Cosmetic pearlescent effect optical detection system and method

Abstract

The present application relates to the technical field of optical detection, and in particular to a cosmetic pearl effect optical detection system and method, which adopts dynamic multi-spectrum modulation technology, virtual multi-angle fusion technology and nonlinear coupling optimization method to comprehensively obtain multi-dimensional optical data of a sample, and solves the problems of insufficient data source and low detection precision in the prior art. Through the combination of a multi-wavelength light source and virtual viewing angle technology, the present application effectively improves the detection range and accuracy of the pearl effect. After fusing multi-source data, key pearl features are extracted using multi-scale wavelet transform and sparse coding technology, and the data is analyzed in combination with a hybrid intelligent evaluation model to generate accurate pearl effect evaluation data. The present application improves the automation detection precision and efficiency of the cosmetic pearl effect, and has a wide application prospect.

Description

Technical Field

[0001] This invention relates to the field of optical inspection technology, and in particular to an optical inspection system and method for the pearlescent effect of cosmetics. Background Technology

[0002] The pearlescent effect in cosmetics is a crucial indicator of quality, especially in makeup products like foundation and eyeshadow, where it directly impacts the product's luster, longevity, and comfort. Therefore, accurately detecting the pearlescent effect is essential for ensuring product quality and enhancing user experience. Traditional methods for detecting pearlescent effects rely primarily on visual evaluation or simple physical measurements. However, these methods are highly subjective, difficult to implement, and struggle to quantify the microstructure of the pearlescent effect. To address these challenges, automated detection methods based on optical interference and scattering principles are becoming increasingly mainstream. These methods enable precise measurement of the sample's optical properties, achieving efficient detection of the pearlescent effect.

[0003] However, existing technologies primarily rely on single-wavelength light sources for detection and often depend solely on traditional phase unfolding techniques and conventional image processing algorithms. These methods are susceptible to environmental noise and equipment errors during optical measurements and cannot comprehensively acquire multi-angle, multi-wavelength information related to pearlescent effects, resulting in low measurement accuracy and inaccurate effect evaluation. Furthermore, the lack of fusion and optimization of multiple data sources during data processing often leads to information loss or local error accumulation, affecting the reliability and consistency of the final detection results. Summary of the Invention

[0004] To address the numerous problems existing in the prior art, this invention provides an optical detection system and method for the pearlescent effect of cosmetics. Based on dynamic multispectral modulation, virtual multi-angle fusion, and nonlinear optimization algorithms, this invention accurately captures the microstructural information of the pearlescent effect in cosmetics through multi-wavelength and multi-angle data fusion. Pearlescent features are extracted using multi-scale wavelet transform and sparse coding techniques, and intelligent analysis is performed using a hybrid intelligent evaluation model, ultimately generating high-precision pearlescent effect evaluation data and providing accurate and stable optical detection results.

[0005] An optical detection system for the pearlescent effect of cosmetics, the system comprising:

[0006] The acquisition module is used to preprocess cosmetic samples to form standardized sample data, and simultaneously acquire the reflection image and interference image of the sample, and perform noise reduction, background subtraction and phase unfolding processing on the image to generate the first global interference map data;

[0007] The iterative module is used to input the first global interferogram data into the improved Gerchberg-Saxton iterative algorithm, and generate intermediate iterative data by alternating between the Fourier domain and the spatial domain and applying physical constraints on the continuity of the sample surface and boundary conditions.

[0008] The fusion module is used to acquire multispectral data using dynamic multispectral modulation technology and generate virtual perspective data through virtual multi-angle fusion technology. The intermediate iterative data, multispectral data and virtual perspective data are combined using a nonlinear coupled multi-objective iterative optimization method. Local phase errors are corrected by a pre-trained residual convolutional neural network and the data is reduced in dimensionality by multi-scale joint sparse wavelet transform to generate second global interferogram data.

[0009] The evaluation module is used to perform multi-scale feature extraction and data dimensionality reduction on the second global interferogram data to generate pearlescent feature data. The pearlescent feature data reflects the period, phase gradient and three-dimensional scattering angle distribution of the interference fringes. The pearlescent feature data is then input into a hybrid intelligent evaluation model based on random forest and convolutional neural network to perform evaluation and generate pearlescent evaluation data.

[0010] Preferably, the acquisition module uses a CMOS interferometric camera with a resolution of no more than 2048×2048 pixels, and sequentially performs image denoising based on U-Net convolutional network, background subtraction using statistical methods, and improved Goldstein phase unrolling processing on the acquired reflection image and interferometric image to generate the first global interferogram data.

[0011] Preferably, the iterative module uses an improved Gerchberg-Saxton iterative algorithm to perform alternating transformations between the Fourier domain and the spatial domain, and applies physical constraints on the continuity of light intensity and sample thickness on the sample surface during the transformation process to generate intermediate iterative data.

[0012] Preferably, the dynamic multispectral modulation technology controls the laser to output a fixed wavelength sequence according to a predetermined wavelength switching sequence to acquire multispectral data.

[0013] Preferably, the virtual multi-angle fusion technology uses a geometric transformation algorithm in conjunction with a depth prediction network to generate virtual perspective data.

[0014] Preferably, the fusion module adopts a nonlinear coupled multi-objective iterative optimization method based on the alternating direction multiplier method, which combines intermediate iteration data, multispectral data and virtual perspective data, and uses a pre-trained residual convolutional neural network to correct local phase errors. Subsequently, the data is reduced in dimensionality through multi-scale joint sparse wavelet transform to generate the second global interferogram data.

[0015] Preferably, the evaluation module uses Daubechies 4 wavelet transform to perform multi-scale decomposition on the second global interferogram data, and combines sparse coding technology to achieve data dimensionality reduction, generating pearlescent feature data that reflects the period, phase gradient and three-dimensional scattering angle distribution of the interference fringes.

[0016] Preferably, the evaluation module uses a hybrid intelligent evaluation model including random forest and convolutional neural network, wherein random forest is used to determine the weight of each feature in the pearlescent feature data, and convolutional neural network is used to extract local pattern features and generate pearlescent evaluation data.

[0017] Preferably, the system uses standardized data transmission interfaces between its modules to achieve automatic transmission of detection data at each stage, and uploads the detection data to a cloud database to form a detection data chain, supporting continuous optimization across batches.

