Optical method for identifying a contamination state of a surface of a corneal shaping lens

Abstract

The present application relates to the field of corneal molding lens detection, especially to an optical identification method for surface pollution state of corneal molding lens. The method comprises: fixing the lens and controlling the time sequence of multi-angle light source to generate reflection and scattering light signals, synchronously collecting images and completing gray scale normalization processing, extracting high-light pollution areas based on dynamic gradient threshold, extracting scattering abnormal areas by using local gray scale difference, extracting edge texture change areas by using edge detection, then calculating the area proportion of each area, combining spatial random field to extract brightness fluctuation value, quantifying texture continuity by using gray level co-occurrence matrix, then splicing to generate pollution characteristic parameter vector, and finally inputting the vector into deep neural network model to output classification identification and cleaning result. The present application effectively avoids artificial subjective error by multi-modal optical collection and spatial mathematical feature quantization, and improves the noise resistance stability and automatic inference accuracy of lens pollution identification.

Description

Technical Field

[0001] This invention relates to the field of orthokeratology lens inspection, and more particularly to an optical identification method for the surface contamination status of orthokeratology lenses. Background Technology

[0002] Orthokeratology lenses are prone to accumulating tear protein and lipid plaques on their surface during daily wear, necessitating cleanliness testing to ensure eye health. Optical imaging technology is the core method for assessing lens contamination. It primarily involves illuminating the lens with external multi-angle light sources and using image sensors to collect the optical signals generated on the surface. Accurately acquiring the specular reflection component, which characterizes the lens substrate, and the diffuse scattering component, which characterizes microscopic contaminant particles, is fundamental for subsequent feature extraction and high-precision contamination identification.

[0003] However, existing detection schemes mostly use a single static light source and conventional asynchronous exposure mode, which makes it difficult to effectively separate bright reflected light from weak diffused light. This acquisition method is easily affected by stray light from the external environment, resulting in a low signal-to-noise ratio of dual-modal images and weakening the overall noise resistance stability. Summary of the Invention

[0004] To overcome the above shortcomings, this invention provides an optical identification method for the surface contamination status of orthokeratology lenses, aiming to improve the problems of existing solutions being unable to effectively separate reflected and scattered light and being easily interfered with by stray light.

[0005] This invention provides the following technical solution: an optical identification method for the surface contamination status of orthokeratology lenses, comprising:

[0006] S1. Fix the orthokeratology lens to the detection base, and sequentially control the multi-angle light sources distributed according to the preset incident angle to irradiate the lens surface in a time sequence, generating optical reflection and scattering light signals on the lens surface;

[0007] S2. In sync with the time-series illumination, the reflection and scattering images of the lens surface corresponding to the multi-angle light source are acquired respectively, and the pixel gray values ​​corresponding to each image are mapped to a preset gray range using a linear mapping function to complete the gray-level normalization processing of the image.

[0008] S3. From the normalized image, the bright pollution area is segmented based on the dynamic gradient threshold, the scattering abnormal area is extracted using the local gray-level difference method, and the edge texture change area is extracted using the edge detection operator.

[0009] S4. Calculate the percentage of pixel area of ​​the high-brightness contamination area, the scattering abnormal area, and the edge texture change area on the lens surface. Extract the brightness fluctuation value of the high-brightness contamination area based on the spatial random field model, and use the gray-level co-occurrence matrix to quantify the texture continuity of the edge texture change area. Then, splice them together to generate a contamination feature parameter vector.

[0010] S5. Input the pollution feature parameter vector into a preset deep neural network model, output the corresponding pollution level classification identifier, and generate a clean status result based on the classification identifier.

[0011] Preferably, in step S1, the step of sequentially controlling multiple angle light sources distributed at preset incident angles to irradiate the lens surface in a time sequence includes:

[0012] Based on a preset multi-dimensional illumination array, the independent spectral frequency bands and spatial incident angle parameters corresponding to each emission node in the multi-angle light source are retrieved.

[0013] A logic controller generates timing trigger pulses to sequentially illuminate each light source node of the illumination array to output a structured light field;

[0014] By controlling the projection of each beam onto the surface of the lens, the specular reflection component of the lens substrate and the diffuse scattering component of the pollution particles are separated.

[0015] Preferably, in step S2, the step of acquiring the reflection image and scattering image of the lens surface corresponding to the multi-angle light source respectively includes:

[0016] Configure an external synchronization trigger pulse and connect it to the synchronization port of the image sensor to perform physical phase-locked control on light source illumination and shutter exposure;

[0017] During the triggering cycle of the reflected incident angle light source, the reflection acquisition channel is activated to capture the reflection image of the lens surface containing macroscopic morphological features;

[0018] During the scattering incident angle light source triggering cycle, switch to the scattering acquisition channel and adjust the photosensitive gain to capture the scattering image containing fine particle characteristics.

