiment of the application, the area scan spectrum obtained in step S202 and paired with the feature point is used as the input variable, and the target spectral features extracted from the corresponding point scan spectrum in step S203 are used as the target variable. The feature prediction model is trained to establish the mapping relationship from area scan spectrum to point scan spectral features.

[0089] In practical applications, depending on whether standard samples are available, this application provides two model training paths. See also Figure 3 , Figure 3 This is a schematic diagram illustrating the training of a feature prediction model based on standard sample availability, provided in an embodiment of this application. Figure 3 The paper illustrates the knowledge distillation path with standard samples and the direct mapping path without standard samples. In an alternative implementation, for cases without standard samples (i.e., lacking standard samples with known component contents), algorithms such as partial least squares regression, support vector machines, or neural networks are used. Using area scan spectra as input and target spectral features extracted from corresponding point scan spectra as the objective, an end-to-end mapping model is trained as the feature prediction model. This method achieves approximate estimation of elemental content by establishing a nonlinear mapping relationship from color information to spectral features. Its advantage lies in its independence from prior standard sample information, thus having a wider range of applications.

[0090] In another alternative implementation, when standard samples are available, training employs a knowledge distillation strategy. As shown in the figure, this path enhances accuracy from color level to elemental content through precise knowledge transfer. The training process specifically includes the following steps:

[0091] Step 1: Train the teacher model based on the point scan spectrum and the target spectral features extracted from the point scan spectrum.

[0092] In the embodiments of this application, the teacher model is a complex and high-performance model based on standard samples, which can accurately reflect the quantitative relationship between spectral line intensity and elemental content.

[0093] Step 2: Using the area scan spectrum as input and the output of the teacher model as the target to be imitated, the student model is trained to obtain the model.

[0094] In this embodiment, a distillation loss function (such as the Körbeck-Leibler divergence) can be introduced to constrain the student model to mimic the behavior of the teacher model, thereby achieving knowledge transfer from precise measurement to rapid estimation. This knowledge transfer process enables the student model to inherit the accuracy and stability of the teacher model.

[0095] The trained student model is the final feature prediction model. It maintains the efficiency of area scanning technology while achieving accuracy close to point scanning technology through knowledge distillation. It possesses stronger generalization ability and transferability, and can be directly applied to rapid component analysis of similar samples. This step realizes knowledge transfer from area scanning spectroscopy to high-precision point scanning features, enabling the model to predict. The trained feature prediction model can generalize to new samples, avoiding repeated point scanning and significantly improving analytical efficiency.

[0096] S205. Obtain the second spectral data cube of the sample to be tested, use the feature prediction model to predict the predicted spectral features of each pixel in the second spectral data cube, and generate the compositional distribution image of the sample to be tested based on the predicted spectral features of all pixels in the second spectral data cube.

[0097] In this embodiment, for a new sample requiring component analysis, only the high spatial resolution imaging device is needed to quickly scan it to obtain a second spectral data cube of the sample. This second spectral data cube is identical in data structure and attributes to the first spectral data cube obtained in step S201, containing spectral information for each pixel of the sample in multiple bands, and serves as the raw data for generating the global component distribution map. It is worth emphasizing that this process completely eliminates the need for time-consuming measurements using a high-precision point-scanning spectral device, thereby reducing the time for full-sample component imaging from hours or even days of traditional point scanning to minutes or seconds of area scanning, achieving an order-of-magnitude speed improvement while significantly reducing equipment operating costs.

[0098] In practice, the spectral data of each pixel in the second spectral data cube of the sample to be tested is input one by one into the feature prediction model trained in step S204. The feature prediction model outputs one or more predicted spectral features for each pixel in parallel or serially. These features can be quantitative estimates of elemental content, compound concentration, or specific physical properties. For example, in mineral analysis, the model can output the percentage content of iron; in biological tissue analysis, it can output the relative concentration of specific proteins or metabolites.

[0099] Then, based on the prediction results of all pixels, a multidimensional component feature matrix with the same spatial resolution as the original data cube can be constructed. Using specialized image processing and visualization techniques, the feature values ​​in this matrix are mapped to different pseudo-color or grayscale levels, generating an intuitive, high spatial resolution image of the target component / content distribution. In this image, the hue, saturation, or brightness of the colors directly encode the two-dimensional spatial distribution and relative abundance information of the target component on the sample surface, providing crucial visualization information for applications such as phase distribution observation in materials analysis, defect identification in industrial testing, lesion area localization in biomedicine, and pollutant diffusion assessment in environmental monitoring.

