A glue velvet intelligent detection method based on morphological features and spectral information

Abstract

The present application relates to the technical field of textile product quality detection, and discloses a glue velvet intelligent detection method based on morphological characteristics and spectral information, which constructs a glue velvet recognition model by fusing fiber morphological characteristic analysis and spectral information analysis, sets model evaluation indexes, optimizes model recognition accuracy, and finally realizes automatic recognition of glue velvet, effectively reduces the uncertainty of manual detection, and overcomes the problem that glue velvet can imitate the fluffy structure of normal velvet in macro morphology and is difficult to accurately distinguish by relying on appearance characteristics alone. The problem that glue in glue velvet forms local adhesion and rigid structure between fibers, destroys the natural branching morphology of fibers, and is not easy to be effectively quantified in conventional detection is solved. The detection efficiency and objectivity of detection results are greatly improved.

Description

Technical Field

[0001] This invention relates to the field of textile product quality testing technology, and in particular to an intelligent detection method for glue-based down based on morphological characteristics and spectral information. Background Technology

[0002] In the quality inspection of down products, whether glued down is mixed into the raw materials is one of the key factors affecting the quality and safety of down. Glued down is usually formed by bonding up broken down, short down or other fibers with glue, spray glue or coating adhesive. Its appearance is similar to natural down, which is highly concealed and seriously affects the loft, resilience and durability of down products, and may even endanger the health of consumers.

[0003] Currently, down inspection mainly relies on manual sorting and experience-based judgment, which is not only inefficient but also highly susceptible to human factors, resulting in inconsistent results. Furthermore, there are still no reliable methods or standards for detecting glued down. While some detection methods attempt to identify down through image analysis with the development of computer vision and optical inspection technologies, existing methods primarily focus on distinguishing down from non-down components, lacking specific detection techniques for the unique abnormal form of glued down.

[0004] Therefore, developing a method that can simultaneously utilize fiber morphological characteristics and spectral information to perform automated and high-precision detection of glued fibers is of great practical significance. Summary of the Invention

[0005] In view of the above-mentioned defects or deficiencies in the existing technology, it is desirable to provide a smart detection method for glue fibers based on morphological features and spectral information.

[0006] This invention provides a smart detection method for glue fibers based on morphological features and spectral information, comprising: S100. After processing, the down sample is placed on the testing platform; S200: Use the visible light information acquisition module to acquire visible light images of the down sample to obtain an initial image; at the same time, use the spectral information acquisition module to acquire spectral data of the down sample to obtain initial spectral data; and perform noise reduction and impurity removal processing on the initial image and initial spectral data respectively to obtain the down image and down spectral data. S300. Perform morphological feature analysis on the down images, filter out abnormal areas, slightly suspected areas, and moderately suspected areas, and mark the remaining areas as normal areas. S400: Retrieve down spectral data corresponding to the spatial location of the moderately suspected area and down spectral data corresponding to the normal area, extract the spectral features of the two, and conduct comparative analysis through a multi-feature fusion judgment strategy to screen out highly suspected areas. S500: The morphological and spectral features of highly suspected areas are combined to form a fused feature vector. The glue-textured wool recognition model is used to judge the fused feature vector and output the judgment result.

[0007] According to the technical solution provided in the embodiments of this application, step S100 specifically involves placing the down sample in a constant temperature and humidity environment for a set time, and then spreading it evenly on a detection stage with a dark background.

[0008] According to the technical solution provided in the embodiments of this application, in step S200, the noise reduction and noise removal process specifically includes image noise reduction and spectral noise removal. Image noise reduction is to perform grayscale conversion, binarization and noise removal on the initial image in sequence to obtain a down image. Spectral noise removal is to smooth, correct and normalize the initial spectral data to obtain down spectral data.

[0009] According to the technical solution provided in the embodiments of this application, step S300 specifically includes the following steps: S310. Perform connected component feature analysis on all regions of the down image, extract and statistically analyze the structural features of each connected region, and screen out several abnormal regions. S320. Perform skeleton branch feature analysis on each abnormal area to screen out several areas with mild suspicion. S330. Perform local morphological feature analysis on the slightly suspected areas to screen out several moderately suspected areas; S340. Mark abnormal areas, slightly suspected areas, and moderately suspected areas as abnormal areas, and mark the remaining areas as normal areas.

[0010] According to the technical solution provided in the embodiments of this application, step S400 specifically includes the following steps: S410. Extract down spectral data from areas with moderate suspected down and from normal areas. S420. Extract spectral features from the down spectral data corresponding to the moderately suspected area and the down spectral data of the normal area to obtain suspected spectral data and normal spectral data containing spectral features. S430. A multi-feature fusion judgment strategy is adopted to conduct a comprehensive analysis of moderately suspected areas and select several highly suspected areas.

