A method and system for correcting color degradation in a fiber optic color wheel

Abstract

The application discloses a kind of optical fiber color ring color degradation correction method and system, the method obtains optical fiber color ring image and extracts to be identified color ring area and reference area, constructs absorbance feature vector based on collection light intensity and reference light intensity;Color intrinsic dictionary and degradation disturbance dictionary are constructed, double-dictionary sparse decoding is carried out under candidate color category, and absorbance feature vector is decomposed into color intrinsic term, degradation disturbance term and residual term;According to reconstruction cost, determine the real color category, and output corrected color intrinsic representation, degradation reason and identification confidence;The method can be used for color degradation correction and cross-batch identification in optical fiber color ring online detection.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method and system for correcting color degradation in fiber optic color rings. Background Technology

[0002] In the optical cable manufacturing process, it is usually necessary to form ink rings of different colors and arrangements on the surface of the optical fiber to achieve coded identification of multi-core optical fibers, loose-tube optical fibers, or high-density optical fiber units. As optical communication systems develop towards high-density, multi-core, and high-speed manufacturing, the number of ink ring combinations is increasing, and production lines are placing higher demands on the color resolution, detection speed, cross-batch stability, and online feedback capabilities of ink ring detection systems.

[0003] Existing methods for detecting color rings in optical fibers mainly include manual visual inspection, color threshold recognition based on industrial cameras, and color classification based on ordinary machine learning or deep learning. These methods still have shortcomings in actual production lines. Since color rings typically form on the surface of transparent or semi-transparent cylindrical optical fibers, factors such as ink layer thickness, curing degree, surface reflection, light source spectrum, camera exposure, and glass substrate transmission all affect the apparent color captured by the camera. The same true ink color may appear whitish, grayish, yellowish, purplish, with residual highlights or blurred edges under different batches or operating conditions.

[0004] Traditional RGB, HSV, or Lab color features directly describe apparent color and are easily affected by variations in brightness, exposure, specular reflection, and ink thickness. When similar colors overlap after degradation, thresholding methods or ordinary classification models tend to mix true color information with information about operational disturbances, leading to confusion between similar colors. While end-to-end neural networks can learn complex mapping relationships, they often require re-labeling and retraining when ink batches change frequently, the number of labeled samples is limited, or production line conditions drift, and they struggle to explain the sources of color shifts.

[0005] Industrial environments not only require identifying the true color category on the current color wheel, but also determining whether color degradation is related to ink thinning, insufficient curing, exposure drift, residual highlights, or changes in light source conditions, in order to provide feedback to inkjet, curing, lighting, or production line control systems. Therefore, it is necessary to provide a technical solution capable of separating the intrinsic information of the true color and degradation perturbation information from the degraded apparent color, enabling the system to stably identify the true color even under conditions of color degradation and batch variation, and to output an explainable cause of degradation. Summary of the Invention

[0006] The present invention aims to at least solve the problems in the prior art where the apparent color of fiber optic color rings degrades due to the influence of light, exposure, reflection, ink thickness and curing state, resulting in unstable true color recognition, easy confusion of similar colors, insufficient generalization ability across batches and difficulty in explaining the causes of degradation, and provides a method and system for color degradation correction of fiber optic color rings.

[0007] To address the aforementioned technical problems, this invention discloses a method for correcting color degradation in optical fiber color rings, comprising the following steps: S01, acquire an optical fiber color ring image, and extract the color ring region to be identified and the reference region from the optical fiber color ring image; S02, construct an absorbance feature vector based on the collected light intensity of the color ring region to be identified and the reference light intensity of the reference region; S03, train or call the color intrinsic dictionary based on standard color wheel samples. The color intrinsic dictionary includes color intrinsic sub-dictionaries corresponding to multiple standard color categories, and limits the dimension, number of atoms, atom normalization conditions, and calculation construction formulas of each dimension element in each color intrinsic atom of the color intrinsic sub-dictionary. S04, based on the degradation residual between the degradation color ring sample and the corresponding color intrinsic reconstruction result, train or call the degradation perturbation dictionary, the degradation perturbation dictionary includes a shared degradation perturbation dictionary and a color conditional degradation perturbation sub-dictionary, and establish a mapping relationship between degradation atoms and degradation causes; S05, perform dual-dictionary sparse decoding on the absorbance feature vector of the color ring sample to be identified, and decompose the absorbance feature vector into color intrinsic terms, degradation perturbation terms and residual terms; S06, determine the true color category based on the sparse reconstruction objective function corresponding to the candidate color category, and output the corrected color intrinsic representation; S07, output the degradation cause based on the activated degradation atoms and their degradation cause labels in the degradation perturbation term, and output the identification confidence based on the reconstruction residual, class interval, sparse coefficient concentration and degradation residual intensity; S08 enables rapid adaptation and updating of the color intrinsic dictionary, degradation perturbation dictionary, or sparse decoding parameters based on a small number of calibration samples when changing ink batches, light source status, or production line conditions.

[0008] In the preferred technical solution, step S02 includes: setting the color ring region to be identified as... The reference area is , Indicates the pixel position within the region. Indicates the first Each wavelength or equivalent color channel The area acquisition light intensity and reference light intensity are calculated as follows:

[0009] in, and These represent the number of pixels included in the statistics within the color wheel region to be identified and the reference region, respectively. Represents pixels exist The grayscale or light intensity collected under the channel; Represents reference area pixels exist The reference grayscale or light intensity under the channel. The light intensity of the region can also be calculated using the median, truncated mean, de-highlighted mean, or weighted mean to reduce the impact of abnormal highlight pixels.

[0010] In the preferred technical solution, step S02 further includes: constructing the absorbance feature as follows:

[0011] in, Indicates the first One sample in Absorbance characteristics under the channel, To prevent small positive numbers where the denominator or numerator is zero.

[0012] In the preferred technical solution, step S02 further includes: the first The absorbance feature vector of each color ring sample is represented as follows:

[0013] in, Indicates the first An additional statistical characteristic, The additional statistical features include at least one of the following: color ring axial absorption curve statistics, radial gray-scale gradient statistics, edge sharpness, local texture, normalized channel ratio, or color features after polarization dereflection. .

[0014] In the preferred technical solution, step S03 includes a color intrinsic dictionary. Represented as:

[0015] in Indicates the number of standard color categories. Indicates the first The color intrinsic sub-dictionary corresponding to the standard color is determined based on the silhouette coefficient, intra-class variance, or validation set recognition accuracy.

[0016] In the preferred technical solution, the color intrinsic sub-dictionary atoms in step S03 are defined as follows:

[0017] in, Indicates the first The number of color intrinsic atoms in the sub-dictionary of color intrinsics. , Indicates the first The first color intrinsic sub-dictionary The intrinsic color atom in the first... Elements in the absorbance feature dimension Preferably, By the The dictionary element is obtained by normalizing the intra-class cluster mean of the standard color wheel samples along the corresponding feature dimensions, so that each dictionary element originates from the stable absorption fingerprint of the standard color samples. The intra-class clustering uses K-means clustering.

[0018] Preferably, the number of intrinsic color atoms Shared number of degenerate atoms and the number of degenerate atoms under color conditions It is not arbitrarily assigned, but rather determined jointly based on the corresponding number of training samples, the number of candidate clusters, the normalized clustering error reduction rate, the minimum number of samples per atom, and the validation set performance. The specific determination process is described in the following detailed implementation.

[0019] In the preferred technical solution, the degradation perturbation dictionary in step S04 is represented as follows:

[0020] in , This indicates a shared degenerate perturbation dictionary. Indicates the number of shared degenerate atoms; Indicates the first A sub-dictionary of color conditional degradation perturbations corresponding to the color class. Indicates the first The number of atoms in the color conditional degeneration.

[0021] In the preferred technical solution, the degenerate perturbation sub-dictionary atoms in step S04 are defined as follows:

[0022] in , Indicates a shared degenerate perturbation dictionary; otherwise, it indicates the first... Color conditional degradation perturbation sub-dictionary; Indicates the first The first degenerate perturbation sub-dictionary The degenerate atom in the first Elements in the absorbance feature dimension. Shared degradation perturbation atoms are obtained by normalizing the cluster mean of the normalized degradation residuals, and color conditional degradation perturbation atoms are obtained by normalizing the intra-class cluster mean of the color conditional residuals after deducting the shared degradation components. The element-by-element construction formula is further defined in the specific implementation.

[0023] In the preferred technical solution, the dictionary atoms in steps S03 and S04 satisfy the normalization constraint:

[0024] in, express The first in Each color intrinsic atom express The first in A degenerate perturbation atom.

[0025] In the preferred technical solution, step S05 includes: assuming the color ring sample to be identified belongs to the first... If the color is a certain type, then its absorbance feature vector is... It is represented as:

[0026] in, Indicates the first Color-related intrinsic sparsity coefficients Indicates the shared degradation sparsity coefficient. Indicates the first Color-type conditional degradation sparsity coefficient, Represents the residual noise term. For color eigenvalues, This is a degenerate perturbation term.