[0100] In an extended application scenario, the area scanning spectral acquisition, point scanning spectral acquisition, and sample data acquisition can all be performed on the same sample. For example, sparse but crucial point scanning data can be acquired only from a representative region (e.g., 5%-10% of the total area) of a geological sample or biological tissue slice. Pure spectral features and local content information can be extracted from these data using methods such as spectral unmixing. Combined with the complete area scanning data of the sample, a trained feature prediction model can be used for spatial extrapolation to accurately predict the spectral feature distribution of all unscanned areas of the sample. This achieves global high-resolution component imaging based on a minimal amount of high-precision measurements. While ensuring overall analytical accuracy, this minimizes the time-consuming high-precision measurements, providing a practical technical path for large-scale, high-throughput sample screening.

[0101] This application's embodiments creatively integrate high spatial resolution area scanning technology with high-precision point scanning technology through the above-described technical solution, and utilize machine learning models to achieve knowledge transfer and mapping from area scanning spectral features to point scanning spectral features. This method effectively overcomes the technical bottleneck of single spectral techniques in balancing scanning speed, spatial resolution, and analytical accuracy. While fully retaining the high efficiency and high spatial resolution advantages of area scanning technology, it significantly improves the accuracy and reliability of component analysis.

[0102] Specifically, this solution generates high-resolution compositional images from samples with only a single rapid surface scan, significantly improving detection efficiency. By "distilling" feature information from high-precision point scan data, the prediction results based on surface scan data approach the accuracy of point scans. Its flexible architecture allows users to adjust the point scan density and equipment combination according to actual needs, combining local point scan and global prediction modes to greatly reduce the dependence on expensive high-precision equipment and the associated costs for high-throughput detection. Furthermore, this solution does not rely on specific spectral techniques, has good equipment compatibility and technological inclusiveness, and can be widely applied in multiple fields such as materials science, biomedicine, geological exploration, environmental monitoring, agricultural product safety, and industrial process quality control.

[0103] Based on the foregoing embodiments, in order to more clearly demonstrate the specific application and significant effects of the technical solution of this application, the following detailed description is provided through two specific embodiments.

[0104] Example 1:

[0105] This application uses a mixed sample of nickel oxide and zirconium oxide as an example to illustrate the specific process of achieving rapid and high-resolution imaging analysis of nickel oxide content based on laser-induced breakdown spectroscopy-short video imaging fusion.

[0106] The experiment used a benchtop LIBS system as a point scanning spectrometer, with the Nd:YAG laser having a wavelength range of 186.9–979.2 nm and a spectral resolution of 0.09 nm. Simultaneously, a smartphone video imaging (SVI) device was used as a surface scanning spectrometer. Gradient color illumination was generated through the phone screen, and the front-facing camera was used to record a 6-second video of the color change on the sample surface, with a resolution of 1080×1920 pixels.

[0107] First, homogeneous samples were prepared by mixing pure nickel oxide and zirconium oxide, with the nickel oxide ratio increasing from 0% to 100% in a 5% concentration gradient. Two samples were prepared for each concentration, for a total of 42 samples, which were used to establish a nickel oxide content prediction model.

[0108] In addition, a non-homogeneous sample was prepared for subsequent imaging analysis to reflect the spatial distribution of nickel oxide content. For the homogeneous sample, LIBS and SVI signals were acquired at 25 locations in a 5×5 region for each sample. The LIBS data underwent preprocessing, including peak identification, baseline removal, and peak area calculation, and spectral lines unrelated to nickel and zirconium were removed, ultimately retaining 648 variables. The SVI data were standardized (values ​​divided by 255), resulting in 606 variables. From the 42 homogeneous samples, four samples with nickel oxide contents of 20%, 40%, 60%, and 80% were randomly selected as the validation set, five samples with contents of 10%, 30%, 50%, 70%, and 90% were selected as the test set, and the remainder were used as the training set.