[0019] Preferably, in step S2, the step of mapping the pixel grayscale values ​​of each image to a preset grayscale range using a linear mapping function includes:

[0020] The pixel intensity distribution of the reflected and scattered images on the lens surface is statistically analyzed, abnormal noise points are removed, and effective gray-level extreme values ​​are extracted.

[0021] Construct a first-order affine transformation model based on the effective gray-level extreme values, and calculate the mapping parameters to the preset gray-level range;

[0022] The mapping parameters are used to perform a linear scaling operation on all pixels to complete pixel recombination under a preset standard.

[0023] Preferably, in step S3, the step of segmenting the bright contamination area based on the dynamic gradient threshold includes:

[0024] The normalized image is spatially convolved using a discrete differential operator to generate a gradient magnitude matrix representing the pixel mutation rate.

[0025] The gradient magnitude matrix is ​​iteratively optimized based on the inter-class variance maximization criterion to obtain a dynamic segmentation threshold adapted to local illumination.

[0026] Based on the dynamic segmentation threshold, the image is subjected to binarization truncation and morphological closing operation to extract the connected and closed bright contamination region.

[0027] Preferably, in step S3, the step of extracting the scattering anomaly region using the local gray-level difference method includes:

[0028] A low-pass smoothing filter is applied to the normalized image to construct a low-frequency background reference image characterizing the lens's underlying optical features;

[0029] The normalized image is compared with the low-frequency background reference image to generate a residual grayscale image that highlights the diffuse reflection features of tiny particles.

[0030] A local grayscale tolerance band is set, and pixel clusters with amplitude exceeding the limit and spatially clustered in the residual grayscale image are extracted to define the boundary of the scattering anomaly region.

[0031] Preferably, in step S4, the step of extracting the brightness fluctuation value within the high-brightness pollution area based on the spatial random field model includes:

[0032] A Markov random field network is constructed in the high-brightness pollution area, and the topological association structure of pixel nodes and their neighborhood system is defined.

[0033] A potential function characterizing the optical dependence between nodes is constructed, and the autocorrelation of local gray levels is evaluated by calculating the neighborhood conditional probability distribution.

[0034] The second-order statistical moments of the potential function in its energy convergence state are extracted and their dimensionality reduced and mapped to the brightness fluctuation values ​​representing the unevenness of the pollution thickness.

[0035] Preferably, in step S4, the step of quantizing the texture continuity of the edge texture change region using a gray-level co-occurrence matrix includes:

[0036] Multi-scale spatial vector parameters are set in the edge texture change area to establish spatial relative position association rules for pixel pairs;

[0037] The joint probability of pixel pairs that satisfy the association rule is calculated, the gray-level co-occurrence matrix of the target dynamic range is constructed, and normalization processing is performed.

[0038] The texture continuity is output using the local stationarity and inverse difference moment parameter of the calculated gray-level co-occurrence matrix.

[0039] Preferably, in step S5, the step of inputting the pollution feature parameter vector into a preset deep neural network model includes:

[0040] A deep neural network architecture comprising an input layer, multiple fully connected hidden layers, and an output layer is constructed, and the pollution feature parameter vector is fed into the input layer.

[0041] The nonlinear activation function of the multilayer fully connected hidden layer is used to perform layer-by-layer feature abstraction and regularization processing.

[0042] The output layer uses a normalized exponential function to calculate the probability distribution of each pollution level, extracts the node corresponding to the maximum probability value, and triggers and outputs the corresponding pollution level classification identifier.

[0043] The present invention has the following beneficial effects:

[0044] 1. In this invention, a multi-dimensional illumination array is constructed and combined with a physical phase-locked loop control mechanism to achieve synchronization between the temporal light field and the sensor exposure, effectively separating the reflective component of the lens substrate and the diffuse scattering component of the contaminant particles, eliminating external stray light interference from the physical acquisition source, and improving the signal-to-noise ratio and overall noise resistance stability of the dual-modal image.

[0045] 2. In this invention, a spatial random field model and a gray-level co-occurrence matrix are used to perform depth quantization of the contaminated target area, transforming the fuzzy unevenness of contamination thickness and edge morphology into rigorous brightness fluctuation values ​​and texture continuity parameters. This breaks through the limitations of traditional detection methods that rely on subjective experience, and achieves a scientific mapping from scattered image pixels to structured features, thereby improving the objective accuracy of feature extraction.