[0027] In the preferred technical solution, step S06 includes: the sparse decoding objective function corresponding to the candidate color category is:

[0028] in, This indicates that the sample belongs to the first... The cost of candidate category reconstruction when dealing with color. This represents the intrinsic sparsity regularization coefficient of color. This represents the sparse regularization coefficient of the degradation perturbation. Describing the L2 norm, This represents the L1 norm. The true color category is determined as follows:

[0029] In the preferred technical solution, step S05 further includes: introducing conditional coupling constraints in dual-dictionary sparse decoding; when using full-dictionary decoding, for any sample and any color category The following constraints are applied to all:

[0030] in, Indicates the first The sample at the th Degeneracy sparsity coefficients on the sub-dictionary of color conditional degradation perturbation. Indicates the first The sample at the th Color eigenvalue sparsity coefficients in the class color eigenvalue sub-dictionary. The conditional coupling parameter is used to limit the maximum allowable amplitude ratio of the color conditional degradation perturbation sparse coefficient to the color intrinsic sparse coefficient. Its initial range is determined based on the empirical ratio of the degradation perturbation coefficient to the color intrinsic coefficient under the true color category, and then calibrated by the recognition accuracy, category interval, reconstruction residual and degradation interpretation consistency on the validation set.

[0031] In the preferred technical solution, the dictionary training in steps S03 and S04 introduces low coherence constraints:

[0032] in, Indicates low coherence loss. This represents the maximum mutual coherence between the color eigenvalue dictionary and the degenerate perturbation dictionary. This represents the Frobenius norm.

[0033] In the preferred technical solution, step S07 includes: defining a set of degradation causes. ,in This indicates a cause of degradation, which may include at least one of the following: whitening, graying, yellowing, purplishening, thin ink, insufficient curing, exposure drift, residual highlights, light source color temperature drift, or ink batch shift. For each degradation atom in the degradation perturbation dictionary... Set degradation reason labels .

[0034] For online samples, the degradation coefficients corresponding to the selected candidate categories are concatenated as follows: And denote the corresponding set of degenerate atoms as For any cause of degradation Its activation score is defined as:

[0035] The main causes of degradation are output as follows:

[0036] when Less than the preset degradation cause threshold When the cause of degradation is not obvious, the output is: when multiple causes of degradation satisfy the condition relative to the threshold. At that time, multiple reasons for degradation will be output.

[0037] The degradation cause threshold Used to determine the cause of degradation The activation score is determined based on whether it exceeds the weak response range caused by normal sample fluctuations, image noise, and sparse solution errors. The degradation cause threshold is determined based on the activation score distribution of validation samples containing and not containing the corresponding degradation cause, in order to balance the detection rate and false alarm rate of degradation causes; different degradation cause thresholds can be set for different degradation causes.

[0038] The relative threshold This is used to determine whether other degradation causes contribute sufficiently significantly relative to the primary degradation cause. The relative threshold is determined based on the relative relationship between the activation scores of each degradation cause in single-factor degradation samples and mixed degradation samples, in order to reduce the erroneous output of single-factor degradation samples as multiple degradation causes, while retaining valid secondary degradation causes in mixed degradation samples.

[0039] During online judgment, when the maximum activation score is lower than the corresponding degradation cause threshold, the degradation cause is not obvious; when the maximum activation score reaches the corresponding degradation cause threshold, its corresponding degradation cause is output as the primary degradation cause. Other degradation causes are only output as secondary degradation causes when their activation scores reach their corresponding degradation cause thresholds and their activation scores relative to the activation scores of the primary degradation cause reach the relative threshold.

[0040] In the preferred technical solution, step S07 further includes: identifying the confidence level. The calculation is performed using category spacing, color intrinsic sparsity coefficient concentration, normalized residual strength, and minimum reconstruction cost. Let... , Then the category interval Concentration of intrinsic sparsity coefficients of color Normalized residual strength They are represented as follows:

[0041] The confidence level for identification is defined as:

[0042] in, This represents the Sigmoid function. , , , , This represents the confidence weight coefficient determined by the validation sample set. When When this happens, the system will mark the current sample as a suspected abnormal sample. The confidence threshold is determined based on the validation sample set. Specifically, the corresponding recognition confidence set is calculated for normal color ring validation samples, and the initial threshold is determined based on the mean and standard deviation of the confidence scores of normal samples.

[0043] in and These represent the mean and standard deviation of the confidence level for normal samples, respectively. This is the threshold adjustment coefficient.

[0044] In the preferred technical solution, step S08 includes defining different ink batches, different light source conditions, or different production line conditions as different tasks. Each task includes a support set and query set The dictionary initialization parameters and sparse decoding parameters that can be quickly adapted to new tasks are obtained through meta-learning; when a new batch is launched, the dictionary initialization parameters and sparse decoding parameters are fine-tuned using a small number of calibration samples, and the update range is limited by anti-drift constraints.

[0045] This invention also discloses a fiber optic color ring color degradation correction system, comprising: an image acquisition module for acquiring fiber optic color ring images; a color ring region extraction module for extracting a color ring region to be identified and a reference region from the fiber optic color ring image; an absorbance feature construction module for constructing an absorbance feature vector based on the acquired light intensity of the color ring region to be identified and the reference light intensity of the reference region; a color intrinsic dictionary module for storing, training, or updating color intrinsic sub-dictionaries corresponding to multiple standard color categories; and a degradation perturbation dictionary module for storing, training, or updating a shared degradation perturbation dictionary and color... The system includes: a color conditional degradation perturbation sub-dictionary; a dual-dictionary sparse decoding module, used to decompose the absorbance feature vector into color intrinsic terms, degradation perturbation terms, and residual terms based on the color intrinsic dictionary and degradation perturbation dictionary; an online recognition output module, used to output the true color category, the corrected color intrinsic characterization, and the recognition confidence based on the reconstruction cost of the candidate color category; a degradation cause explanation module, used to output the degradation cause based on the activated degradation atoms in the degradation sparse coefficients; and a fast adaptation module, used to update the dictionary or sparse decoding parameters based on a small number of calibration samples when changing ink batches, light source states, or production line conditions.

[0046] Preferably, the system further includes a process feedback interface module, which is used to output the degradation cause, true color category, identification confidence level or anomaly mark to the inkjet equipment, curing equipment, lighting controller or production line control system.

[0047] Compared with the prior art, the present invention has at least the following beneficial effects: First, this invention converts the apparent color of the fiber color ring into absorbance characteristics and clearly defines the calculation methods for regional light intensity, reference light intensity, absorbance channels, and additional statistical characteristics, so that color characterization is related to ink absorption characteristics, which can reduce the impact of light source intensity, exposure changes, and reflective residue on color recognition.

[0048] Second, the present invention uses a dual-dictionary sparse decoupling model consisting of a color intrinsic dictionary and a degradation perturbation dictionary to decompose degraded apparent color into color intrinsic terms and degradation perturbation terms, enabling the system to both identify the true color category and explain the source of color degradation.

[0049] Third, this invention limits the dimensions, number of atoms, sparsity coefficient dimension, and atom normalization conditions of the color intrinsic sub-dictionary, the shared degradation perturbation dictionary, and the color conditional degradation perturbation sub-dictionary, thereby improving the model's implementability and parameter interpretability.

[0050] Fourth, this invention establishes a correspondence between degradation coefficients and degradation causes by using degradation atom tags and degradation cause activation scores, and can output degradation causes such as whitening, graying, thin ink, insufficient curing, exposure drift, and highlight residue.

[0051] Fifth, this invention restricts the activation conditions of the color conditional degradation perturbation sub-dictionary by limiting the conditional coupling constraints, which can reduce the situation where the degradation perturbation dictionary replaces the color intrinsic dictionary for classification and improve the separability of color intrinsic information and degradation information.

[0052] Sixth, this invention reduces feature overlap between the intrinsic color dictionary and the degenerate perturbation dictionary by using low coherence constraints, thereby improving the physical interpretability and engineering traceability of the dictionary training and online decoding process.

[0053] Seventh, this invention improves the feasibility of online verification, anomaly labeling, and few-sample adaptation triggering by integrating the class interval, sparse coefficient concentration, normalized residual strength, and minimum reconstruction cost through a confidence function.

[0054] Eighth, this invention enables the system to update the dictionary and sparse decoding parameters using a small number of calibration samples when changing ink batches, light source status or production line conditions, thereby reducing the cost of re-labeling and retraining. Attached Figure Description

[0055] Figure 1 This is the overall flowchart of the fiber optic color ring color degradation correction method and system of the present invention; Figure 2 This is an enlarged view of a representative optical fiber color ring sample and reference area, and a color ring area to be identified in an embodiment of the present invention. Figure 3 This is a table of RGB truncated mean values ​​and relative absorbance of the reference area and four axial color ring blocks in the embodiments of the present invention. Figure 4 This is a diagram illustrating the process of calculating and arranging a 12-dimensional absorbance vector from the statistical values ​​of the RGB regions channel by channel in an embodiment of the present invention. Figure 5 This is a RGB absorbance response distribution diagram of eight intrinsic color dictionary atoms in an embodiment of the present invention; Figure 6 This is a comparison chart of the uncorrected baseline and the dual-dictionary color recognition rate under six imaging states in the embodiments of the present invention; Figure 7 This is a comparison chart of the observation intrinsic error and the corrected intrinsic error for two sets of samples in an embodiment of the present invention; Figure 8 These are representative optical fiber color ring sample images of different colors and imaging states in the embodiments of the present invention. Detailed Implementation

[0056] The following combination Figure 1The present invention will be further described below. It should be noted that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. Unless otherwise specified, the following embodiments and their technical features can be combined with each other.