[0109] A neural network model is used to first normalize the input features through layers, then compress them to 32 dimensions through fully connected layers. After passing through a Rectified Linear Unit (ReLU) activation function, layer normalization is applied again, finally outputting a 1-dimensional prediction result. Furthermore, a model based on LIBS data is used as the teacher model, and a model based on SVI data is used as the student model. Linear layers and a Hyperbolic Tangent (Tanh) activation function are used to align the student features (32 dimensions) with the teacher features (32 dimensions). The loss function uses Mean Absolute Error (MAE) as the task loss and cosine loss as the distillation loss, with a weight ratio of 0.9:0.1. Figure 4The image shows a comparison of the quantitative analysis results of nickel oxide content provided in this application embodiment. From left to right, the results are LIBS measurement, SVI measurement, and the fusion measurement result of this method. The LE-SVI shown in the image is the imaging result generated by the feature prediction model trained by the aforementioned LIBS-SVI fusion method and knowledge distillation strategy. It represents a knowledge-enhanced smartphone short video imaging analysis capability. It can be seen that the performance of LIBS-SVI fusion measurement is significantly better than that of single SVI measurement.

[0110] Further imaging analysis was performed on the non-uniform samples. LIBS acquired 25×25 scan points within a specific region, while SVI acquired images of 315×315 pixels within the same region. The two types of data were processed using the aforementioned preprocessing methods and a neural network model. Figure 5 This document presents a comparative image of non-uniform sample imaging using different measurement methods, as provided in an embodiment of this application. LIBS can predict nickel oxide content with high accuracy, but its spatial resolution is low and the scanning time is approximately 5 minutes. SVI has poor prediction accuracy, but its spatial resolution is high, the scanning speed is fast (only 6 seconds), and the cost is low. By fusing LIBS and SVI using the component imaging method based on point-area spectral information fusion provided in this embodiment of the application, the resulting nickel oxide content imaging maintains high spatial resolution while effectively improving its quantitative accuracy.

[0111] Example 2:

[0112] In this embodiment of the application, a gneiss slice is used as an example to illustrate a specific implementation method for achieving rapid and high-resolution imaging analysis of magnesium content based on LIBS-SVI fusion.

[0113] LIBS data measurement used the same desktop LIBS system as in Example 1; SVI data measurement used a mobile phone screen to generate gradient color lighting, and simultaneously used the front-facing camera to record a 6-second video of the color change on the sample surface, with a resolution of 1080×1920 pixels.

[0114] Twenty rock slice samples were prepared, with 14 used as training samples, 3 as validation samples, and 3 as test samples. Within the effective area (7 mm²) of each slice sample, LIBS acquired 18×18 spectra, and SVI acquired images containing 138×138 pixels. First, image registration techniques were used to extract the corresponding SVI signals based on the actual locations of the LIBS scan points. After preprocessing the LIBS data, including peak identification, baseline removal, and peak area calculation, only the Mg I 383.8 nm spectral line intensity was retained as a characterization of magnesium content. The SVI data was then standardized. Subsequently, a neural network model was used to establish a mapping relationship from SVI data to LIBS spectral line intensity (Mg I 383.8 nm) on the training set. This network treats the input data as a three-channel RGB structure, first uniformly dividing the input features into three parts, then extracting features through sub-networks with shared weights, and finally concatenating the three-channel features and fusing them through a fully connected network to output the predicted value.

[0115] Figure 6 This is a schematic diagram of the predicted and measured values ​​of the intensity of all characteristic spectral lines of a test sample provided in an embodiment of this application. The characteristic spectral line is Mg I 383.8 nm, with a mean absolute error of prediction (MAEP) of 0.036 and a root mean square error of prediction (RMSEP) of 0.07. Figure 7 This is a comparison image of the imaging results of spectral line intensity in a test sample provided in an embodiment of this application, wherein the spectral line is Mg I 383.8 nm. Experiments show that although color information itself is not directly related to the content of a specific element, by establishing a mapping relationship between the SVI signal and the LIBS spectral line intensity, rapid, high-resolution, and low-cost spatial distribution imaging of element content can be achieved.

[0116] Based on the component imaging method based on point-area spectral information fusion provided in the foregoing embodiments, this application also provides a component imaging system based on point-area spectral information fusion. Figure 8 This is a schematic diagram of the structure of a component imaging system based on point-area spectral information fusion, provided as an embodiment of this application. Figure 8 As shown, the component imaging system based on point-area spectral information fusion includes: an acquisition module 801, a correspondence module 802, an extraction module 803, a training module 804, and an imaging module 805.