[0018] An optical detection method for the pearlescent effect of cosmetics, used to perform the optical detection system for the pearlescent effect of cosmetics, the method comprising:

[0019] After preprocessing, cosmetic samples are made into standardized sample data. Multiple CMOS interferometric cameras and single-frequency solid-state lasers are used to acquire reflection and interference images of the samples from a predetermined angle under synchronous control. The images are then processed sequentially to perform image denoising, background subtraction, and phase unwrapping to generate the first global interferogram data.

[0020] The first global interferogram data is input into the improved Gerchberg-Saxton iterative algorithm. During the alternating transformation between the Fourier domain and the spatial domain, physical constraints on the continuity of the sample surface and boundary conditions are applied to generate intermediate iterative data.

[0021] Multispectral data is acquired by switching at a predetermined wavelength using dynamic multispectral modulation technology, and virtual perspective data is generated by virtual multi-angle fusion technology. The intermediate iterative data, multispectral data and virtual perspective data are combined using a nonlinear coupled multi-objective iterative optimization method. Local phase errors are corrected by a pre-trained residual convolutional neural network and the data is dimensionality reduced by multi-scale joint sparse wavelet transform data to generate second global interferogram data.

[0022] Multi-scale feature extraction and dimensionality reduction are performed on the second global interferogram data to generate pearlescent feature data reflecting the period, phase gradient, and three-dimensional scattering angle distribution of the interference fringes. The pearlescent feature data is then input into a hybrid intelligent evaluation model based on random forest and convolutional neural network for evaluation to generate pearlescent evaluation data.

[0023] Compared with the prior art, the advantages and beneficial effects of the present invention are as follows:

[0024] This invention achieves high-precision optical detection of the pearlescent effect of cosmetics by introducing dynamic multispectral modulation technology, virtual multi-angle fusion technology, and a nonlinear coupled multi-objective iterative optimization method based on alternating direction multiplier method.

[0025] This invention solves the problem that a single wavelength cannot fully reflect the pearlescent effect in traditional technology by acquiring reflection and interference images of samples at multiple wavelengths; at the same time, it expands the detection angle range by using virtual multi-angle fusion technology, overcoming the problem of insufficient viewing angle caused by hardware limitations.

[0026] By combining multiple data sources, this invention effectively reduces the impact of noise and achieves high-precision data reconstruction during iterative optimization, ultimately obtaining more accurate and stable pearlescent feature data. Furthermore, the use of a hybrid intelligent evaluation model for intelligent analysis of the feature data significantly improves the system's detection accuracy and efficiency. Attached Figure Description

[0027] Figure 1 This is a structural block diagram of the system of the present invention;

[0028] Figure 2 This is a schematic diagram of the virtual reference channel structure in this invention;

[0029] Figure 3 This is a schematic diagram of the data fusion and optimization process in this invention;

[0030] Figure 4 This is a flowchart illustrating the method of the present invention. Detailed Implementation

[0031] The embodiments of the present disclosure will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the disclosure. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of the present disclosure for ease of explanation. However, it will be apparent that one or more embodiments may be practiced without these specific details. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concepts of the present disclosure.

[0032] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit this disclosure. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0033] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0034] like Figure 1 As shown, an optical detection system for the pearlescent effect of cosmetics includes:

[0035] The acquisition module is used to preprocess cosmetic samples to form standardized sample data, and simultaneously acquire the reflection image and interference image of the sample, and perform noise reduction, background subtraction and phase unfolding processing on the image to generate the first global interference map data;

[0036] The acquisition module in this system is responsible for the primary data acquisition task in the optical detection of the pearlescent effect of cosmetics. Its overall design goal is to standardize the cosmetic samples through preprocessing, and then simultaneously acquire the reflection and interference images of the samples under precise optical conditions. The acquired image data is then sequentially processed with noise suppression, background subtraction, and phase unfolding to ultimately generate the first global interferogram data. Specifically, before entering the acquisition module, the cosmetic samples undergo preprocessing, which includes uniformly coating the sample surface, controlling the coating thickness, and initially smoothing the microstructure of the sample surface to form standardized sample data. This process ensures high consistency of the optical properties of the sample in subsequent detection, providing a stable physical basis for data acquisition.

[0037] After sample pretreatment, the acquisition module uses multiple optical sensors to collect data from the sample. The sensors used here are primarily CMOS interferometers, whose main task is to capture the optical signals reflected by the sample at different angles. Through a pre-set synchronization control mechanism, the system can simultaneously acquire reflection and interference images of the sample from multiple predetermined angles. The reflection image reflects the overall optical reflection characteristics of the sample, while the interference image contains interference fringe information caused by the sample's internal microstructure; this information directly relates to the pearlescent effect of cosmetics.

[0038] The acquired raw image data often contains noise and background signals due to factors such as ambient light interference and equipment noise. Therefore, image denoising and background subtraction are essential processing steps. The specific processing flow includes preprocessing the raw image using image processing algorithms. First, noise suppression techniques are applied to filter out random noise in the image data. Then, statistical methods or techniques based on local mean calculation are used to subtract the image background to eliminate interference from non-sample signals. Next, addressing the issue of periodic packaging of phase information in the interferometric image, the system employs phase unwrapping techniques to process the image, restoring a continuous phase distribution and forming a first global interferogram data that accurately reflects the microscopic optical structure of the sample.

[0039] Overall, this acquisition module, through standardized preprocessing of cosmetic samples, synchronous multi-angle data acquisition, and a series of image processing procedures, ensures that the obtained first global interferogram data has a high signal-to-noise ratio and accuracy, providing reliable basic data for subsequent data processing modules. The design of this acquisition module not only meets high requirements in physical optical measurement but also fully utilizes modern image processing technology in the data processing stage, enabling the generated first global interferogram data to accurately reflect the optical characteristics of the sample, thereby accurately detecting the pearlescent effect. This design of the system provides a solid technical foundation for the objective and accurate detection of the pearlescent effect in cosmetics and plays a crucial role in subsequent data iteration, fusion, and intelligent evaluation, ensuring the efficiency and stability of the entire detection method.

[0040] Preferably, the acquisition module uses a CMOS interferometric camera with a resolution of no more than 2048×2048 pixels, and sequentially performs image denoising based on U-Net convolutional network, background subtraction using statistical methods, and improved Goldstein phase unrolling processing on the acquired reflection image and interferometric image to generate the first global interferogram data.