[0057] The following also combines Figures 2 to 8 The calculation examples of RGB region statistics to absorbance vector, color intrinsic dictionary response, and procedural verification results corresponding to steps S05 to S07 are further explained.

[0058] To ensure that the formulas, matrices, and processing procedures involved in this invention have clear meanings, this embodiment directly explains the meanings of the main terms, variables, and matrices in the following order: image region, absorbance feature, color category, sample set, dictionary matrix, sparse coefficient, loss function, confidence level, and few-sample fitting parameters.

[0059] In terms of regional, channel, and light intensity statistics, the first The color ring regions to be identified are denoted as It is the region containing color ring pixels or local color ring blocks, located from the fiber color ring image; the first The reference region corresponding to each sample is denoted as This can be a region without color rings, a transparent fiber region, a standard white field region, or an online reference region. Pixel positions are... This indicates that it is used to perform summation, mean, median, or weighted statistics on the pixels participating in the statistics within a region. Each wavelength or equivalent color channel is denoted as , You can number the true spectral wavelength, RGB channels, multispectral channels, polarization channels, or equivalent color channels. ,in Indicates the number of absorbance channels. The color ring area to be identified is in The regional statistical light intensity under the channel is denoted as Reference area is The regional statistical light intensity under the channel is denoted as Small positive numbers Used to prevent the denominator or numerator from being zero during logarithmic, division, or normalization processes.

[0060] Regarding the construction of absorbance eigenvectors, the first... One sample in The absorbance calculated from the acquired light intensity and the reference light intensity in the channel is denoted as . In addition to the absorbance of each channel, the sample may also include additional statistical characteristics. , ,in Indicates the number of additional statistical features. This can include axial absorption curve statistics, radial grayscale gradient, edge sharpness, texture features, channel ratio, or polarization dereflection features. From the above... Absorbance of each channel and The additional statistical features are spliced ​​together to form the first Absorbance feature vectors ,and , .

[0061] Regarding color categories, sample indexes, and sample sets, the number of standard color categories is denoted as: Standard color categories may include red, orange, yellow, green, blue, purple, brown, black, or other ink color categories defined by the production line; the color category index is denoted as... , The sample index is denoted as... The total number of samples is denoted as , . No. The true color category label of each training sample or calibration sample is denoted as . The true color category determined based on the reconstruction cost of candidate categories during online detection is denoted as... . No. The standard sample set is denoted as:

[0062] No. The set of degenerate samples is denoted as:

[0063] No. The number of standard samples is denoted as The total number of degenerate samples participating in the training of the degenerate dictionary is denoted as . , No. The number of degenerate samples is denoted as .

[0064] Regarding the color eigenvalue dictionary, the first... The dictionary of color intrinsic sub-sub ... , , used to characterize the Stable absorption of fingerprints by color-like particles in the absorbance space, among which Indicates the first The number of color eigenato atoms in the sub-dictionary of color eigenvalues. The color eigenvalue dictionary is denoted as... , , . No. Class 1 Each color intrinsic atom is denoted as , , it is The The _th atom, used to represent the _th _th atom A stable absorber sub-mode similar to standard colors. The feature dimension index is denoted as... , , This represents the first eigenatom of the color. The element is the first element. Class Standard Samples in the 1st The mean of intra-class clusters is obtained by normalizing the mean of clusters on each feature dimension.

[0065] In the element-wise construction of color intrinsic atoms, the first Class 1 The normalized standard samples are denoted as follows: ,in The index of the atom belonging to a class is denoted as... , which indicates the first Class 1 Each standard sample belongs to which color intrinsic atom or intraclass cluster center? The set of sample indices of each intrinsic color atom is denoted as The first before normalization Each color eigenvalue is denoted as . , its origin Chinese standard sample The mean of the dimensional features is calculated; subsequently, the mean of each dimensional feature is obtained. The column vectors are normalized using the L2 norm to obtain the final dictionary elements. .

[0066] Regarding the color eigensample matrix and update matrix, the first... The standard sample matrix is ​​denoted as , Each column is a first column. The absorbance feature vector of a standard color wheel sample. The color-class intrinsic sparse coefficient matrix is ​​denoted as Specifically:

[0067] Used for color intrinsic sub-dictionary Least squares update. The unnormalized color eigenvalue update matrix is ​​denoted as... It can be derived from The calculation yielded, where Update the regularization parameters for the dictionary. for An identity matrix of dimensionality; The Line 1 After normalizing the columns, we get .

[0068] Regarding the degradation perturbation dictionary, the shared degradation perturbation dictionary is denoted as... , This is used to characterize a common degradation pattern that may occur in multiple color categories, among which Indicates the number of shared degenerate atoms. The color conditional degradation perturbation sub-dictionary corresponding to the color class is denoted as , , used to characterize the The degradation offset mode unique to color classes, in which Indicates the first The number of degenerate atoms under color-conditional degradation. The degradation perturbation dictionary is denoted as... Specifically:

[0069] in .

[0070] To unify the representation of the shared degradation perturbation dictionary and the color condition degradation perturbation sub-dictionary, the first Each degenerate perturbation sub-dictionary is denoted as ,in Indicates a shared degenerate perturbation dictionary; otherwise, it indicates the first... Color conditional degradation perturbation sub-dictionary . No. The first degenerate sub-dictionary Each degenerate atom is denoted as , , it is The List. The first degenerate atom The element is the normalized degenerate residual or the color conditional residual at the th element. The cluster mean is obtained by normalizing the cluster mean across the feature dimensions.

[0071] Regarding the degenerate residuals and the construction of shared degenerate atoms, the eigenvectors of the degenerate color ring samples are denoted as... It is constructed from degraded color ring samples according to absorbance characteristics, and The degenerate residual vector is denoted as... , used to represent the offset of a degraded sample relative to the color intrinsic reconstruction result, where Indicates the first The color eigenvalue reconstruction results of each degenerate color wheel sample on its true color category corresponding color eigenvalue sub-dictionary, the specific calculation steps are as follows: The normalized degenerate residual vector is denoted as... This is used for element-wise construction of the shared degenerate perturbation dictionary. The shared degenerate atom attribution index is denoted as... , which indicates the first The degenerate residual belongs to which shared degenerate atom; the first The sample set of shared degenerate atoms is denoted as Shared degenerate atoms before normalization. The elements are denoted as , its origin Normalized degenerate residuals The The mean value is calculated.

[0072] Regarding the residuals of color condition degradation and the atomic structure of color condition degradation, the first... The color-conditional residual vector is denoted as , used to construct the first A sub-dictionary of color conditional degradation perturbations. The normalized color conditional residual vector is denoted as... The color condition degenerates the atom attribution index as . , which indicates the first The first The color conditional residual belongs to which conditional degenerate atom? The sample set of color conditionally degenerate atoms is denoted as Color condition degeneration atoms before normalization The elements are denoted as , its origin Normalized color condition residual The The mean value is calculated.

[0073] Regarding the degenerate dictionary update matrix, the shared degenerate dictionary update matrix and the corresponding coefficient matrix are denoted as follows: and The color conditional degradation dictionary update matrix and the corresponding coefficient matrix are denoted as follows: and The matrix described above is used to update the degenerate perturbation dictionary using least squares with Tikhonov regularization, and then renormalizes the degenerate atoms in each column after the update.

[0074] Regarding dual-dictionary sparse decoding and coefficient matrices Indicates the first Color-related intrinsic sparsity coefficients ; Indicates the shared degradation sparsity coefficient. ; Indicates the first Color-type conditional degradation sparsity coefficient, The residual noise term is denoted as... This is used to represent unmodeled noise, contamination, edge defects, or anomalous perturbations. During the joint training phase, the third... The color intrinsic sparse coefficients of each sample are expanded into a global vector. The degradation perturbation sparse coefficients are expanded into a global vector. And define the global color eigenvalue sparse coefficient matrix. Global degenerate perturbation sparse coefficient matrix .

[0075] Regarding the loss function and constraint parameters, and These are the intrinsic color sparsity regularization coefficient and the degradation perturbation sparsity regularization coefficient, respectively. This is a conditional coupling parameter used to restrict the color conditional degradation perturbation sub-dictionary to be activated only when the corresponding color intrinsic sub-dictionary is activated; This is the category margin parameter, used to discriminate the loss, ensuring that the reconstruction residual of the true color category is smaller than the reconstruction residual of other categories; , , These represent the weights of the low coherence constraint, the discrimination loss, and the conditional coupling loss, respectively. The low coherence projection step size is used to reduce the overlap between the degenerate perturbation dictionary and the color intrinsic dictionary. The double dictionary reconstruction error of the training samples is denoted as... It is the sum of the reconstruction errors of the training samples on the color eigenvalue dictionary and the degenerate perturbation dictionary. The dictionary cross-correlation matrix is ​​denoted as... , , its first Line 1 Column elements are .