[0117] The acquisition module 801 is used to acquire a first image and a first spectral data cube of the sample, and to acquire point scan spectra of multiple feature points of the sample and a second image containing the feature points; wherein, the first image and the first spectral data cube are acquired by surface scanning, and the point scan spectra are acquired by point measurement.

[0118] The corresponding module 802 is used to determine the surface scan spectrum corresponding to the point scan spectral space from the first spectral cube based on the second image and the first image;

[0119] Extraction module 803 is used to extract target spectral features from point scan spectrum; the target spectral features are spectral features used to characterize the target components;

[0120] Training module 804 is used to train a feature prediction model by taking the surface scan spectrum as the input variable and the target spectral features extracted from the point scan spectrum corresponding to the surface scan spectrum as the target variable.

[0121] The imaging module 805 is used to acquire the second spectral data cube of the sample to be tested, predict the predicted spectral features of each pixel in the second spectral data cube using a feature prediction model, and generate a compositional distribution image of the sample to be tested based on the predicted spectral features of all pixels in the second spectral data cube.

[0122] This embodiment of the application achieves a complementary and effective integration of point scanning and area scanning technologies by combining the functions of the acquisition module 801, the corresponding module 802, the extraction module 803, the training module 804, and the imaging module 805. The imaging module 805, as the core output unit of the system, rapidly converts the spectrum of each pixel in the second spectral data cube acquired by the area scanning of the sample into quantitatively significant predicted spectral features by calling the trained feature prediction model. Based on this, it generates a high spatial resolution component distribution image, ultimately achieving component analysis capabilities approaching the accuracy of point scanning technology while maintaining the high efficiency and spatial resolution advantages of area scanning. Through the streamlined collaboration of the above modules, the entire system effectively solves the industry challenge of traditional single-technology approaches failing to balance detection speed, spatial resolution, and analytical accuracy, providing a reliable system-level solution for rapid component distribution analysis in fields such as materials science, biomedicine, and environmental monitoring.

[0123] In the optional implementation, module 802 is specifically used for:

[0124] Based on the second image, determine the spatial coordinates of the feature points corresponding to the scanned spectra of each point;

[0125] Spatial registration is performed between the second image and the first image to obtain the coordinate mapping relationship;

[0126] Based on the coordinate mapping relationship, and using the spatial coordinates of the feature points corresponding to each point's scan spectrum, the surface scan spectrum corresponding to the point scan spectrum space is determined from the first spectral data cube.

[0127] In the optional implementation, the target spectral features extracted by the extraction module 803 include the characteristic peak intensity, peak area, and spectral line ratio of the point scan spectrum, or the feature vector obtained after preprocessing and feature extraction of the point scan spectrum.

[0128] In an optional implementation, the extraction module 803 includes a first extraction unit and a second extraction unit.

[0129] The first extraction unit is used to extract target spectral features related to the target component from the point scan spectrum based on prior knowledge.

[0130] The second extraction unit is used to extract target spectral features related to the target component from the mixed spectrum using a spectral demixing method when the point scan spectrum is a mixed spectrum containing the target component and interfering components, and to calculate the content distribution of the target component in the sample points corresponding to the mixed spectrum.

[0131] In optional implementations, spectral unmixing methods include multivariate curve-resolved alternating least squares or nonnegative matrix factorization.

[0132] In the optional implementation, training module 804 is specifically used for:

[0133] A teacher model is trained based on point scan spectra and target spectral features extracted from point scan spectra.

[0134] Using area scanning spectrum as input and the output of the teacher model as the target to be imitated, the student model is trained; the student model is the feature prediction model.

[0135] In an optional implementation, the area scanning in the acquisition module 801 is achieved by a high spatial resolution imaging device, which includes at least one of a hyperspectral imaging device, a multispectral imaging device, or a short video imaging device.

[0136] In an optional implementation, multiple light sources are used for illumination when performing surface scanning with a high spatial resolution imaging device.

[0137] In an optional implementation, the point measurement in the acquisition module 801 is achieved by a high-precision spectroscopic device, which includes at least one of a laser-induced breakdown spectroscopy device, a Raman spectroscopy device, a visible / near-infrared spectroscopy device, an X-ray fluorescence spectroscopy device, or an inductively coupled plasma mass spectrometry device.