[0041] To further improve the performance and data quality of the acquisition module, the CMOS interferometer camera used in the acquisition module is required to have a spatial resolution strictly controlled within 2048×2048 pixels, thus ensuring image detail while also considering data transmission and processing speed. Specifically, after the pre-processed cosmetic sample enters the acquisition module, the CMOS interferometer camera first captures the sample's reflection image and the interference image caused by the sample's internal microstructure. The interference image directly reflects the physical basis of the pearlescent effect of the cosmetic, namely the phase interference phenomenon generated by the interaction between particles or thin film layers in the sample.

[0042] To effectively eliminate random noise and background interference introduced by the optical system, this invention employs an image denoising technique based on a U-Net convolutional neural network. The U-Net structure consists of an encoder and a decoder. During the encoding stage, it extracts multi-scale features of the image, and during the decoding stage, it gradually recovers high-resolution information through skip connections, ensuring that no crucial information is lost during denoising. In this embodiment, after the input reflection and interference images are processed by the U-Net network, the resulting image noise is significantly reduced, and image details are preserved, thus providing a cleaner data foundation for subsequent background subtraction.

[0043] The background subtraction process employs statistical methods. By analyzing the local statistical characteristics of each pixel region in the image, the background signal is automatically calculated and subtracted from the original image, thus retaining only the portion reflecting the true optical information of the sample. This background subtraction method is calculated based on the statistical mean or median of local pixel intensity, effectively reducing background deviations caused by ambient light and device errors, ensuring that the generated image data has a higher signal-to-noise ratio.

[0044] Phase unwrapping plays a crucial role in interferometric image data processing. Since the phase information in interferometric images typically exhibits a periodic packaging phenomenon, meaning the phase values ​​are confined to the range of -π to π, unwrapping is necessary to obtain a continuous phase distribution. This invention employs an improved Goldstein phase unwrapping method, which optimizes the conventional Goldstein algorithm to address noise interference. It utilizes local phase gradient calculation and iterative correction mechanisms to minimize phase jumps between adjacent pixels. The specific calculation expression is as follows:

[0045] in, Indicates the location The packaging phase at the location, An integer multiple factor ensuring phase continuity after unfolding is used, determined by a local phase gradient minimization criterion. The improved Goldstein method can stably recover continuous phase data even in the presence of noise, providing accurate phase information for generating the first global interferogram data reflecting the microstructure of the sample.

[0046] Overall, this invention significantly improves data quality while maintaining image detail and measurement accuracy by optimizing each processing step of the acquisition module. The use of a resolution-limited CMOS interferometric camera ensures image data standardization, while the U-Net-based denoising method, statistical background subtraction, and improved Goldstein phase expansion constitute an efficient data preprocessing workflow. This workflow not only improves image clarity and stability but also lays a reliable foundation for subsequent core detection steps such as data iteration, fusion, and intelligent evaluation. Through this preferred embodiment, the system can acquire high-quality first global interferogram data in actual detection, ensuring that subsequent processing modules can fully utilize this data, ultimately achieving objective and accurate detection of the pearlescent effect of cosmetics and providing strong support for continuous optimization of the detection data.

[0047] The iterative module is used to input the first global interferogram data into the improved Gerchberg-Saxton iterative algorithm, and generate intermediate iterative data by alternating between the Fourier domain and the spatial domain and applying physical constraints on the continuity of the sample surface and boundary conditions.

[0048] The iterative module in this system is the core component of the optical inspection process for the pearlescent effect in cosmetics. Its main task is to iteratively process the first global interferogram data obtained by the acquisition module to recover the true optical phase information and internal microstructural features of the sample. Specifically, the iterative module feeds the input first global interferogram data into an improved Gerchberg-Saxton iterative algorithm. This algorithm continuously alternates between the Fourier domain and the spatial domain to achieve the conversion from packaging phase data to continuous phase data. During this process, by applying physical constraints to the image, such as sample surface continuity and boundary conditions, it ensures that each iteration update can refine local structural details while maintaining global phase consistency.

[0049] The iterative module is designed to eliminate phase packaging and noise interference introduced during image acquisition. Through iterative optimization, it generates high-quality intermediate iteration data, providing a solid foundation for subsequent data fusion, feature extraction, and intelligent evaluation. The core principle of this module is based on Fourier transform theory. Its application in image processing primarily involves converting spatial domain images to the frequency domain, analyzing spectral information to identify periodic features within the sample's internal structure, and then converting the frequency domain information back to the spatial domain to recover phase information.

[0050] The entire iterative process requires a balanced consideration of global and local information in each transformation, and the imposition of physical constraints on the transformation results to ensure that the generated intermediate iterative data accurately reflects the physical properties of the sample. This process relies not only on the convergence and stability of the iterative algorithm itself, but is also affected by factors such as input data quality, physical constraint parameters, and iteration step size. Overall, the iterative module, as a key technical component of this system, aims to utilize advanced iterative algorithms and physical constraint techniques to transform the original data, limited to the packaging phase range, into continuous, non-jumping phase data. Through continuous iterative optimization, it achieves the best data reconstruction effect, thereby providing accurate optical phase and structural information for the detection of pearlescent effects in cosmetics, ensuring a solid foundation and reliable data for subsequent data fusion and intelligent evaluation modules.

[0051] Preferably, the iterative module uses an improved Gerchberg-Saxton iterative algorithm to perform alternating transformations between the Fourier domain and the spatial domain, and applies physical constraints on the continuity of light intensity and sample thickness on the sample surface during the transformation process to generate intermediate iterative data.