[0076] In terms of candidate category identification and confidence calculation This indicates that the sample belongs to the first... Minimize the sparse reconstruction objective function for color classes; the minimum reconstruction cost is denoted as . The cost of the second smallest reconstruction is denoted as The category interval normalization factor is denoted as... The concentration of the intrinsic sparsity coefficients of color is denoted as The normalized residual intensity is denoted as , which is the normalized value of the residual relative to the sample energy after dual-dictionary reconstruction. The recognition confidence is denoted as . The confidence threshold is denoted as ,when When this happens, the system will mark the current sample as a suspected abnormal sample.

[0077] In terms of explaining the causes of degradation and fitting to small samples, the set of causes of degradation is denoted as . The degradation reason label corresponding to each degenerate atom is denoted as Causes of degradation The activation score is recorded as The system through , and The degradation sparsity coefficient is mapped to degradation causes such as whitening, graying, thin ink, insufficient curing, exposure drift, or highlight residue. Rapid adaptation with a small number of samples is underway. Individual learning tasks are denoted as Support sets and query sets are denoted as follows: and The set of parameters to be optimized by meta-learning is denoted as... It includes a dictionary, sparse parameters, threshold, and normalization parameters; This represents the step size of the meta-learning inner loop. The anti-drift regularization weight in the new batch adaptation is denoted as... , and The maximum allowed dictionary update size is denoted as This is used to limit the update magnitude of the color intrinsic dictionary, degradation perturbation dictionary, and decoding parameters on a small number of calibration samples.

[0078] The fiber optic color ring color degradation correction system of the present invention includes an image acquisition module, a color ring region extraction module, an absorbance feature construction module, a color intrinsic dictionary module, a degradation perturbation dictionary module, a dual-dictionary sparse decoding module, an online identification output module, a degradation cause explanation module, a fast adaptation module, and an optional process feedback interface module.

[0079] The image acquisition module may include an industrial camera, lens, light source, acquisition card, and trigger control unit. The light source may be a ring light, coaxial light, multi-angle illumination, multispectral light source, or polarized illumination assembly. The camera may be a color camera, multispectral camera, polarized camera, or a combination of multiple cameras. For high-speed moving optical fibers, the image acquisition module may use short exposure, high frame rate acquisition, stroboscopic illumination, or motion compensation methods to obtain color ring images.

[0080] The color ring region extraction module is used to extract the color ring region to be identified based on the fiber edge, fiber centerline, axial projection, color change, or candidate color ring detection results. And extract adjacent non-color wheel reference regions. Alternatively, an online reference area. The absorbance feature construction module is used to convert the collected light intensity of the color ring region to be identified and the reference light intensity of the reference region into absorbance feature vectors. The color intrinsic dictionary module stores and updates stable absorption fingerprints for each standard color category. The degradation perturbation dictionary module stores and updates the shared degradation perturbation dictionary and the color conditional degradation perturbation sub-dictionary. The dual-dictionary sparse decoding module solves for the color intrinsic sparsity coefficients and degradation sparsity coefficients. The online recognition output module outputs the true color category, the corrected color intrinsic characterization, and the recognition confidence score. The degradation cause explanation module outputs degradation causes such as whitening, graying, yellowing, thinning, insufficient curing, exposure drift, or highlight retention based on the degradation sparsity coefficients.

[0081] Combination Figure 1 The overall process is as follows: preferably, after passing through an online coating or inkjet printer, the optical fiber enters the inspection station, where continuous image acquisition of the fiber's color rings is performed. The color ring region extraction module first locates the fiber edge and fits the fiber centerline; then, it calculates the color change, brightness change, or absorbance change along the fiber axis to determine candidate color ring bands; subsequently, the color ring region to be identified is obtained based on the axial boundary and radial range of the candidate color ring bands. .

[0082] Combination Figure 8 As shown, fiber optic color ring samples of different colors and under different imaging conditions have appearance characteristics such as thin optical fibers, narrow color rings, low-contrast backgrounds, discontinuous coverage of color rings, and local brightness variations. Figure 8 The images sequentially display the blue standard state, the whitish state, the thin ink state, the high-gloss residue state, the exposure drift state, the color temperature drift state, the hue shift state, and the yellow standard state state. The image acquisition module uses these samples as input, and the color ring region extraction module locates the color ring band along the fiber optic axis and selects reference areas before and after the color ring band to construct subsequent absorbance features.

[0083] Preferably, the color ring region to be identified Including the color ring region extending along the fiber axis, and the reference region. This includes the transparent fiber regions adjacent to the color ring region. The reference region can be obtained in the same frame as the color ring region to be identified, or it can be obtained from online white field, standard fiber samples, or periodic calibration images.

[0084] To reduce edge blurring caused by motion, the system can perform motion compensation based on the fiber optic movement speed, camera exposure time, and trigger period. For multi-frame images, fiber axial registration can be performed first, followed by statistical fusion of multiple local block features within the same color ring to improve the stability of absorbance features.

[0085] Combination Figure 1The overall process shown involves the absorbance feature construction module calculating the absorbance features of pixels, local blocks, or statistical regions within the color wheel region to be identified. For the first... One sample, calculate the light intensity of the area. Reference light intensity and absorbance characteristics .

[0086] Under the ideal absorption model, absorbance satisfies:

[0087] in, Indicates the ink material in The equivalent absorption coefficient under the channel, Indicates ink concentration or equivalent absorption intensity. Indicates the ink layer thickness or equivalent optical path. This indicates a disturbance term introduced by residual reflection, exposure drift, light source color temperature drift, or sensor noise. It mainly reflects the absorption of fingerprints by the ink color material. and It reflects the ink concentration, thickness, or equivalent absorption intensity.

[0088] For a color wheel sample, multiple wavelengths, multiple color channels, multiple sampling points, or multiple statistical features are concatenated into an absorbance feature vector. Additional statistical features may include the mean of the axial absorption curve, the variance of the axial absorption curve, the mean of the radial gray-level gradient, the peak value of the radial gray-level gradient, edge sharpness, local texture energy, normalized channel ratio, or color features after polarization dereflection. For example, the normalized channel ratio can be expressed as:

[0089] Edge sharpness can be expressed as the mean value of the gray level or absorbance gradient amplitude near the axial boundary of the color wheel:

[0090] in, Indicates the first The system converts the apparent color into a stable representation related to ink absorption characteristics through the above processing, reducing the impact of light source intensity, exposure variations, specular reflection residue, and camera gain variations on subsequent color recognition.

[0091] Combination Figures 2 to 4 As shown, to illustrate the specific relationship between RGB image pixel values, regional statistical light intensity, absorbance vector, and dictionary input, a representative sample of blue color ring specular residue is used for recalculation. Figure 2This shows a representative fiber color ring sample and a reference area, as well as the color ring area to be identified. Figure 3 The RGB truncation mean and relative absorbance of the corresponding region are shown. Figure 4 This illustrates the channel-by-channel calculation and 12-dimensional vector arrangement. The reference region and four axial color ring blocks expand the pixels within the region according to the R, G, and B channels, respectively. After deleting the lowest and highest 8% of pixels for each channel, the truncated mean is calculated for the remaining pixels. Therefore, the following values ​​are the channel statistical mean of multiple image pixels within the region, ranging from 0 to 255, and are not individual pixel values, absorbance values, or dictionary coefficients.

[0092] The RGB truncated mean of the reference region is The RGB truncated mean values ​​of the four axial color ring blocks are as follows: , , and For the first Channels of each axial color ring block Calculate the absorbance using the following formula:

[0093] In the formula, For reference area in the passage The truncated mean; For the first Each axial color ring block in the channel The truncated mean; To prevent the logarithmic independent variable from reaching zero, the numerical stability constant in this example is taken based on the gray levels of an 8-bit image. For example, the absorbance of the R channel of color ring block 2 is... .

[0094] Arrange the R, G, and B channels sequentially in blocks 1 to 4 to obtain a 12-dimensional absorbance vector:

[0095] The absorbance is a relative absorbance calculated relative to a neighboring reference region. When the... Channel statistical light intensity of each color ring block Slightly higher than the light intensity of the reference area hour, It can be a negative value, which indicates an increase in channel brightness relative to the reference area, but does not indicate that the color ring material has a negative absolute physical absorbance.

[0096] The axial standard deviation of each channel is calculated from the absorbance of the four axial blocks to obtain... , , Additional statistical characteristics to The 12-dimensional absorbance vector is calculated using the aforementioned method based on the proportion of highlight pixels, local texture amount, reference area chromaticity ratio, luminance dispersion, and chromaticity dispersion, respectively. to The 20-dimensional observation features are assembled in a fixed order. This order is consistent with the row order of each 20-dimensional atom in the color intrinsic dictionary and the degenerate perturbation dictionary, and is then input into the dual-dictionary sparse decoding.