[0046] 3. In this invention, the feature vector is fed into a deep neural network containing a regularization mechanism to fit the boundary of complex contamination distribution and output a rigorous probability distribution label. While effectively avoiding human judgment errors, it realizes the automatic classification of lens cleanliness status, providing reliable intelligent decision support for users' safe wearing and scientific cleaning. Attached Figure Description

[0047] Figure 1 This is a flowchart of an optical identification method for the surface contamination status of an orthokeratology lens proposed in this invention;

[0048] Figure 2 This is a diagram illustrating the temporal optical field and optical signal generation mechanism proposed in this invention;

[0049] Figure 3 This is a flowchart of the dual-modal image synchronization phase-locking and normalization processing proposed in this invention;

[0050] Figure 4 This is a flowchart of the multi-dimensional contamination target area physical stripping and feature segmentation proposed in this invention;

[0051] Figure 5 This is the feature quantization map based on spatial random fields and co-occurrence matrices proposed in this invention;

[0052] Figure 6 This is the state inference and logic decision diagram based on the DNN architecture proposed in this invention. Detailed Implementation

[0053] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] In embodiments of the present invention, the present invention provides an optical identification method for the surface contamination status of orthokeratology lenses, such as... Figure 1 As shown, it includes:

[0055] S1. Fix the orthokeratology lens to the detection base, and sequentially control the multi-angle light sources distributed according to the preset incident angle to irradiate the lens surface in a time sequence, generating optical reflection and scattered light signals on the lens surface;

[0056] Further, in step S1, the step of sequentially controlling multiple angle light sources distributed according to a preset incident angle to irradiate the lens surface in a time sequence includes:

[0057] Based on a preset multi-dimensional illumination array, the independent spectral frequency bands and spatial incident angle parameters corresponding to each emission node in the multi-angle light source are retrieved;

[0058] A logic controller generates timing trigger pulses to sequentially illuminate each light source node of the illumination array to output a structured light field.

[0059] By controlling the projection of each beam onto the lens surface, the specular reflection component of the lens substrate and the diffuse scattering component of the pollution particles are separated.

[0060] Specifically, the orthokeratology lens to be tested is first placed on a testing base with high-precision curved surface fitting features, ensuring that the concave or convex surface of the lens is physically fitted to the base, thus maintaining a static reference position for the lens during subsequent optical scanning. A multi-dimensional illumination array is arranged around the periphery of the testing base. After the system starts, the central control unit retrieves the independent spectral bands and spatial incident angle parameters corresponding to each emission node within the multi-angle light source from a preset hardware configuration file. The independent spectral bands here are selected from specific narrow-band wavelengths that can excite protein and lipid contaminants, and the spatial incident angle parameters are strictly distinguished into a small bright-field angle for acquiring the specular spot and a large dark-field angle for highlighting tiny particles.

[0061] Subsequently, a high-frequency timing trigger pulse is generated using a hardware logic controller. This timing pulse signal is characterized as a periodic square wave sequence in the time domain, driving the light source nodes of the illumination array to light up alternately in sequence, thereby outputting a structured light field on the lens surface whose spatial distribution and spectral characteristics alternate with time. The specific timing trigger pulse signal is represented as follows:

[0062] ;

[0063] In the above formula, Indicates time The timing trigger pulse signal of the output is controlled at all times. This represents the total number of light source nodes in a multi-dimensional illumination array. Indicates the first The driving current amplitude of each light source node, This indicates the trigger cycle duration for switching between adjacent light sources. Indicates the first Each light source node maintains the exposure pulse width for emitting light. This represents a standard rectangular pulse function. This pulse controls the sequential projection of light beams with different spatial incident angles onto the surface of the orthokeratology lens.

[0064] After the light beam reaches the lens surface, it undergoes complex optical interactions with the lens substrate material and contaminant particles such as tear proteins and lipids attached to it. The system separates the specular reflection component of the lens substrate from the diffuse scattering component of the contaminant particles using a predefined physical optical model. The total reflected light intensity of the orthokeratology lens surface under a specific spatial incident angle is expressed as:

[0065] ;

[0066] In the above optical separation formula, Represents the two-dimensional coordinates of the lens surface and The total reflected light intensity received at that point, This represents the initial incident light intensity of the currently lit light source node. This indicates the preset incident angle of the multi-angle light source space. This indicates the reflection reception angle of a fixed image sensor. The part before the plus sign represents the specular reflection component of the lens substrate. This indicates the Fresnel reflectance coefficient of the high-permeability substrate material for orthokeratology lenses. This indicates the surface roughness parameter of the clean lens substrate. The part after the plus sign represents the diffuse scattering component of the contaminant particles. Represents the two-dimensional coordinates of the lens surface and The albedo at a given location is positively correlated with the thickness of pollutant deposition and the particle density at that location.