[0052] The iterative module employs an improved Gerchberg-Saxton iterative algorithm, the core of which lies in repeatedly alternating between the Fourier domain and the spatial domain, while strictly applying physical constraints during each transformation. Specifically, the system first performs a Fourier transform on the input first global interferogram data, denoted as... ,in This represents the phase distribution after packaging. In the frequency domain, the algorithm corrects the amplitude information, filters out noise and high-frequency interference, and then returns to the spatial domain through an inverse Fourier transform. During the spatial domain transformation, to ensure data continuity in local regions, the system forces the local phase gradient to meet smoothness requirements through preset physical constraints, namely, the continuity constraint of light intensity on the sample surface and boundary condition restrictions. Here, smoothness can be achieved by minimizing the second-order difference of the local phase gradient, and the calculation expression can be written as:

[0053]

[0054] in, Indicates the location Phase value at that point, This represents the local phase smoothness loss, which is used to adjust the local phase change in each iteration. By... Minimizing the phase transition error (PTR) ensures higher phase continuity and stability in the generated intermediate iteration data. In practice, the algorithm sets a fixed iteration step size and a preset stopping criterion, such as terminating the iteration when the phase update error falls below a certain fixed threshold (e.g., 0.01 radians), thus guaranteeing algorithm convergence and obtaining optimal intermediate iteration data. In the preferred embodiment, the improved algorithm also introduces an adaptive correction mechanism to address the abrupt transitions in the packaged phase data. This involves detecting local phase abrupt change points after each iteration and adjusting the corresponding phase compensation amount based on local gradient information to ensure a smooth transition of the overall phase.

[0055] Through this improvement, the system can effectively recover continuous phase and generate high-quality intermediate iterative data even in the presence of noise and local discontinuities. In a preferred embodiment, the algorithm is implemented on a modern high-performance computing platform, capable of completing a large number of iterative calculations in a short time, suitable for real-time detection needs, and its practicality and reliability in the optical detection of pearlescent effects in cosmetics have been verified through extensive experiments. Overall, the improved Gerchberg-Saxton iterative algorithm plays a key role in this system. During the alternating Fourier and spatial domain transformation, it achieves the generation of high-quality intermediate iterative data through precise physical constraints and adaptive correction mechanisms, thereby providing accurate and continuous phase information for subsequent data fusion and feature extraction, effectively improving the performance and data reconstruction accuracy of the entire detection system.

[0056] The iterative module pays particular attention to achieving the continuity constraint of light intensity on the sample surface, which plays a crucial role in ensuring that intermediate iteration data accurately reflects the surface characteristics of the sample. Specifically, the formation of the pearlescent effect in cosmetics is closely related to the interaction between particles on the sample surface, and this interaction typically manifests as local continuity of surface light intensity in optical detection. Therefore, this invention introduces a continuity constraint on light intensity during the alternating transformation between the Fourier and spatial domains to adjust the phase change between local pixels. This constraint is determined by calculating the difference in light intensity between adjacent pixels before and after the transformation, as expressed in the following formula:

[0057]

[0058] in, Indicates the location The light intensity value at that location, This represents the difference in light intensity between adjacent pixels at that location. A threshold is set... (For example, determined experimentally to be a fixed value within a certain range), if If a discontinuity in light intensity is detected, the system applies a smoothing correction at that location. This correction employs a local mean or weighted average algorithm to adjust the light intensity at the outlier point to the weighted average of the light intensity values ​​within its neighborhood, thereby maintaining the smoothness of the light intensity on the sample surface throughout the conversion process. In the examples, researchers conducted multiple tests on different cosmetic samples to determine the method suitable for most samples. The value is fixed in the algorithm to achieve automatic correction.

[0059] This method allows the system to monitor and adjust the local light intensity continuity in real time during the Fourier domain to spatial domain transformation. This ensures that the generated intermediate iterative data reflects sample details while avoiding errors introduced by abrupt changes in light intensity, thus providing high-quality data for subsequent multi-technology fusion. The application of this light intensity continuity constraint not only guarantees the overall smoothness of the data but also improves the stability and convergence speed of the phase unfolding algorithm in local regions, effectively solving the problem of local phase discontinuities that easily occur in the traditional Gerchberg-Saxton algorithm under high noise conditions. Overall, this implementation method, through precise control of the light intensity continuity on the sample surface, enables the detection system to more accurately capture key surface structure information in the pearlescent effect, thereby providing more convincing data support for cosmetic quality testing.

[0060] Furthermore, in this invention, to ensure that the intermediate iteration data can fully reflect the actual physical properties of the cosmetic sample, the iterative module introduces a sample thickness continuity constraint during the Fourier domain to spatial domain conversion process. The coating thickness of the cosmetic sample affects its optical interference phenomenon to a certain extent, thus affecting the formation of the pearlescent effect. Therefore, accurately recovering the coating thickness information is crucial to the performance of the overall detection system. To achieve this goal, this invention models and constrains the sample thickness continuity during the iteration process to ensure that the thickness change in the local area remains smooth. In specific implementation, the system first estimates the sample coating thickness using the phase information obtained from the preliminary reconstruction. The thickness calculation can be based on the following formula:

[0061]

[0062] in, Indicates the location The estimated sample thickness, The wavelength of the laser. This represents the local phase difference. To ensure the continuity of the estimated thickness, the system smooths the calculated thickness data using local weighted averaging or median filtering methods. Local thickness outliers are replaced with the weighted average of the thickness data within their neighborhood, thus eliminating errors caused by measurement noise and local structural changes. Through this thickness continuity constraint, the iterative module not only improves the accuracy of thickness information in intermediate iteration data but also enhances the sensitivity to thickness changes during phase unfolding, enabling the final reconstructed data to more realistically reflect the sample's microstructure.

[0063] In the embodiments, experimental data show that when thickness continuity constraints are adopted, the thickness error in each region of the sample is significantly reduced, and the smoothness and consistency of the phase data are significantly improved, thus providing higher quality input data for the multi-technology fusion module. Overall, by introducing thickness continuity constraints into the iterative module, this invention effectively solves the reconstruction instability problem caused by local thickness abrupt changes in traditional algorithms, ensuring that the generated intermediate iteration data has high continuity and accuracy globally, thereby further improving the reliability and repeatability of cosmetic pearlescent effect detection.

[0064] The fusion module is used to acquire multispectral data using dynamic multispectral modulation technology and generate virtual perspective data through virtual multi-angle fusion technology. The intermediate iterative data, multispectral data and virtual perspective data are combined using a nonlinear coupled multi-objective iterative optimization method. Local phase errors are corrected by a pre-trained residual convolutional neural network and the data is reduced in dimensionality by multi-scale joint sparse wavelet transform to generate second global interferogram data.

[0065] like Figure 3 As shown, this fusion module plays a crucial role in the optical inspection system for the pearlescent effect of cosmetics. Its overall design goal is to effectively fuse intermediate iterative data with other auxiliary data obtained through the data acquisition module during the optical inspection process, so as to accurately reconstruct the second global interferogram data of the sample.