[0097] Combination Figure 1 The overall process shown sets the standard color category to... Class. For the first Standard color wheel sample set Color intrinsic dictionary module training Color intrinsic sub-dictionary The specific training process is as follows: Order No. The sample set of standard color wheels is represented as follows:

[0098] in, Indicates the first Number of standard color wheel samples Indicates the first Class 1 The absorbance feature vectors of a standard color wheel sample. To eliminate the overall intensity differences between different samples, the standard samples are first normalized:

[0099] No. The dictionary of color intrinsics consists of Composed of intrinsic atoms of each color:

[0100] in , Indicates the first Atomic indices in the class color intrinsic sub-dictionary This represents the dimension index of the absorbance feature vector. For the normalized standard samples... Perform intra-class clustering to obtain the atom affiliation index for each sample:

[0101] in Indicates the first Class Standard Samples in the 1st Intra-class cluster centers Indicates belonging to the first A standard sample index set of eigenatomes of _ ... The first color intrinsic atom The element is first determined by the corresponding standard sample. The mean of each feature dimension is calculated to obtain:

[0102] in express The Each element. Then, the vector consisting of the means of each dimension... After normalizing the columns, we obtain the final elements of the color eigenatomes: any element in All by the first Standard color wheel samples in the 1st Stable statistical values ​​on each absorbance feature dimension are calculated, rather than being directly given by degraded samples or online samples to be identified. hour, Including the All standard samples of the class, at this time Degenerate into the first Normalized class mean of standard color absorbs fingerprints; when At different times These represent stable absorber modes of the same standard color under different standard thicknesses, standard curing states, standard lighting conditions, or permissible manufacturing fluctuations.

[0103] After completing the above initialization, sparse dictionary learning can be used to iteratively refine the color intrinsic sub-dictionary. Let... For the first Class Standard Samples The sparse coefficient matrix on is then fixed. A dictionary update at a given time can be represented as:

[0104] in, This represents a small regularization parameter used to ensure matrix invertibility or improve numerical stability. Represents the identity matrix. This indicates that the unnormalized update of the dictionary is in the first position. line, number The elements of the column. Through the above calculations, the color intrinsic sub-dictionary further approximates the first element while maintaining atomic normalization. Stable absorption structures for a standard sample set.

[0105] Preferably, the first The training objective of the class color intrinsic sub-dictionary is:

[0106] in, Indicates the first The standard sample at the th The sparse coefficients on the sub-dictionary of color eigenvalues. The element-by-element construction of the color eigenvalue atoms is performed according to the above process, specifically by the first... The mean of the intra-class clusters of the standard color wheel samples is obtained by normalization, and can be further iteratively corrected through sparse dictionary learning.

[0107] Combination Figure 5 As shown, in a set of runnable dictionary instances, the standard color categories include red, orange, yellow, green, blue, purple, brown, and black, with each color having four intrinsic color atoms. The first 12 dimensions of each atom correspond to the R, G, and B absorbance of the four axial blocks, respectively, and the last 8 dimensions correspond to... to . Figure 3 The first 12 dimensions of the four atoms for each color are summarized: after merging the four axial blocks of the four atoms, 16 trained response values ​​are obtained for each color and each channel, and their mean and standard deviation are plotted.

[0108] In the example, the mean absorbance responses of the R, G, and B channels for the red intrinsic atom are 0.0525, 0.1058, and 0.1043, respectively; for orange, 0.0528, 0.0833, and 0.1047; for yellow, 0.0625, 0.0687, and 0.1218; for green, 0.1215, 0.0698, and 0.1192; for blue, 0.1030, 0.0847, and 0.0535; for purple, 0.0842, 0.1076, and 0.0714; for brown, 0.0883, 0.1083, and 0.1144; and for black, 0.1159, 0.1173, and 0.1170. These values ​​were obtained by statistically analyzing the first 12 dimensions of the actual trained color intrinsic atoms. Since each dictionary atom undergoes column normalization, Figure 3 It reflects the relative channel response in the 20-dimensional dictionary space and should not be interpreted as the absolute spectral absorption coefficient of the ink material.

[0109] from Figure 5It can be seen that the relative absorbance distributions of different colors in the three equivalent channels R, G, and B are different. During online processing, the 20-dimensional observation features of the sample to be identified have the same feature dimensions and arrangement order as the intrinsic sub-dictionaries of each candidate color. Therefore, the intrinsic sparsity coefficient, degradation sparsity coefficient, and reconstruction cost of the candidate colors can be calculated separately, instead of directly comparing the RGB region mean with the dictionary coefficients.

[0110] To enhance the separability between similar colors, discriminative constraints can be further introduced. For the samples... Define its in the first The residuals of color eigenvalue reconstruction on the class color eigenvalue sub-dictionary are:

[0111] For the true category For the samples, the true class residuals must be smaller than the residuals of other classes, and a class margin parameter must be set. :

[0112] The corresponding discriminant loss is defined as:

[0113] Through the discriminative loss, the color intrinsic dictionary not only has reconstruction capabilities but also color category discrimination capabilities.

[0114] For the A sample of degenerate color rings The system first utilizes its true color category. The corresponding color intrinsic sub-dictionary reconstructs the color intrinsic components:

[0115] Then calculate the degenerate residuals:

[0116] in, This represents the corresponding color eigenvalue reconstruction result. This represents the degradation residual. The degradation residual is used to represent the shift of the degraded sample relative to the intrinsic characterization of the standard color. This shift can be caused by variations in ink thickness, curing degree, reflective residue, exposure drift, light source color temperature drift, or batch-to-batch differences in ink.

[0117] The degradation perturbation dictionary adopts a structure combining a shared degradation perturbation dictionary and a color-conditional degradation perturbation sub-dictionary. Used to characterize general degradation patterns, such as overall whitening, exposure drift, highlight retention, or graying; Used to characterize the The fading direction is unique to each color, such as blue turning purple, yellow turning white, red turning orange, or brown turning black.

[0118] Let the total number of degenerate samples participating in the training of the degenerate dictionary be . For degraded residuals Normalization is performed:

[0119] The shared degradation perturbation dictionary is constructed from common degradation residuals that may occur in all color categories. This is applied to the normalized degradation residuals. Clustering is performed to obtain the shared degenerate atom attribution index and sample set:

[0120] in, Indicates the first A shared degenerate clustering center, Indicates belonging to the first A set of degenerate sample indices sharing degenerate atoms. The first shared degenerate atom The elements are first determined by the first normalized degenerate residual. The mean of each feature dimension is calculated to obtain:

[0121] Then on Perform column normalization to obtain the final elements sharing the degenerate atom: any element All are derived from the degenerate residual set in the th The common offset statistics on each feature dimension are calculated to represent common degradation patterns that may occur in multiple color categories, such as whitening, graying, exposure drift, highlight retention, or changes in light source status.

[0122] Color Condition Degradation Perturbation Sub-dictionary Used to indicate the first A degradation offset specific to color classes. Construction. First, the shared degradation coefficients are solved using the shared degradation perturbation dictionary, and common degradation components are subtracted:

[0123] in, Indicates the first Color conditional residuals of degraded samples after subtracting shared degraded components. This indicates its normalization result. For the same color category... Intra-class clustering is performed on the normalized color conditional residuals:

[0124] No. The first The first color conditional degenerate atom The element is first determined by the first residual of the corresponding color condition. The mean of each feature dimension is calculated to obtain:

[0125] Then normalization yields the final elements of the color-conditionally degenerate atoms: any element All by the first The color conditional offset statistics of the degraded sample after deducting the shared degraded components are calculated and used to represent the direction of change in color intensity, such as purple, orange, white, black or thickness, which are unique to that color category.

[0126] If the degraded sample has a label indicating the cause of degradation. Then the above cluster set or Alternatively, it can be replaced or constrained by a sample set with the same degradation cause label, and the label of the corresponding degradation atom can be set. or If the same atom is composed of multiple degradation cause samples, the degradation cause with the highest sample number or the highest residual energy percentage will be used as the label for that atom.

[0127] After completing the above initialization, the degenerate perturbation dictionary can be updated using an alternating least squares approach. Let... , Then the unnormalized update matrix of the shared degenerate perturbation dictionary is:

[0128] For the Let the color conditional degradation perturbation sub-dictionary be... , ,but:

[0129] For any Degenerate perturbation dictionary update matrix The Line 1 The column elements are normalized to obtain the final dictionary elements:

[0130] in, Represents the first element in the unnormalized update matrix. Line 1 Column elements.

[0131] Preferably, for degraded residuals The training objective of the degenerate perturbation dictionary is:

[0132] If the degradation type of the degraded samples is not clearly labeled, the set of degradation residuals can be clustered. The cluster centers can be used as the initial atoms of the degradation perturbation dictionary. The degradation mode can be named as whitening, graying, yellowing, thinning, insufficient curing, exposure drift, or highlight retention based on the working condition records or image features of the clustered samples. If the degraded samples have a degradation cause label, the degradation residuals belonging to the same degradation cause can be used to initialize or train the corresponding degradation atoms.