[0067] Through the above steps, the strong reflection signal characterizing the lens background material and the weak scattering signal characterizing micro-pollutants were accurately separated, providing a physical basis for obtaining high signal-to-noise ratio optical image data.

[0068] S2. In sync with the time-series illumination, acquire the reflection and scattering images of the lens surface corresponding to the multi-angle light source, and use the linear mapping function to map the pixel gray values ​​of each image to the preset gray range to complete the gray normalization processing of the image.

[0069] Further, step S2, which involves acquiring the reflection and scattering images of the lens surface corresponding to the multi-angle light sources, includes:

[0070] Configure an external synchronization trigger pulse and connect it to the synchronization port of the image sensor to perform physical phase-locked control on light source illumination and shutter exposure;

[0071] Within the triggering cycle of the reflected incident angle light source, the reflection acquisition channel is activated to capture the reflection image of the lens surface containing macroscopic morphological features;

[0072] During the scattering incident angle light source trigger cycle, switch to the scattering acquisition channel and adjust the photosensitive gain to capture scattering images containing fine particle characteristics.

[0073] Further, in step S2, the step of mapping the pixel grayscale values ​​of each image to a preset grayscale range using a linear mapping function includes:

[0074] The pixel intensity distribution of the reflected and scattered images on the lens surface is statistically analyzed, abnormal noise points are removed, and effective gray-level extreme values ​​are extracted.

[0075] Construct a first-order affine transformation model based on the effective gray-level extreme values, and calculate the mapping parameters to the preset gray-level range;

[0076] By using mapping parameters to perform a linear scaling operation on all pixels, pixel recombination under a preset standard is completed.

[0077] Specifically, an external synchronous trigger pulse is configured and directly connected to the hardware synchronization port of the high-resolution image sensor to establish a physical phase-locked loop control mechanism that strictly aligns the periodic illumination action of the multi-dimensional illumination array with the global exposure time of the image sensor shutter. When the hardware receives a high-level pulse within the trigger cycle of the reflected incident angle light source, the central processing unit activates the reflection acquisition channel and uses the system's basic photosensitive gain value to capture image data of the bright areas on the lens surface, generating a lens surface reflection image that includes the overall morphological structure of the orthokeratology lens and macroscopic morphological features such as the edges of large-area liquid films. When the hardware timing progresses to the trigger cycle of the scattered incident angle light source, the central processing unit automatically switches to the scattering acquisition channel and synchronously increases the analog photosensitive gain of the image sensor, extending the physical exposure time to capture weak dark-field scattered photons, generating a scattering image that includes fine particle characteristics such as micron-level protein deposits and lipid plaques.

[0078] After acquiring the image data, the spatial distribution frequency of the intensity of all pixels in the entire image is statistically analyzed for the two-dimensional pixel matrices of the reflected and scattered images on the lens surface. To eliminate extreme interference caused by sensor thermal noise and ambient stray light, upper and lower threshold cutoffs are set based on the cumulative probability distribution of pixel intensity. Abnormal noise pixels distributed in the statistical edge interval are stripped out, and the true effective minimum and maximum grayscale values ​​of the current physical channel image are extracted.

[0079] Subsequently, a first-order affine transformation model based on the extracted effective gray-level extrema is constructed, and mapping parameters that map the original pixel intensity to a preset gray-level range are calculated. The system then performs a linear scaling operation on all pixels using these mapping parameters. The corresponding first-order linear mapping operator is expressed as follows:

[0080] ;

[0081] In the above formula for first-order affine transformation, Represents the coordinates in two-dimensional space after normalization. and The output of the target grayscale value of the lens. This represents the original pixel grayscale value of the orthokeratology lens surface at the corresponding coordinate point acquired by the image sensor. This represents the minimum effective grayscale value of the current channel image of the lens after removing abnormal noise. This represents the maximum effective grayscale value of the current channel image of the lens after removing abnormal noise. This represents the lower bound reference value of the preset grayscale range set according to the bit depth precision of the digital image processing system. This represents the upper bound of the preset grayscale range. The system uses this function matrix to perform point-by-point translation and scaling on the two-dimensional pixels corresponding to each modality image, completing pixel reconstruction under the preset standard.

[0082] The above steps effectively eliminate data distribution drift caused by external stray light interference and differences in hardware response of different photosensitive channels, and unify the numerical dimensions of reflected and scattered images.

[0083] S3. From the normalized image, the bright pollution area is segmented based on the dynamic gradient threshold, the scattering abnormal area is extracted by the local gray-level difference method, and the edge texture change area is extracted by the edge detection operator.

[0084] Furthermore, in step S3, the step of segmenting the bright contamination region based on the dynamic gradient threshold includes:

[0085] The normalized image is spatially convolved using discrete differential operators to generate a gradient magnitude matrix representing the pixel mutation rate.