[0066] The basic workflow of this module is as follows: First, dynamic multispectral modulation technology is used to acquire optical response data of cosmetic samples at different wavelengths, thereby obtaining multispectral data. Simultaneously, virtual multi-angle fusion technology is used to generate virtual perspective data that complements the actual acquisition angle, supplementing the insufficient perspective caused by limitations in the actual hardware acquisition angle. Next, intermediate iterative data, multispectral data, and virtual perspective data are combined using a nonlinear coupled multi-objective iterative optimization method. Local phase errors are corrected using a pre-trained residual convolutional neural network, and then multi-scale joint sparse wavelet transform is used to achieve data dimensionality reduction, ultimately generating the second global interferogram data. Overall, this module, through deep fusion and iterative optimization of multi-source data, ensures that the reconstructed data not only reflects the optical characteristics of the cosmetic sample globally but also presents a continuous, non-abrupt phase distribution in local details, providing reliable and accurate basic data for subsequent feature extraction and intelligent evaluation.

[0067] The system's fusion scheme considers the differences in response data across wavelengths in optical measurements, expands the coverage of acquisition angles using virtual perspective technology, and achieves collaborative optimization between data through nonlinear coupling algorithms. This overcomes the limitations of incomplete information from a single data source, thus realizing a data fusion effect of "1+1>2" under multi-dimensional data interaction. The overall module design fully leverages the advantages of modern optical detection technology and, through advanced algorithm models and deep learning methods, achieves efficient conversion from multispectral and virtual perspective data to the final reconstructed data. It also effectively reduces noise and errors in data processing, improving the accuracy and stability of the entire detection system.

[0068] Preferably, the dynamic multispectral modulation technology controls the laser to output a fixed wavelength sequence according to a predetermined wavelength switching sequence to acquire multispectral data.

[0069] In a preferred embodiment, dynamic multispectral modulation technology precisely controls the laser to output a fixed wavelength sequence through a predetermined wavelength switching sequence, thereby achieving multispectral data acquisition of the sample. Specifically, this technology utilizes a pre-programmed wavelength switching scheme, causing the laser to output several preset wavelengths in a fixed order during each acquisition cycle, such as 450 nm, 500 nm, 550 nm, and 600 nm. These wavelengths cover the key bands related to color and brightness in the pearlescent effect of cosmetics.

[0070] The acquired data for each wavelength in the system can reflect the optical response of the sample in that specific band, thereby revealing the influence of internal particles, thin film structure, and coating characteristics on light scattering and interference phenomena. The advantage of using dynamic multispectral modulation technology is that it can complete wavelength switching in a very short time, and because the laser output wavelength is fixed, the multispectral data obtained during data acquisition has high repeatability and accuracy, thus providing a stable input signal for subsequent data fusion.

[0071] Experimental verification shows that this technology can significantly improve the resolution of spectral data, ensuring that the reflected and interferometric images acquired at each wavelength have a high signal-to-noise ratio, thus laying a solid foundation for the fusion of multispectral data. For example, when cosmetic samples have complex pearlescent effects, the optical responses at different wavelengths may differ significantly. A predetermined wavelength switching scheme can accurately capture these differences. By comparing the interferograms and their phase information at different wavelengths, the microstructure distribution and optical interface characteristics of the sample can be revealed. In summary, the application of dynamic multispectral modulation technology provides a systematic and highly operable data acquisition method for the entire detection system. Through precise control and rapid switching of the laser output wavelength, it ensures high consistency and repeatability of multispectral data, providing sufficient and accurate spectral information support for subsequent multi-data fusion processing.

[0072] Preferred, such as Figure 2 As shown, the virtual multi-angle fusion technology uses a geometric transformation algorithm in conjunction with a depth prediction network to generate virtual perspective data.

[0073] Virtual multi-angle fusion technology, as a crucial component of the fusion module, primarily aims to compensate for insufficient angular information caused by hardware and field-of-view limitations during actual data acquisition. This technology employs a geometric transformation algorithm combined with a depth prediction network to generate virtual viewpoint data. The specific operational process is as follows: First, based on the intermediate iteration data acquired by the acquisition module, theoretical views of the sample at different angles are calculated using known optical system parameters and acquisition angles through geometric projection. Subsequently, a pre-trained depth prediction network optimizes these theoretical views, correcting deviations caused by errors in the actual optical system, ultimately generating high-quality virtual viewpoint data. The core of virtual multi-angle fusion technology lies in its use of a deep learning model to correct the results of traditional geometric transformations, enabling the generated virtual viewpoint data to more realistically reflect the optical characteristics of the sample at angles not actually acquired. Specifically, the depth prediction network employs a multi-layer convolutional structure; the input data consists of the image after preliminary geometric transformation and its corresponding acquisition parameters, while the output is the corrected virtual viewpoint image.

[0074] In this way, the system can expand the field of view, so that the final fused data includes both the actually acquired data and optical information inferred from other angles by virtual perspective technology. Experimental results show that after applying virtual multi-angle fusion technology, the generated virtual perspective data can effectively improve the overall integrity and continuity of the data, providing richer input for subsequent nonlinear iterative optimization. The introduction of this technology greatly improves the problem of missing image information caused by the limited actual acquisition angles and plays a key role in the data fusion process, ensuring that the fusion module can fully utilize multi-angle data to achieve higher-quality reconstruction of the second global interferogram data.

[0075] Preferably, the fusion module adopts a nonlinear coupled multi-objective iterative optimization method based on the alternating direction multiplier method, which combines intermediate iteration data, multispectral data and virtual perspective data, and uses a pre-trained residual convolutional neural network to correct local phase errors. Subsequently, the data is reduced in dimensionality through multi-scale joint sparse wavelet transform to generate the second global interferogram data.

[0076] The fusion module further employs a nonlinear coupled multi-objective iterative optimization method based on the Alternating Direction Multiplier Method (ADMM). The main purpose of this method is to deeply integrate intermediate iteration data, multispectral data, and virtual viewpoint data, and in the process, achieve accurate correction of local phase errors. The specific operation procedure is as follows: First, the intermediate iteration data is denoted as... Multispectral data is denoted as (in (representing different wavelengths), virtual viewpoint data is denoted as (in (Representing the viewpoint parameters). Using the ADMM method, the objective function is designed as a weighted sum of errors among the data points, and the global optimum is iteratively solved by introducing Lagrange multiplier constraints. Mathematically, this can be expressed as minimizing the objective function:

[0077]

[0078] in, This represents the data consistency loss function. This represents the regularization term, used to ensure that the output data... The smoothness and continuity of the ADMM method are achieved through alternating updates. The system employs Lagrange multipliers to jointly minimize both global and local data errors. Next, to further correct local phase errors, a pre-trained residual convolutional neural network is used. This network takes the current iteration data as input and outputs the corrected local phase shift. By weightedly feeding the correction results back into the ADMM iteration process, fine-tuning of the overall data is achieved.