[0133] To establish a clear correspondence between the degradation dictionary and the causes of degradation, this invention defines a set of degradation causes. And for each degenerate atom in the degenerate perturbation dictionary Assignment Degradation Reason Labels The degradation cause label can be determined by the operating condition annotation of the degraded sample, manual annotation, residual clustering category, or production line sensor record. Specifically, the degradation cause label is determined according to the following priority order: First, the process parameters directly adjusted in the controlled degradation experiment and the corresponding production line sensor records should be used as the degradation cause labels. Second, when the target parameters are not directly measured by the production line sensors, control commands recorded by inkjet equipment, curing equipment, cameras or lighting controllers are used, and the degradation cause labels are determined in conjunction with the calibration results. Third, when complete working condition records are lacking, manual annotation is performed based on the color wheel image and known production process. The manual annotation results should preferably be independently confirmed by at least two annotators. Fourth, degradation residuals are clustered only when there are no working condition annotations or manual confirmation results. The categories obtained from residual clustering are initially used as candidate degradation patterns, not directly as labels for degradation causes with definite physical meanings. After confirmation based on the working condition records, image performance, or reproduction experiments of the corresponding samples, they are then mapped to ink thinning, insufficient curing, exposure drift, highlight residue, or other degradation causes.

[0134] For single-factor controlled degradation samples, the degradation cause corresponding to the actively adjusted target process parameter is used as a single degradation cause label. For samples containing multiple controlled degradation factors, a set of degradation cause labels is established.

[0135]

[0136] in Indicates the first The number of degradation causes contained in each sample. When the first sample... The cluster sample set corresponding to each degenerate atom is: In this case, the degradation cause label of the degraded atom can be determined by weighted voting based on residual energy. For clustering patterns whose physical meaning cannot be determined through operating condition records, reproduction experiments, or manual confirmation, they are marked as unmodeled degradation or unknown degradation. The unknown degradation samples can be used for unsupervised degradation dictionary updates, but not for the initialization of atoms with determined degradation cause labels or for degradation cause identification accuracy statistics.

[0137] Combination Figure 1 The overall process shown illustrates that for any color ring sample to be identified, its absorbance feature vector... It can be composed of color intrinsic terms, degradation perturbation terms, and residual terms.

[0138] The system achieves separation between true color eigenvalues ​​and degradation perturbations through this decomposition. Color eigenvalue sparsity coefficients. Used to determine the true color category, degradation sparsity coefficient and The residual term is used to explain why the current sample deviates from the standard state. Used to absorb unmodeled noise, localized contamination, or abnormal defects.

[0139] To avoid misinterpretations of color category shifts due to degradation perturbation in the dictionary, this invention introduces a conditional coupling constraint between the intrinsic color sparsity coefficients and the degradation sparsity coefficients. This constraint states that only when the eigenvalue is... When the class color intrinsic sub-dictionary is activated, the first Only the color conditional degradation perturbation sub-dictionary is allowed to be activated.

[0140] When it is necessary to incorporate conditional coupling constraints into the optimizable loss, the following conditional coupling loss can be used:

[0141] To prevent the intrinsic color dictionary and the degenerate perturbation dictionary from substituting for each other in the feature space, this invention introduces low coherence constraints during dictionary training. In actual iterations, the following projection correction method can be used to reduce the overlap between the degenerate perturbation dictionary and the intrinsic color dictionary:

[0142] in, This indicates the low-coherence projection correction step size. After performing the projection correction, the atoms in the degenerate perturbation dictionary are re-normalized.

[0143] Furthermore, during the joint training phase, the first... The color intrinsic sparse coefficients of each sample are expanded into a global vector. The degradation perturbation sparse coefficients are expanded into a global vector. And construct the global coefficient matrix:

[0144] in, Non-zero blocks are located in the true color category or the candidate color category. Sub-dictionary position, Shared non-zero blocks are located in Position, color condition: non-zero blocks are located in the corresponding Location.

[0145] low coherence matrix The dimension is , its first Line 1 Column elements are represented as:

[0146] Therefore, the low coherence loss can be expanded element-wise as follows:

[0147] Double dictionary reconstruction error of training samples Specifically defined as:

[0148] When the training samples are decoded using candidate classes instead of true labels Can be replaced with the current candidate category Or a category determined by the minimum reconstruction cost .

[0149] Integrating color reconstruction, degradation reconstruction, sparsity constraints, conditional coupling constraints, discriminative constraints, and low coherence constraints, the joint training objective of this invention can be expressed as:

[0150] in, This represents the double dictionary reconstruction error of the training samples. Represents the set of all intrinsic sparse coefficients of colors. This represents the set of all degenerate perturbation sparse coefficients. , and These are the weight parameters for the discrimination loss, conditional coupling loss, and low coherence loss, respectively.

[0151] Combination Figure 1 The overall process shown is as follows: for new color wheel samples acquired online, the system first extracts the color wheel region to be identified. and reference area Calculate the absorbance eigenvector Then, for each candidate color category... The system solves for the cost of candidate category reconstruction. And determine the true color category.

[0152] The corrected intrinsic color characteristics are as follows:

[0153] Degradation perturbations are characterized as follows:

[0154] in, , and They are respectively true color categories The corresponding optimal color intrinsic sparsity coefficient, shared degenerate sparsity coefficient, and color conditional degenerate sparsity coefficient.

[0155] The system is based on the set of degradation causes. Degenerate atomic tags and degradation coefficient Calculate the activation score for the cause of degradation and output the main causes of degradation. If the corresponding coefficient of the whitening degradation atom is large, the output will show whitening degradation; if the corresponding coefficient of the exposure drift degradation atom is large, the output will show exposure drift; if the corresponding coefficient of the thickness thinning degradation atom is large, the output will show ink thinning; if the corresponding coefficient of the highlight residue degradation atom is large, the output will show reflective interference; if the corresponding coefficient of the insufficient curing degradation atom is large, the output will show curing abnormality.

[0156] Then the recognition confidence level was calculated. Category interval The larger the value, the more obvious the optimal candidate class is relative to other classes; the concentration of color intrinsic sparsity coefficients. The larger the value, the more likely the intrinsic reconstruction is to be explained by a small number of stable atoms; the normalized residual strength The larger the value, the more likely it is to indicate the presence of anomalous perturbations that are not explained by the current model. Therefore, this invention will... , , The minimum reconstruction cost is used together with the confidence score calculation. Below the confidence threshold When this happens, the system will mark the current sample as a suspected abnormal sample and trigger a manual review, online calibration, or rapid adaptation process.

[0157] Combination Figure 1 The dictionary training process of this invention, as shown in the figure, includes the following steps.

[0158] Step S101: Construct a training sample set. Collect standard color wheel samples and degraded color wheel samples. Standard color wheel samples are used to train the intrinsic color dictionary, and degraded color wheel samples are used to train the degradation perturbation dictionary. Degraded color wheel samples may include samples that are whitish, grayish, yellowish, purplish, have thin ink, are under-cured, have exposure shifts, or have residual highlights.

[0159] Step S102: Extract absorbance features. For each sample, extract the color ring region and reference region, calculate the reference light intensity, and calculate the absorbance features to form an absorbance feature vector.

[0160] Step S103: Initialize the color intrinsic dictionary. Group the standard samples according to color category, and then... A sample matrix is ​​built from standard samples, and the samples are normalized. Finally, intra-class clustering is performed, and initial values ​​for the color intrinsic sub-dictionary are constructed element by element. .

[0161] Step S104: Train the color intrinsic dictionary. Solve the sparse reconstruction model for each color class and iteratively update the color intrinsic sub-dictionary. At the same time, use the discriminant loss to update the color intrinsic dictionary and color intrinsic sparse coefficients, so that color samples of the same class can be stably reconstructed by a small number of intrinsic atoms, and there is a reconstruction residual interval between different color classes.

[0162] Step S105: Construct the degenerate residual. For the degenerate color wheel sample, reconstruct the color intrinsic components using the corresponding color intrinsic dictionary, and calculate the degenerate residual.

[0163] Step S106: Initialize the degenerate perturbation dictionary. Normalize the degenerate residuals and construct the initial values ​​of the shared degenerate perturbation dictionary element by element. Then construct the initial values ​​of the color conditional degradation perturbation sub-dictionary element by element. And set a degradation reason label for each degenerate atom. .

[0164] Step S107: Train the degenerate perturbation dictionary. Fix the intrinsic color dictionary and train the degenerate perturbation dictionary using the degenerate residuals, so that the degenerate residuals are reconstructed from a small number of shared degenerate atoms and a small number of color conditional degenerate atoms; update the matrix elements of the shared degenerate perturbation dictionary and the color conditional degenerate perturbation sub-dictionary in alternating optimization, and perform atom normalization after each update.

[0165] Step S108, dual-dictionary joint optimization. The following sub-steps are executed iteratively using an alternating optimization approach: [Fixed] and Solve and ;fixed and Update the intrinsic color dictionary after deducting degenerate terms; fix and After subtracting the intrinsic color terms, update the degenerate perturbation dictionary; normalize the dictionary atoms; perform low coherence correction; and perform conditional coupling constraint correction.

[0166] Step S109: Determine the convergence condition. Training stops when the decrease in total reconstruction error is less than a preset threshold, the recognition accuracy on the validation set no longer improves, the change in dictionary atoms is less than a preset threshold, or the number of iterations reaches its maximum value. The preset threshold and the maximum number of iterations can be determined based on the training sample size, online real-time requirements, and validation set performance.

[0167] When solving for sparse coefficients, orthogonal matching pursuit, LASSO, FISTA, ADMM, Group Lasso, or combinations thereof can be used. Exact sparse solutions can be used during the training phase, while lightweight approximation solutions or neural network expansion structures can be used during the online phase to meet real-time detection requirements.