[0086] The gradient magnitude matrix is ​​iteratively optimized based on the inter-class variance maximization criterion to obtain a dynamic segmentation threshold that adapts to local illumination.

[0087] Binarization truncation and morphological closing operations are performed on the image based on a dynamic segmentation threshold to extract connected and closed bright contamination regions.

[0088] Furthermore, step S3, the step of extracting the scattering anomaly region using the local gray-level difference method, includes:

[0089] A low-pass smoothing filter is applied to the normalized image to construct a low-frequency background reference image that characterizes the lens's underlying optical features;

[0090] The normalized image is compared with the low-frequency background reference image to generate a residual grayscale image that highlights the diffuse reflection features of tiny particles.

[0091] A local grayscale tolerance band is set, and pixel clusters with amplitude exceeding the limit and spatially concentrated in the residual grayscale image are extracted to define the boundary of the scattering anomaly region.

[0092] Specifically, after acquiring the normalized bimodal image of the orthokeratology lens, the system uses a direction-sensitive two-dimensional discrete differential operator to perform spatial convolution on the image's two-dimensional matrix. By calculating the spatial partial derivatives in the horizontal and vertical directions respectively, a gradient magnitude matrix representing the abrupt change rate of pixel grayscale on the lens surface is generated. The corresponding spatial gradient calculation formula is as follows:

[0093] ;

[0094] In the gradient calculation formula above, This represents the pixel gradient magnitude at a two-dimensional coordinate point on the lens surface. This represents the normalized corneal reshaping lens image data input to the system. and Let represent the discrete differential operator convolution kernels in the horizontal and vertical directions, respectively. This represents the two-dimensional spatial convolution operator. After generating the gradient magnitude matrix, the system iteratively optimizes the matrix based on the criterion of maximizing inter-class variance. By traversing all possible gray levels, it seeks the optimal threshold that maximizes the variance between large-area contaminated patches and the clean lens background area. The iterative objective function is expressed as:

[0095] ;

[0096] In the above optimization formula, This represents the dynamic segmentation threshold calculated to adapt to local illumination on the lens surface. This represents the candidate grayscale threshold for iterative traversal. and These represent the area probabilities occupied by bright lipid-contaminated pixels and clean background pixels within the effective area of ​​the lens, respectively, under the current threshold division. and These represent the gradient mean values ​​of the two types of pixel regions, respectively. The system performs binarization truncation on the normalized image based on the solved dynamic segmentation threshold, and then uses a morphological closing operation of dilation followed by erosion to fill in the tiny digital holes inside the connected regions of the pixels, finally extracting the bright contamination areas representing large areas of tear lipid membrane with connected closure.

[0097] To extract micron-sized particulate contaminants from the lens surface, the system performs a two-dimensional low-pass smoothing filter on the normalized image. A Gaussian smoothing kernel is used to perform spatial weighted averaging on neighboring pixels, filtering out local high-frequency particle scattering details and constructing a low-frequency background reference image characterizing the smooth material of the orthokeratology lens substrate. Then, pixel-level matrix difference operations are performed between the normalized image and this low-frequency background reference image; the residual calculation formula is as follows:

[0098] ;

[0099] In the above difference operation formula, This represents the feature pixel magnitude in the generated residual grayscale image. This represents the low-pass smoothing filter kernel matrix used to filter out high-frequency features on the lens surface. The residual grayscale image eliminates background interference caused by the lens substrate curvature, effectively highlighting the abnormal diffuse reflection characteristics caused by tiny protein particles attached to the lens surface. The system sets a local grayscale tolerance band based on the physical test threshold of a contaminated standard lens, traverses the residual grayscale image, and extracts pixel clusters whose pixel amplitude exceeds this tolerance band and meets the preset neighborhood connectivity density in two-dimensional spatial coordinates, thereby accurately defining the boundaries of isolated scattering anomaly regions.

[0100] After extracting the bright contamination area and the scattering anomaly area, the system simultaneously uses the edge detection operator to perform spatial abrupt change response calculation on the normalized image pixels, and extracts the edge texture change area on the lens surface caused by the boundary of scratches or stubborn attachments.

[0101] The above steps enable multi-dimensional physical stripping and precise segmentation of complex contamination target areas on the surface of orthokeratology lenses, providing a high-precision data foundation with strong noise resistance and stability for subsequent quantitative calculation of contamination parameters.

[0102] S4. Calculate the percentage of pixel area of ​​the bright contamination area, the scattering anomaly area and the edge texture change area on the lens surface. Extract the brightness fluctuation value in the bright contamination area based on the spatial random field model, and use the gray-level co-occurrence matrix to quantify the texture continuity of the edge texture change area. Then, splice them to generate a contamination feature parameter vector.