[0079] In this process, the system sets a fixed iteration step size and convergence criterion. For example, iteration stops when the phase change in each iteration is lower than a predetermined threshold (e.g., 0.01 radians). Experimental data shows that this invention effectively reduces local phase errors while maintaining the overall continuity of the data, significantly improving the reconstruction quality of the second global interferogram data. Through this optimization method, the system can collaboratively fuse multi-source data under multiple objective functions, fully leveraging the complementary advantages between data to form a reconstruction result with higher accuracy and lower noise levels, thus laying a solid foundation for subsequent data dimensionality reduction and feature extraction.

[0080] The evaluation module is used to perform multi-scale feature extraction and data dimensionality reduction on the second global interferogram data to generate pearlescent feature data. The pearlescent feature data reflects the period, phase gradient and three-dimensional scattering angle distribution of the interference fringes. The pearlescent feature data is then input into a hybrid intelligent evaluation model based on random forest and convolutional neural network to perform evaluation and generate pearlescent evaluation data.

[0081] This evaluation module aims to perform multi-scale feature extraction and dimensionality reduction on the second global interferogram data obtained after preliminary processing, in order to generate pearlescent feature data that accurately reflects the pearlescent effect of cosmetic samples. This module comprehensively achieves in-depth mining of the microstructural information contained in complex optical interferograms. Its core objective is to use mathematical transformations and data compression methods to convert high-dimensional optical data into low-dimensional but information-dense feature descriptions, thereby making the subsequent evaluation of the pearlescent effect using intelligent evaluation models more efficient and accurate.

[0082] Specifically, this module first performs multi-scale decomposition on the second global interferogram data, separating the original data according to different frequency levels to extract key optical features such as interference fringes, phase gradients, and three-dimensional scattering angles. Next, data dimensionality reduction techniques are used to sparsely encode and compress the decomposition results to eliminate redundant information while maintaining sensitivity to the sample's microstructure. The entire process not only leverages the advantages of wavelet transform in time-frequency analysis but also significantly reduces the data volume of the output pearlescent feature data while preserving key information through a meticulously designed dimensionality reduction algorithm, thus providing high-quality input data for subsequent machine learning-based intelligent evaluation. Overall, this module achieves efficient conversion from high-dimensional complex data to low-dimensional feature representation, ensuring that the generated pearlescent feature data fully reflects the key parameters of the sample in optical detection. This provides data support for the quantitative detection of the pearlescent effect in cosmetics, demonstrating high practical value and promising application prospects.

[0083] Preferably, the evaluation module uses Daubechies 4 wavelet transform to perform multi-scale decomposition on the second global interferogram data, and combines sparse coding technology to achieve data dimensionality reduction, generating pearlescent feature data that reflects the period, phase gradient and three-dimensional scattering angle distribution of the interference fringes.

[0084] This evaluation module employs the Daubechies 4 wavelet transform to perform multi-scale decomposition of the second global interferogram data. This method, based on wavelet theory, uses the Daubechies 4 wavelet basis to perform time-frequency domain decomposition of the original data, allowing for the simultaneous preservation of details and overall trends at different scales. In specific implementation, the system first processes the input data... Performing a discrete wavelet transform yields a set of wavelet coefficients. ,in Indicates the scale factor. To represent spatial displacement, use the formula:

[0085]

[0086] in, This represents the Daubechies 4 wavelet basis function, through which data features at different scales are extracted layer by layer. After wavelet decomposition, the system further processes the wavelet coefficients using sparse coding techniques. Here, an L1-regularized least squares optimization method is used to sparsify the coefficient matrix, with the goal of solving the following optimization problem:

[0087]

[0088] in, This indicates data to be reduced in dimensionality. Describes the wavelet basis matrix. This represents the low-dimensional representation obtained after dimensionality reduction. This is a regularization parameter used to balance data reconstruction errors and sparsity requirements. Through this process, the pearlescent feature data generated by the system not only retains key information such as interference fringe period, phase gradient, and three-dimensional scattering angle distribution, but also significantly reduces data dimensionality, lowers computational load, and improves the input data quality of subsequent intelligent evaluation models.

[0089] In the experiment, interferogram data from various cosmetic samples were tested. The results showed that after adopting this invention, the pearlescent feature data achieved a good balance between data compression rate and information fidelity, effectively suppressing noise interference, and exhibiting good robustness and stability in multi-scale feature extraction. This technical solution can achieve real-time processing via GPU acceleration during implementation, ensuring high computational efficiency and response speed in large-scale detection tasks, providing strong support for the accurate detection of pearlescent effects in cosmetics.

[0090] Preferably, the evaluation module uses a hybrid intelligent evaluation model including random forest and convolutional neural network, wherein random forest is used to determine the weight of each feature in the pearlescent feature data, and convolutional neural network is used to extract local pattern features and generate pearlescent evaluation data.

[0091] The evaluation module employs a hybrid intelligent evaluation model based on a joint evaluation system composed of random forest and convolutional neural network. The model's design goal is to fully utilize the nonlinear modeling capabilities of random forest in processing high-dimensional feature data and the advantages of convolutional neural networks in extracting local pattern features, thereby comprehensively evaluating pearlescent feature data and generating pearlescent evaluation data with high discriminative power. Specifically, the system first inputs the pearlescent feature data obtained through multi-scale wavelet dimensionality reduction into the random forest model. The random forest consists of multiple decision trees, and during its training, the Bagging method is used to randomly extract samples and features from the pearlescent feature data, determining the initial evaluation score for each sample through majority decision. The role of the random forest model here is to rank the importance of each feature, thereby determining which features have a greater impact on the pearlescent effect. During this process, the model outputs a weight vector. Each component in this vector corresponds to a feature in the pearlescent feature data.