[0168] Combination Figure 1 The overall process shown introduces a few-sample rapid adaptation mechanism to adapt to different ink batches, light source conditions, and production line conditions. Each ink batch, light source condition, or production line condition is defined as a task.

[0169] in, Indicates the first The support set for each task contains a small number of calibration samples; Indicates the first The query set for each task is used to evaluate the recognition performance after adaptation; This indicates the number of meta-learning tasks.

[0170] Let the set of parameters to be optimized by meta-learning be:

[0171] in, This represents the absorbance normalization parameter, reference intensity smoothing parameter, or online calibration parameter. For any batch of samples... Batch loss is defined as:

[0172] The optimization objective of meta-learning is:

[0173] in, Indicates based on support set Calculated loss, Indicates based on query set Calculated loss, This indicates that the inner loop of the meta-learning algorithm updates the step size. Represents the parameter set Find the gradient.

[0174] When a new batch is launched, the system uses a small number of calibration samples. Constrained fine-tuning of the dictionary and sparse decoding parameters. Let... This represents the sparse decoding parameters, threshold parameters, and normalization parameters other than the dictionary, with the following adaptation target:

[0175] in, and These represent the update amounts for the intrinsic color dictionary and the degeneracy perturbation dictionary, respectively. This indicates the update amount of the sparse decoding parameters, threshold parameters, or normalization parameters. , and Indicates the anti-drift regularization weights. This indicates the maximum allowed dictionary update range. After adaptation, the system updates the dictionary, threshold, and sparse decoding parameters used for online detection to enable rapid deployment under new ink batches or new production line conditions.

[0176] In one embodiment, the absorbance characteristics can be composed of RGB three channels. If an industrial color camera is used to acquire RGB three-channel images, the R, G, and B channels are respectively used as equivalent color channels, with wavelengths of... , , The absorbance eigenvector is then expressed as: In another embodiment, the absorbance characteristics may be constituted by multispectral channels, polarized dereflected images, or multi-angle imaging characteristics.

[0177] In one embodiment, the color intrinsic dictionary and the degradation perturbation dictionary are trained offline; in another embodiment, the dictionary atoms can be updated using an online incremental learning method.

[0178] In one embodiment, sparse solution employs orthogonal matching pursuit; in another embodiment, sparse solution employs LASSO, FISTA, ADMM, Group Lasso, or a neural network approximation expansion structure.

[0179] In one implementation, fast adaptation employs MAML meta-learning; in another implementation, Reptile, prototype networks, conditional normalization networks, or supernetworks can be used to achieve fast adaptation with few samples.

[0180] In one embodiment, the degradation perturbation dictionary includes a shared degradation perturbation dictionary and a color condition degradation perturbation sub-dictionary; in another embodiment, only the shared degradation perturbation dictionary may be used, or the degradation perturbation dictionary may be further subdivided into a lighting degradation dictionary, an ink thickness degradation dictionary, a curing degradation dictionary, and a reflective degradation dictionary.

[0181] In one embodiment, the cause of degradation is used only for result interpretation; in another embodiment, the cause of degradation may be fed back to the coating equipment, inkjet equipment, curing equipment, or light source controller for online process adjustment.

[0182] Specifically, the workflow of this invention is illustrated below using an online color ring inspection production line for optical fibers as an example. After inkjet printing or coating the color rings, a color industrial camera, a ring light source, and a trigger control unit are set up at the inspection station. The optical fiber continuously passes through the inspection field of view along the axial direction. The system pre-collects standard color ring samples of red, orange, yellow, green, blue, purple, brown, and black. Following the controlled degradation sample generation method described in step S101, samples with thin ink are obtained by reducing the equivalent inkjet volume; samples with insufficient curing are obtained by reducing the ultraviolet irradiation dose; samples with exposure drift are obtained by adjusting the camera exposure time or camera gain; and samples with residual highlights are obtained by adjusting the polarization state or illumination incident angle. Multiple degradation intensity levels are set for each degradation cause, and the corresponding inkjet control quantity, ultraviolet irradiation dose, camera exposure parameters, illumination parameters, and acquisition time are recorded to form a standard sample set, a single-factor degradation sample set, and a mixed degradation verification sample set.

[0183] During the training phase, the system first extracts the color ring region and adjacent transparent fiber reference region from each standard sample, calculates the absorbance features corresponding to the RGB channels or multispectral channels, and uses the absorbance features of the same real color category to train the color intrinsic sub-dictionary. Subsequently, the system reconstructs the color intrinsic components of the degraded samples using the color intrinsic sub-dictionary, and uses the difference between the absorbance features of the degraded samples and the color intrinsic reconstruction results as the degradation residual. Based on the degradation residual, a shared degradation perturbation dictionary and a color conditional degradation perturbation sub-dictionary are trained. The sparse regularization coefficient, conditional coupling parameter, low coherence weight, discriminant loss weight, degradation cause threshold, and confidence threshold are determined based on the recognition accuracy of the validation sample set, the candidate category reconstruction cost interval, the consistency of degradation cause annotation, and the online processing latency.

[0184] During the online detection phase, the system performs fiber optic centerline positioning and color ring band positioning on the newly acquired color ring image to obtain the color ring region to be identified. and reference regions in the same frame or adjacent frames For the color ring sample to be identified, the system calculates the absorbance feature vector. And respectively assume that it belongs to each candidate color category. The sparsity coefficients are solved on the corresponding color intrinsic sub-dictionary, shared degradation perturbation dictionary, and color conditional degradation perturbation sub-dictionary. If the blue color wheel appears purplish or whitish due to highlights and thin ink, the system still determines the true color category using the color intrinsic sub-dictionary with the lowest candidate category reconstruction cost, and assigns the color intrinsic terms... As the corrected color representation output.

[0185] When the degradation sparsity coefficient primarily activates specular residue atoms and thinning atoms, the system outputs the corresponding degradation reason and sends the degradation reason, true color category, and recognition confidence level to the production line control system. If the recognition confidence level of multiple consecutive samples is lower than the confidence level threshold... If the same degradation cause continues to occur within a preset time window, the system triggers an online review or a rapid adaptation process with a small number of samples. After the operator or the automatic calibration unit provides a small number of new batches of calibration samples, the rapid adaptation module updates the color intrinsic dictionary, degradation perturbation dictionary, or sparse decoding parameters in a limited manner, and uses the updated parameters for subsequent color ring detection.

[0186] To verify the operability of the aforementioned S01 to S08 processing links, a two-stage procedural verification was conducted after the scheme in this application was finalized. The first stage used simulated samples with simplified fiber and color ring geometry to check whether sample generation, ROI positioning, absorbance feature construction, dual-dictionary training, sparse decoding, intrinsic correction, and output evaluation could operate continuously. The second stage used samples with narrower color rings, low-contrast backgrounds, and feathered edges and discontinuous ink blots. Each stage generated 1248 samples, and the training, validation, and testing partitions were determined independently for each stage. Each stage's independent test set consisted of 384 samples, including 8 colors, 6 types of degradation, and 8 test samples for each color-degradation combination. The two stages were trained and evaluated separately, and the test results from the two stages were not combined into a single accuracy rate.

[0187] Combination Figure 6 As shown, in the second stage, under six types of degradation—whitening, thin ink, residual highlights, exposure drift, color temperature drift, and hue shift—the color recognition rates of the uncorrected baseline were 60.94%, 51.56%, 76.56%, 75.00%, 59.38%, and 53.12%, respectively. After using a dual-dictionary sparse decoding with both the intrinsic color dictionary and the degradation perturbation dictionary, the color recognition rates were 93.75%, 93.75%, 93.75%, 87.50%, 78.12%, and 59.38%, respectively. The overall uncorrected color recognition rate in the second stage was 62.76%, while the dual-dictionary color recognition rate was 84.38%, an improvement of 21.61 percentage points. These results support the trend that the inclusion of the degradation dictionary in decoding can improve color determination; meanwhile, the hue shift condition, at only 59.38%, indicates that the continuous mixing of adjacent color absorption fingerprints remains a weakness.

[0188] Combination Figure 7 As shown, the median observation intrinsic error in the first stage was 0.2055, and the median corrected intrinsic error was 0.0626, a relative reduction of 69.53%; in the second stage, the median observation intrinsic error was 0.1110, and the median corrected intrinsic error was 0.0623, a relative reduction of 43.85%. The intrinsic error is the dimensionless normalized distance between the observed or corrected characterization and the corresponding standard color intrinsic characterization. Both stages showed that the corrected intrinsic error was lower than the observed intrinsic error, consistent with the color intrinsic term output in step S06. This aligns with the design objectives.

[0189] The embodiments described above are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. For those skilled in the art, equivalent substitutions, parameter adjustments, module combinations, or algorithm modifications made to the above embodiments without departing from the technical concept of the present invention should all fall within the scope of protection of the present invention.