[0103] Further, step S4, the step of extracting the brightness fluctuation values ​​within the high-brightness contamination area based on the spatial random field model, includes:

[0104] In the high-brightness contamination area, a Markov random field network is constructed, and the topological association structure of pixel nodes and their neighborhood system is defined.

[0105] A potential function characterizing the optical dependence between nodes is constructed, and the autocorrelation of local gray levels is evaluated by calculating the neighborhood conditional probability distribution.

[0106] The second-order statistical moments of the potential function at the energy convergence state are extracted and their dimension reduction mapping is used to represent the brightness fluctuation values ​​that represent the unevenness of the pollution thickness.

[0107] Further, step S4, which involves quantizing the texture continuity of the edge texture variation region using the gray-level co-occurrence matrix, includes:

[0108] In the region of edge texture change, set multi-scale spatial vector parameters and establish spatial relative position association rules for pixel pairs;

[0109] Calculate the joint probability of pixel pairs that satisfy the association rule, construct the gray-level co-occurrence matrix of the target dynamic range, and perform normalization processing;

[0110] The local stationarity of the calculated gray-level co-occurrence matrix and the inverse difference moment parameter are used to output texture continuity.

[0111] Specifically, after identifying the high-brightness contamination area, the abnormal scattering area, and the edge texture change area, the system counts the total number of effective pixels contained within each physical region. The total number of pixels in each region is then divided by the preset total number of effective pixels on the orthokeratology lens surface to directly calculate the percentage of pixel area on the lens surface corresponding to the target contamination area.

[0112] To analyze the optical feature distribution within the high-brightness contamination region, a two-dimensional Markov random field network is constructed within this region. Each pixel within the high-brightness contamination region is defined as a random variable node in the network, and the topological association structure between the pixel node and its neighboring systems is established using the 8-connectivity rule. To evaluate the autocorrelation of local gray levels, a potential function characterizing the optical dependence between nodes is constructed, and the degree of gray level jump between adjacent pixels is quantified by calculating the neighborhood conditional probability distribution. The energy calculation formula for the potential function is expressed as:

[0113] ;

[0114] In the above formula for the potential function energy... Indicates the center pixel node within the bright contamination area of ​​the lens. The local neighborhood energy state This represents the spatial smoothing penalty coefficient set in the network model. Represents the center pixel node The set of eight-connected neighborhood systems and These represent the grayscale values ​​of tear lipid contamination at the corresponding two-dimensional spatial coordinates on the surface of the orthokeratology lens. After the global situation function energy reaches convergence, the system extracts the second-order statistical moments characterizing the spatial non-uniformity of distribution. Specifically, it calculates the variance of local energy within the high-brightness contamination region and maps it to brightness fluctuation values ​​representing the uneven thickness of lipid contamination on the lens surface.

[0115] For the edge texture variation region on the lens surface, the system sets a specific spatial distance step size and orientation angle as spatial vector parameters to establish spatial relative position association rules for pixel pairs within this region. The joint probability of pixel pairs at different gray levels is calculated under the set association rules, constructing a two-dimensional gray-level co-occurrence matrix for the target dynamic range. Probability sum normalization is then performed on all elements within this matrix. To output quantization metrics, the system calculates the local stationarity and inverse difference moment parameters of the normalized gray-level co-occurrence matrix. The texture continuity calculation formula is as follows:

[0116] ;

[0117] In the above texture continuity formula, The calculated output represents the texture continuity at the edges of physical scratches or stubborn protein deposits on the surface of the orthokeratology lens. This represents the number of dynamic gray levels of the target image. This indicates the set distance step size. and direction deflection The probability values ​​of the elements in the gray-level co-occurrence matrix after normalization are as follows. and These correspond to the row and column gray-level indices of the gray-level co-occurrence matrix, respectively. This inverse difference moment parameter assigns high weights to the diagonal elements, effectively filtering out non-physical texture breakpoints caused by high-frequency stray light.

[0118] Finally, the system concatenates the calculated pixel area ratio, brightness fluctuation value, and texture continuity of the three regions into a one-dimensional floating-point array according to the set dimensions, generating a pollution feature parameter vector with high-dimensional mathematical representation.

[0119] This step completes the mapping from unstructured optical images to structured feature values, providing quantized data with high noise resistance and stability for subsequent classification algorithms.

[0120] S5. Input the pollution feature parameter vector into the preset deep neural network model, output the corresponding pollution level classification label, and generate the clean status result based on the classification label.