[0092] Next, the system inputs the pearlescent feature data along with the weight vector output by the random forest into a convolutional neural network. The convolutional neural network employs a multi-layer convolutional structure, extracting local pattern features from the data through local convolution operations, thereby capturing subtle variations in the pearlescent feature data. The specific network structure may include several convolutional layers, activation layers, and pooling layers, and finally, a fully connected layer maps the extracted features onto the evaluation score. The loss function used in the model is typically the mean squared error function to ensure high accuracy in regression prediction of the evaluation score during training.

[0093] By combining random forest and convolutional neural network, the system fully leverages the advantages of both, preserving global statistical information while extracting local detailed features to generate pearlescent evaluation data. Experimental data shows that this hybrid model accurately reflects the actual optical properties of different cosmetic samples when evaluating their pearlescent effects. The output evaluation results demonstrate high stability and consistency across multiple sample tests, providing a scientific basis for cosmetic quality control. Overall, this hybrid intelligent evaluation model uses random forest to determine feature weights and convolutional neural network to extract local patterns, forming a comprehensive evaluation mechanism. This provides quantitative evaluation indicators for pearlescent effect detection and enables efficient and accurate automated evaluation in practical applications.

[0094] Preferably, the system uses standardized data transmission interfaces between its modules to achieve automatic transmission of detection data at each stage, and uploads the detection data to a cloud database to form a detection data chain, supporting continuous optimization across batches.

[0095] Data transmission and integration between modules utilize standardized data transmission interfaces to ensure the stability and efficiency of data exchange between processing modules. Specifically, the acquisition module, iteration module, fusion module, and evaluation module all transmit data through a unified data interface protocol, ensuring consistency in data format, timestamps, and precision parameters across modules. This interface protocol supports high-throughput data transmission and employs data verification and error correction mechanisms during data exchange to ensure the integrity and accuracy of transmitted data.

[0096] Furthermore, the system achieves centralized storage and cross-batch data integration by uploading detection data from each stage to a cloud database in real time, forming a complete detection data chain. This data chain not only records data from the entire process—from sample preprocessing, image acquisition, iteration, fusion to evaluation—but also provides rich data resources for subsequent continuous system optimization and model updates. Through big data analysis and trend prediction of cloud data, the system can automatically identify deviations occurring during the detection process and dynamically adjust algorithm parameters to achieve continuous optimization across batches.

[0097] Specifically, the cloud-based analysis module utilizes statistical methods and machine learning algorithms to analyze historical testing data, extracting features such as data distribution, error variations, and system drift. This data is then fed back to the local system to adjust the operating parameters of each module. This design not only improves data transmission and processing efficiency but also creates a self-regulating and continuously optimizing closed-loop mechanism during the testing process, ensuring the long-term stable operation of the cosmetic pearlescent effect optical testing system. Experimental results show that, after adopting a standardized data transmission interface and a cloud-based data integration scheme, the system exhibits high real-time performance and robustness when handling large-scale testing tasks. Data at each stage can be accurately synchronized, providing technical support for the efficient collaborative operation of the entire testing process, thereby significantly improving testing accuracy and overall performance.

[0098] like Figure 4 As shown, an optical detection method for the pearlescent effect of cosmetics is used to perform the optical detection system for the pearlescent effect of cosmetics. The method includes:

[0099] After preprocessing, cosmetic samples are made into standardized sample data. Multiple CMOS interferometric cameras and single-frequency solid-state lasers are used to acquire reflection and interference images of the samples from a predetermined angle under synchronous control. The images are then processed sequentially to perform image denoising, background subtraction, and phase unwrapping to generate the first global interferogram data.

[0100] The first global interferogram data is input into the improved Gerchberg-Saxton iterative algorithm. During the alternating transformation between the Fourier domain and the spatial domain, physical constraints on the continuity of the sample surface and boundary conditions are applied to generate intermediate iterative data.

[0101] Multispectral data is acquired by switching at a predetermined wavelength using dynamic multispectral modulation technology, and virtual perspective data is generated by virtual multi-angle fusion technology. The intermediate iterative data, multispectral data and virtual perspective data are combined using a nonlinear coupled multi-objective iterative optimization method. Local phase errors are corrected by a pre-trained residual convolutional neural network and the data is dimensionality reduced by multi-scale joint sparse wavelet transform data to generate second global interferogram data.

[0102] Multi-scale feature extraction and dimensionality reduction are performed on the second global interferogram data to generate pearlescent feature data reflecting the period, phase gradient, and three-dimensional scattering angle distribution of the interference fringes. The pearlescent feature data is then input into a hybrid intelligent evaluation model based on random forest and convolutional neural network for evaluation to generate pearlescent evaluation data.

[0103] This invention performs high-precision analysis of the pearlescent effect of cosmetic samples through a series of precise optical detection and data processing steps. First, the cosmetic samples are preprocessed to form standardized sample data. Then, multiple CMOS interferometer cameras and a single-frequency solid-state laser are used to acquire reflection and interference images of the samples from predetermined angles under synchronous control. Next, the acquired images are processed for image denoising, background subtraction, and phase unwrapping to generate the first global interferogram data.

[0104] Then, the generated first global interferogram data is input into an improved Gerchberg-Saxton iterative algorithm. This algorithm applies physical constraints on the sample surface continuity and boundary conditions through alternating transformations between the Fourier and spatial domains to generate intermediate iterative data. Next, multispectral data is acquired by switching at predetermined wavelengths using dynamic multispectral modulation technology, and virtual multi-angle fusion technology is used to generate virtual viewpoint data. These intermediate iterative data, multispectral data, and virtual viewpoint data are combined and further processed using a nonlinear coupled multi-objective iterative optimization method. Local phase errors are corrected via a pre-trained residual convolutional neural network, and data dimensionality reduction is performed through multi-scale joint sparse wavelet transform to generate the second global interferogram data.

[0105] Finally, multi-scale feature extraction and dimensionality reduction are performed on the second global interferogram data to generate pearlescent feature data reflecting the interference fringe period, phase gradient, and three-dimensional scattering angle distribution. This pearlescent feature data is then input into a hybrid intelligent evaluation model based on random forest and convolutional neural network for real-time evaluation, generating pearlescent evaluation data. This method provides strong technical support for the efficient and accurate detection of pearlescent effects in cosmetics.

[0106] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects.