Claims

1. A method for correcting color degradation in fiber optic color rings, characterized in that, include: Acquire an optical fiber color ring image, and extract the color ring region to be identified and the reference region from the optical fiber color ring image; Based on the light intensity of the acquired light in the color ring region to be identified and the reference light intensity of the reference region, an absorbance feature vector of the color ring sample to be identified is constructed. The construction of the absorbance feature vector of the color ring sample to be identified includes: Let the first The color ring regions to be identified are: The corresponding reference area is , No. Each wavelength or equivalent color channel is Then, the light intensity of the region to be identified (color ring region) and the reference light intensity of the reference region are respectively... ， ； No. The color ring sample to be identified is in the first... The absorbance characteristics at each wavelength or equivalent color channel are as follows: ； The absorbance feature vector is ， in, Indicates the pixel position within the region. and These represent the number of pixels included in the statistics within the corresponding area. This indicates a small positive number that prevents the numerator or denominator from being zero. Indicates the first An additional statistical characteristic, ; Constructing a color intrinsic dictionary and a degradation perturbation dictionary, wherein the color intrinsic dictionary includes color intrinsic sub-dictionaries corresponding to multiple standard color categories, and the degradation perturbation dictionary includes a shared degradation perturbation dictionary and color conditional degradation perturbation sub-dictionaries corresponding to each standard color category; the construction of the color intrinsic dictionary and the degradation perturbation dictionary includes: The color intrinsic dictionary is represented as ， in, Indicates the number of standard color categories. Indicates the first The color intrinsic sub-dictionary corresponding to the class standard color; The degradation perturbation dictionary is represented as ， in, This indicates a shared degenerate perturbation dictionary. Indicates the first The sub-dictionary of color conditional degradation perturbations corresponding to the standard colors; For the true color category is The A degenerate color ring sample Solve using the corresponding color intrinsic sub-dictionary ， Obtain color eigenvalue reconstruction results ， And calculate the degradation residuals. ； The shared degradation perturbation dictionary is trained based on the degradation residuals of multiple color categories; And based on the color conditional residuals after deducting shared degradation components, the corresponding color conditional degradation perturbation sub-dictionary is trained; For each candidate color category, the absorbance feature vector is solved by a dual-dictionary sparse solution based on the color intrinsic sub-dictionary and the degradation perturbation dictionary corresponding to the candidate color category, and the absorbance feature vector is decomposed into color intrinsic terms, degradation perturbation terms and residual terms. The true color category is determined based on the reconstruction cost corresponding to each candidate color category, and the corrected color intrinsic representation is output. The degradation cause is output based on the activated degradation atom and the corresponding degradation cause label in the degradation perturbation term, and the identification confidence level is also output.

2. The fiber optic color ring color degradation correction method according to claim 1, characterized in that, The step of determining the true color category based on the reconstruction cost corresponding to each candidate color category and outputting the corrected color intrinsic representation includes: assuming that the color wheel sample to be identified belongs to the first... If the color is similar, then the absorbance feature vector satisfy ， in, Indicates the first Color-related intrinsic sparsity coefficients Indicates the shared degradation sparsity coefficient. Indicates the first Color-type conditional degradation sparsity coefficient, Represents the residual noise term; No. The reconstruction cost for each candidate color category is determined according to the following objective function: ； The true color category is according to It is determined that the corrected color intrinsic characterization is as follows: 。 3. The fiber optic color ring color degradation correction method according to claim 2, characterized in that: The dual-dictionary sparse solution introduces conditional coupling constraints, which are as follows: ， in, Indicates the first The sample at the th Degeneracy sparsity coefficients on the sub-dictionary of color conditional degradation perturbation. Indicates the first The sample at the th Color eigenvalue sparsity coefficients in the class color eigenvalue sub-dictionary. Indicates conditional coupling parameters; The construction, training, or joint optimization of the color intrinsic dictionary and the degradation perturbation dictionary introduces low coherence constraints, which include minimizing... Or reduce ， in, The first element in the intrinsic color dictionary represents the color. Each color intrinsic atom The first in the degenerate perturbation dictionary represents the first... A degenerate perturbation atom.

4. The fiber optic color ring color degradation correction method according to claim 1, characterized in that, The reasons for output degradation and the confidence level of output identification include: Define the set of causes of degradation And set a degradation cause label for each degradation atom in the degradation perturbation dictionary. ; The optimal shared degradation sparsity coefficient and the optimal color conditional degradation sparsity coefficient corresponding to the real color category are concatenated as follows: For any cause of degradation ,according to Calculate the activation score for the cause of degradation, and according to... Main reasons for output degradation; Let the minimum reconstruction cost be The cost of the second smallest reconstruction is Then the category interval Concentration of intrinsic sparsity coefficients of color and normalized reconstructed residual strength respectively ， ， ； The recognition confidence level is calculated based on the category interval, the concentration of color intrinsic sparsity coefficients, the intensity of normalized reconstruction residuals, and the minimum reconstruction cost. ， in, , This represents the confidence weight coefficient determined by the validation sample set. make As the confidence threshold, when When the time comes, the corresponding color ring sample to be identified will be marked as a suspected abnormal sample, and a manual review, online calibration or rapid adaptation process will be triggered.

5. The fiber optic color ring color degradation correction method according to claim 1, characterized in that, The fiber optic color ring color degradation correction method also includes: Different ink batches, different light source conditions, or different production line operating conditions are defined as meta-learning tasks. ,in, Indicates the first Support set for individual learning tasks Indicates the first Query set for individual learning tasks; Based on multiple meta-learning tasks, we obtain color intrinsic dictionary initialization parameters, degradation perturbation dictionary initialization parameters, and sparse decoding parameters that can be quickly adapted to new tasks. When a new batch of ink, a new light source, or a new production line is launched, the intrinsic color dictionary, the degradation perturbation dictionary, or the sparse decoding parameters are updated in a limited manner using a small number of calibration samples, and the update range is limited by anti-drift constraints.

6. A fiber optic color ring color degradation correction system, characterized in that, include: The image acquisition module is used to acquire color ring images of optical fibers; A color ring region extraction module is used to extract the color ring region to be identified and the reference region from the fiber color ring image; An absorbance feature construction module is used to construct an absorbance feature vector of the color ring sample to be identified based on the collected light intensity of the color ring region to be identified and the reference light intensity of the reference region; including: Let the first The color ring regions to be identified are: The corresponding reference area is , No. Each wavelength or equivalent color channel is Then, the light intensity of the region to be identified (color ring region) and the reference light intensity of the reference region are respectively... ， ； No. The color ring sample to be identified is in the first... The absorbance characteristics at each wavelength or equivalent color channel are as follows: ； The absorbance feature vector is ， in, Indicates the pixel position within the region. and These represent the number of pixels included in the statistics within the corresponding area. This indicates a small positive number that prevents the numerator or denominator from being zero. Indicates the first An additional statistical characteristic, ; The color intrinsic dictionary module is used to store, train, or update color intrinsic sub-dictionaries corresponding to multiple standard color categories; including: The color intrinsic dictionary is represented as ， in, Indicates the number of standard color categories. Indicates the first The color intrinsic sub-dictionary corresponding to the class standard color; The degenerate perturbation dictionary is represented as ， in, This indicates a shared degenerate perturbation dictionary. Indicates the first The sub-dictionary of color conditional degradation perturbations corresponding to the standard colors; For the true color category is The A degenerate color ring sample Solve using the corresponding color intrinsic sub-dictionary ， Obtain color eigenvalue reconstruction results ， And calculate the degradation residuals. ； The shared degradation perturbation dictionary is trained based on the degradation residuals of multiple color categories; And based on the color conditional residuals after deducting shared degradation components, the corresponding color conditional degradation perturbation sub-dictionary is trained; The degradation perturbation dictionary module is used to store, train, or update the shared degradation perturbation dictionary and the color conditional degradation perturbation sub-dictionary corresponding to each standard color category based on the degradation residual between the degradation color wheel sample and the corresponding color intrinsic reconstruction result. The dual-dictionary sparse decoding module is used to perform dual-dictionary sparse solution on the absorbance feature vector based on the color intrinsic dictionary and the degradation perturbation dictionary, and decompose the absorbance feature vector into color intrinsic terms, degradation perturbation terms and residual terms; The online recognition output module is used to determine the true color category based on the reconstruction cost corresponding to each candidate color category, and output the corrected color intrinsic representation. The degradation cause and confidence assessment module is used to output the degradation cause based on the activated degradation atoms and corresponding degradation cause labels in the degradation perturbation term, and to calculate the identification confidence based on the category interval, the concentration of color intrinsic sparse coefficients, the intensity of normalized reconstruction residuals, and the minimum reconstruction cost. The rapid adaptation module is used to update the color intrinsic dictionary, the degradation perturbation dictionary, or the sparse decoding parameters based on a small number of calibration samples when changing ink batches, light source status, or production line conditions. The process feedback interface module is used to output the degradation cause, the true color category, the identification confidence level or anomaly mark to the inkjet equipment, curing equipment, lighting controller or production line control system.

7. An electronic device, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, it implements the fiber optic color ring color degradation correction method according to any one of claims 1 to 5.

8. A non-volatile computer-readable storage medium, characterized in that, The non-volatile computer-readable storage medium stores a computer program, which, when executed by a processor, implements the fiber optic color ring color degradation correction method according to any one of claims 1 to 5.

Images