[0121] Further, step S5, which involves inputting the pollution feature parameter vector into a preset deep neural network model, includes:

[0122] Construct a deep neural network architecture that includes an input layer, multiple fully connected hidden layers, and an output layer, and feed the contamination feature parameter vector into the input layer;

[0123] Layer-by-layer feature abstraction is performed using the nonlinear activation function of multi-layer fully connected hidden layers, followed by regularization processing.

[0124] In the output layer, the probability distribution of each pollution level is calculated using a normalized exponential function, and the node corresponding to the maximum probability value is extracted to trigger and output the corresponding pollution level classification label.

[0125] Specifically, a DNN architecture is constructed, comprising one input layer, three fully connected hidden layers, and one output layer. The system feeds a contamination feature parameter vector, generated in the preceding steps and containing pixel area proportions, brightness fluctuation values, and texture continuity, into the input layer of the DNN architecture. This feature parameter vector serves as the initial input data for the network, and its tensor dimension strictly corresponds to the number of neurons in the input layer, thus establishing a data flow path between the optical quantization parameters and the deep network architecture.

[0126] Within the network, the system utilizes nonlinear activation functions in three fully connected hidden layers to perform layer-by-layer feature abstraction on the fed contamination data. To accurately fit the complex nonlinear boundary of the contamination distribution on the orthokeratology lens surface, the hidden layers employ a linear rectified function as the activation operator and perform temporary deregulation to randomly disconnect some neuron connections, thereby suppressing overfitting of the network to stray light in a specific optical acquisition environment. The feature abstraction algorithm formula within the hidden layers is as follows:

[0127] ;

[0128] In the formula, Indicates the first High-order contamination feature tensor of orthokeratology lenses extracted from a hidden layer and Let these represent the weight matrix and bias vector of the fully connected network layer, respectively. This represents the feature data passed from the previous layer. Represents the linear rectification activation operator. This represents the random mask matrix used to perform regularization. Through the above layer-by-layer calculations, the network completes the mapping representation from basic optical parameters to the deep contamination features of orthokeratology lenses.

[0129] The higher-order contamination feature tensor, abstracted from the hidden layer, is passed to the end of the network. In one output layer, a normalized exponential function is used to calculate the probability distribution of each contamination level. The pre-defined contamination levels are categorized as clean lens, mild protein adhesion, and severe lipid deposition. The probability distribution formula for each level is as follows:

[0130] ;

[0131] In the formula, This indicates that the current orthokeratology lens belongs to the [number] category. The probability of determining the pollution level category. This represents the high-order contamination feature tensor output by the final hidden layer. and These represent the output layer corresponding to the first... The feature mapping weight matrix and classification bias scalar for each level category This represents the total number of preset pollution level categories. The system compares the generated probability distribution sequences and extracts the node corresponding to the highest probability value, assigning its network judgment state as a value. That is, the logical truth value, which triggers and outputs the corresponding pollution level classification label. Finally, the system directly generates a cleanliness status result based on the classification label, prompting the user to perform deep cleaning or wear the device as usual.

[0132] This step completes the model inference from multidimensional quantitative features to specific pollution classification, effectively improving the overall noise resistance stability and automated decision-making accuracy of the orthokeratology lens pollution detection system while completely eliminating the error of human visual judgment.

[0133] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An optical identification method for the surface contamination status of orthokeratology lenses, characterized in that, include: S1. Fix the orthokeratology lens to the detection base, and sequentially control the multi-angle light sources distributed according to the preset incident angle to irradiate the lens surface in a time sequence, generating optical reflection and scattering light signals on the lens surface; S2. In sync with the time-series illumination, the reflection and scattering images of the lens surface corresponding to the multi-angle light source are acquired respectively, and the pixel gray values ​​corresponding to each image are mapped to a preset gray range using a linear mapping function to complete the gray-level normalization processing of the image. S3. From the normalized image, the bright pollution area is segmented based on the dynamic gradient threshold, the scattering abnormal area is extracted using the local gray-level difference method, and the edge texture change area is extracted using the edge detection operator. S4. Calculate the percentage of pixel area of ​​the high-brightness contamination area, the scattering abnormal area, and the edge texture change area on the lens surface. Extract the brightness fluctuation value of the high-brightness contamination area based on the spatial random field model, and use the gray-level co-occurrence matrix to quantify the texture continuity of the edge texture change area. Then, splice them together to generate a contamination feature parameter vector. S5. Input the pollution feature parameter vector into a preset deep neural network model, output the corresponding pollution level classification identifier, and generate a clean status result based on the classification identifier.

2. The optical identification method for the surface contamination status of an orthokeratology lens according to claim 1, characterized in that, In step S1, the step of sequentially controlling multiple angle light sources distributed according to a preset incident angle to irradiate the lens surface in a time sequence includes: Based on a preset multi-dimensional illumination array, the independent spectral frequency bands and spatial incident angle parameters corresponding to each emission node in the multi-angle light source are retrieved. A logic controller generates timing trigger pulses to sequentially illuminate each light source node of the illumination array to output a structured light field; The beams are controlled to be projected onto the surface of the lens, separating the specular reflection component of the lens substrate from the diffuse scattering component of the pollution particles.