[0107] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. An optical detection system for the pearlescent effect of cosmetics, characterized in that, The system includes: The acquisition module is used to preprocess cosmetic samples to form standardized sample data, and simultaneously acquire the reflection image and interference image of the sample, and perform noise reduction, background subtraction and phase unfolding processing on the image to generate the first global interference map data; The iterative module is used to input the first global interferogram data into the improved Gerchberg-Saxton iterative algorithm, and generate intermediate iterative data by alternately transforming between the Fourier domain and the spatial domain and applying physical constraints on the continuity of the sample surface and boundary conditions. The improved Gerchberg-Saxton iterative algorithm applies physical constraints on the continuity of light intensity and sample thickness on the sample surface during the transformation process. During the spatial domain transformation process, it forces the local phase gradient to meet the smoothness requirements through preset physical constraints, namely the continuity constraints of light intensity on the sample surface and boundary conditions. It sets a fixed iteration step size and a preset stopping criterion. When the phase update error is lower than a certain fixed threshold, the iteration is terminated. After each iteration, local phase abrupt change points are detected, and the corresponding phase compensation amount is adjusted according to the local gradient information. The fusion module is used to acquire multispectral data using dynamic multispectral modulation technology and generate virtual perspective data through virtual multi-angle fusion technology. The intermediate iterative data, multispectral data and virtual perspective data are combined using a nonlinear coupled multi-objective iterative optimization method. Local phase errors are corrected by a pre-trained residual convolutional neural network and the data is reduced in dimensionality by multi-scale joint sparse wavelet transform to generate second global interferogram data. The evaluation module is used to perform multi-scale feature extraction and data dimensionality reduction on the second global interferogram data to generate pearlescent feature data. The pearlescent feature data reflects the period, phase gradient and three-dimensional scattering angle distribution of the interference fringes. The pearlescent feature data is then input into a hybrid intelligent evaluation model based on random forest and convolutional neural network to perform evaluation and generate pearlescent evaluation data.

2. The system according to claim 1, characterized in that, The acquisition module uses a CMOS interferometric camera with a resolution of no more than 2048×2048 pixels. It sequentially performs image denoising based on a U-Net convolutional network, background subtraction using statistical methods, and improved Goldstein phase unrolling processing on the acquired reflection and interferometric images. The improved Goldstein phase unrolling processing optimizes the conventional Goldstein algorithm for noise interference by employing local phase gradient calculation and iterative correction mechanisms to minimize phase jumps between adjacent pixels. Furthermore, the integer multiple factor is determined by the local phase gradient minimization criterion to stably recover continuous phase data even in the presence of noise, thereby generating the first global interferogram data.

3. The system according to claim 1, characterized in that, The dynamic multispectral modulation technology controls the laser to output a fixed wavelength sequence according to a predetermined wavelength switching sequence, thereby acquiring multispectral data.

4. The system according to claim 1, characterized in that, The virtual multi-angle fusion technology uses a geometric transformation algorithm in conjunction with a depth prediction network to generate virtual perspective data.

5. The system according to claim 1, characterized in that, The fusion module employs a nonlinear coupled multi-objective iterative optimization method based on the alternating direction multiplier method, combining intermediate iterative data, multispectral data, and virtual perspective data. It also utilizes a pre-trained residual convolutional neural network to correct local phase errors, followed by multi-scale joint sparse wavelet transform to achieve data dimensionality reduction and generate the second global interferogram data.

6. The system according to claim 1, characterized in that, The evaluation module uses Daubechies 4 wavelet transform to decompose the second global interferogram data at multiple scales, and combines sparse coding technology to achieve data dimensionality reduction, generating pearlescent feature data that reflects the period, phase gradient and three-dimensional scattering angle distribution of the interference fringes.

7. The system according to claim 1, characterized in that, The evaluation module employs a hybrid intelligent evaluation model that includes random forest and convolutional neural network. Random forest is used to determine the weights of each feature in the pearlescent feature data, while convolutional neural network is used to extract local pattern features and generate pearlescent evaluation data.

8. The system according to claim 1, characterized in that, The system uses standardized data transmission interfaces between its modules to enable automatic transmission of detection data at each stage, and uploads the detection data to a cloud database to form a detection data chain, supporting continuous optimization across batches.

9. A method for optically detecting the pearlescent effect of cosmetics, used to execute the optical detection system for the pearlescent effect of cosmetics according to any one of claims 1 to 8, characterized in that, The method includes: After preprocessing, cosmetic samples are made into standardized sample data. Multiple CMOS interferometric cameras and single-frequency solid-state lasers are used to acquire reflection and interference images of the samples from a predetermined angle under synchronous control. The images are then processed sequentially to perform image denoising, background subtraction, and phase unwrapping to generate the first global interferogram data. The first global interferogram data is input into the improved Gerchberg-Saxton iterative algorithm. During the alternating transformation between the Fourier domain and the spatial domain, physical constraints on the continuity of the sample surface and boundary conditions are applied to generate intermediate iterative data. The improved Gerchberg-Saxton iterative algorithm applies physical constraints on the continuity of light intensity and sample thickness on the sample surface during the transformation. During the spatial domain transformation, the local phase gradient is forced to meet the smoothness requirements through preset physical constraints, namely, the continuity constraints of light intensity on the sample surface and the boundary conditions. A fixed iteration step size and a preset stopping criterion are set. The iteration is terminated when the phase update error is lower than a certain fixed threshold. After each iteration, local phase abrupt change points are detected, and the corresponding phase compensation amount is adjusted according to the local gradient information. Multispectral data is acquired by switching at a predetermined wavelength using dynamic multispectral modulation technology, and virtual perspective data is generated by virtual multi-angle fusion technology. The intermediate iterative data, multispectral data and virtual perspective data are combined using a nonlinear coupled multi-objective iterative optimization method. Local phase errors are corrected by a pre-trained residual convolutional neural network and the data is dimensionality reduced by multi-scale joint sparse wavelet transform data to generate second global interferogram data. Multi-scale feature extraction and dimensionality reduction are performed on the second global interferogram data to generate pearlescent feature data reflecting the period, phase gradient, and three-dimensional scattering angle distribution of the interference fringes. The pearlescent feature data is then input into a hybrid intelligent evaluation model based on random forest and convolutional neural network for evaluation to generate pearlescent evaluation data.

Images