[0138] It should be noted that the various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for the system embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method embodiments. The system embodiments described above are merely illustrative, and the units described as separate components may or may not be physically separate. The components indicated as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the solution in this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0139] The above description is merely one specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A component imaging method based on point-plane spectral information fusion, characterized in that, The method comprises the following steps: acquiring a first image and a first spectral data cube of a sample, and acquiring point scanning spectra of a plurality of feature points of the sample and a second image containing the feature points; wherein the first image and the first spectral data cube are acquired by a surface scanning mode, and the point scanning spectra are acquired by a point measurement mode; determining a surface scanning spectrum corresponding to the point scanning spectrum from the first spectral cube based on the second image and the first image; extracting a target spectral feature from the point scanning spectrum; the target spectral feature is a spectral feature used to represent a target component; training a feature prediction model by taking the surface scanning spectrum as an input variable and taking the target spectral feature extracted from the point scanning spectrum corresponding to the surface scanning spectrum as a target variable; acquiring a second spectral data cube of a sample to be measured, predicting a prediction spectral feature of each pixel in the second spectral data cube by using the feature prediction model, and generating a component distribution image of the sample to be measured based on the prediction spectral features of all pixels in the second spectral data cube.

2. The method of claim 1, wherein, The method comprises the following steps: determining the spatial coordinates of the feature points corresponding to each point scanning spectrum based on the second image; spatially registering the second image and the first image to obtain a coordinate mapping relationship; determining the surface scanning spectrum corresponding to the point scanning spectrum from the first spectral data cube based on the spatial coordinates of the feature points corresponding to each point scanning spectrum according to the coordinate mapping relationship.

3. The method of claim 1, wherein, The target spectral feature comprises a feature peak intensity, a peak area, a spectral line ratio of the point scanning spectrum, or a feature vector obtained after preprocessing and feature extraction of the point scanning spectrum.

4. The method of claim 1, wherein, The method comprises the following steps: extracting a target spectral feature related to the target component from the point scanning spectrum based on prior knowledge; or, in the case that the point scanning spectrum is a mixed spectrum containing a target component and an interference component, extracting a target spectral feature related to the target component from the mixed spectrum by using a spectral unmixing method, and calculating the content distribution of the target component in the sample point corresponding to the mixed spectrum.

5. The method of claim 4, wherein, The spectral unmixing method comprises a multivariate curve resolution-alternating least squares method or a non-negative matrix factorization method.

6. The method of claim 1, wherein, The method comprises the following steps: training a teacher model based on the point scanning spectrum and the target spectral feature extracted from the point scanning spectrum; training a student model by taking the surface scanning spectrum as an input and taking the output of the teacher model as a mimic target; wherein the student model is the feature prediction model.

7. The method of claim 1, wherein, The surface scanning is realized by using a high spatial resolution imaging device; the high spatial resolution imaging device comprises at least one of a hyperspectral imaging device, a multispectral imaging device, or a short video imaging device.

8. The method of claim 7, wherein, When the surface scanning is performed by using the high spatial resolution imaging device, a multi-light source illumination is adopted.

9. The method of claim 1, wherein, The point measurement is implemented by a high-precision spectral device, and the high-precision spectral device includes at least one of a laser-induced breakdown spectroscopy device, a Raman spectroscopy device, a visible / near-infrared spectroscopy device, an X-ray fluorescence spectroscopy device or an inductively coupled plasma mass spectrometry device.

10. A component imaging system based on point-plane spectral information fusion, characterized in that, Comprise: An acquisition module is configured to acquire a first image and a first spectral data cube of a sample, and acquire point scanning spectra of a plurality of feature points of the sample and a second image containing the feature points; wherein the first image and the first spectral data cube are acquired by a surface scanning mode, and the point scanning spectra are acquired by a point measurement mode; A corresponding module is configured to determine a surface scanning spectrum corresponding to the point scanning spectrum from the first spectral cube based on the second image and the first image; An extraction module is configured to extract a target spectral feature from the point scanning spectrum; the target spectral feature is a spectral feature for characterizing a target component; A training module is configured to train a feature prediction model by taking the surface scanning spectrum as an input variable and taking the target spectral feature extracted from the point scanning spectrum corresponding to the surface scanning spectrum as a target variable; An imaging module is configured to acquire a second spectral data cube of a to-be-measured sample, predict a predicted spectral feature of each pixel in the second spectral data cube by using the feature prediction model, and generate a component distribution image of the to-be-measured sample based on the predicted spectral features of all pixels in the second spectral data cube.

Images