3. The optical identification method for the surface contamination status of an orthokeratology lens according to claim 1, characterized in that, Step S2, the step of acquiring the reflection image and scattering image of the lens surface corresponding to the multi-angle light source respectively, includes: Configure an external synchronization trigger pulse and connect it to the synchronization port of the image sensor to perform physical phase-locked control on light source illumination and shutter exposure; During the triggering cycle of the reflected incident angle light source, the reflection acquisition channel is activated to capture the reflection image of the lens surface containing macroscopic morphological features; During the scattering incident angle light source triggering cycle, switch to the scattering acquisition channel and adjust the photosensitive gain to capture the scattering image containing fine particle characteristics.

4. The optical identification method for the surface contamination status of an orthokeratology lens according to claim 1, characterized in that, In step S2, the step of mapping the pixel grayscale values ​​of each image to a preset grayscale range using a linear mapping function includes: The pixel intensity distribution of the reflected and scattered images on the lens surface is statistically analyzed, abnormal noise points are removed, and effective gray-level extreme values ​​are extracted. Construct a first-order affine transformation model based on the effective gray-level extreme values, and calculate the mapping parameters to the preset gray-level range; The mapping parameters are used to perform a linear scaling operation on all pixels to complete pixel recombination under a preset standard.

5. The optical identification method for the surface contamination status of an orthokeratology lens according to claim 1, characterized in that, In step S3, the step of segmenting the bright contamination area based on the dynamic gradient threshold includes: The normalized image is spatially convolved using a discrete differential operator to generate a gradient magnitude matrix representing the pixel mutation rate. The gradient magnitude matrix is ​​iteratively optimized based on the inter-class variance maximization criterion to obtain a dynamic segmentation threshold adapted to local illumination. Based on the dynamic segmentation threshold, the image is subjected to binarization truncation and morphological closing operation to extract the connected and closed bright contamination region.

6. The optical identification method for the surface contamination state of an orthokeratology lens according to claim 1, characterized in that, Step S3, the step of extracting the scattering anomaly region using the local gray-level difference method, includes: A low-pass smoothing filter is applied to the normalized image to construct a low-frequency background reference image characterizing the lens's underlying optical features; The normalized image is compared with the low-frequency background reference image to generate a residual grayscale image that highlights the diffuse reflection features of tiny particles. A local grayscale tolerance band is set, and pixel clusters with amplitude exceeding the limit and spatially clustered in the residual grayscale image are extracted to define the boundary of the scattering anomaly region.

7. The optical identification method for the surface contamination state of an orthokeratology lens according to claim 1, characterized in that, Step S4, the step of extracting the brightness fluctuation value within the high-brightness pollution area based on the spatial random field model, includes: A Markov random field network is constructed in the high-brightness pollution area, and the topological association structure of pixel nodes and their neighborhood system is defined. A potential function characterizing the optical dependence between nodes is constructed, and the autocorrelation of local gray levels is evaluated by calculating the neighborhood conditional probability distribution. The second-order statistical moments of the potential function in its energy convergence state are extracted and their dimensionality reduced and mapped to the brightness fluctuation values ​​representing the unevenness of the pollution thickness.

8. The optical identification method for the surface contamination state of an orthokeratology lens according to claim 1, characterized in that, In step S4, the step of quantizing the texture continuity of the edge texture change region using the gray-level co-occurrence matrix includes: Multi-scale spatial vector parameters are set in the edge texture change area to establish spatial relative position association rules for pixel pairs; The joint probability of pixel pairs that satisfy the association rule is calculated, the gray-level co-occurrence matrix of the target dynamic range is constructed, and normalization processing is performed. The texture continuity is output using the local stationarity and inverse difference moment parameter of the calculated gray-level co-occurrence matrix.

9. The optical identification method for the surface contamination state of an orthokeratology lens according to claim 1, characterized in that, Step S5, the step of inputting the pollution feature parameter vector into a preset deep neural network model, includes: A deep neural network architecture comprising an input layer, multiple fully connected hidden layers, and an output layer is constructed, and the pollution feature parameter vector is fed into the input layer. The nonlinear activation function of the multi-layer fully connected hidden layer is used to perform layer-by-layer feature abstraction and regularization processing. The output layer uses a normalized exponential function to calculate the probability distribution of each pollution level, extracts the node corresponding to the maximum probability value, and triggers and outputs the corresponding pollution level classification identifier.

Images