[0011] According to the technical solution provided in the embodiments of this application, the multi-feature fusion determination strategy in step S430 specifically includes the following steps: S431. Compare the spectral features of suspected spectral data with the spectral features of normal spectral data. If one of the spectral features of suspected spectral data deviates from the normal range, then it is determined that the spectral feature of that region is abnormal. S432. If the number of spectral features in the same region exceeds the threshold, the region is determined to be a highly suspected region.

[0012] According to the technical solution provided in the embodiments of this application, the method for creating the glue lint recognition model in step S500 specifically includes the following steps: S510. Select several samples of glued down and several samples of pure natural down as down samples. S520, Execute steps S100 to S400 to obtain morphological and spectral feature data of each suspected area, as well as morphological and spectral feature data of the normal areas of glue-bonded down samples and pure natural down samples. S530. Merge morphological feature data and spectral feature data of the same region to form a fused feature vector, and divide multiple fused feature vectors into training set and test set; S540. Based on the training set data, a random forest classification model is used to construct a model for recognizing glue fibers.

[0013] According to the technical solution provided in the embodiments of this application, the following steps are further included after step S540: S550. Use the glue-slip recognition model to identify and judge the fused feature vector of the test set, and verify the performance of the glue-slip recognition model through evaluation indicators. If the verification result is qualified, proceed to step S560; otherwise, proceed to step S570. S560. Based on the successful verification in step S550, different types of down samples, adhesive types, and testing conditions are changed to form multiple combinations. Step S520 is repeated. The morphological feature data and spectral feature data obtained each time are fused to form a fused feature vector. The fused feature vector is used as a test set and returned to step S560 until all combinations are verified. Then the glue-down recognition model is verified. S570. Perform error analysis on the glue lint recognition model that failed verification in step S550, take targeted optimization measures, and return to step S540 to reconstruct and generate the glue lint recognition model.

[0014] According to the technical solution provided in the embodiments of this application, in step S550, the evaluation index specifically includes detection accuracy and F1 value. Detection accuracy is the overall correctness of the glue wool recognition model for all samples, and F1 value is the weighted harmonic average of precision and recall. Precision is the proportion of samples that are identified as glue fluff but are actually glue fluff, while recall is the proportion of samples that are actually glue fluff that are correctly detected by the glue fluff identification model.

[0015] Compared with the prior art, the beneficial effects of the present invention are: This invention constructs a glue-based fiber identification model by integrating fiber morphology analysis and spectral information analysis. By setting model evaluation indicators and optimizing the model's recognition accuracy, it ultimately achieves automated identification of glue-based fibers, effectively reducing the uncertainty of manual detection. It overcomes the problem that glue-based fibers can mimic the fluffy structure of normal fibers in macroscopic morphology, making it difficult to accurately distinguish them based solely on appearance characteristics. It also solves the problem that adhesives form local adhesions and rigid structures between fibers, disrupting the natural branching morphology of the fibers and making them difficult to quantify effectively in conventional testing. This significantly improves detection efficiency and the objectivity of test results.

[0016] Among them, fiber morphology characteristics, through connected component analysis, skeletal branch feature analysis, and local morphology feature analysis, effectively analyze and judge the parts of glued down that deviate significantly from the threshold based on the fiber morphology characteristics of normal down. It can initially screen out the areas suspected of being glued down, which can be used in combination with spectral features to improve the accuracy of judgment.

[0017] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of the present invention, nor is it intended to restrict the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0018] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a schematic diagram of the overall process of an intelligent detection method for glue fibers based on morphological features and spectral information provided in an embodiment of this application. Figure 2 A flowchart illustrating step S300 provided in an embodiment of this application; Figure 3 A flowchart illustrating step S400 provided in an embodiment of this application; Figure 4 A flowchart illustrating step S430 provided in an embodiment of this application; Figure 5 A flowchart illustrating step S500 provided in the embodiments of this application. Figure 1 ; Figure 6 A flowchart illustrating step S500 provided in the embodiments of this application. Figure 2 ; Detailed Implementation The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0019] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] Please refer to Figure 1 The present invention provides an intelligent detection method for glue fibers based on morphological features and spectral information, specifically including the following steps: S100. After processing, the down sample is placed on the testing platform.

[0021] In some embodiments, step S100 specifically involves placing the down sample in a constant temperature and humidity environment for a set time, and then evenly spreading it on a detection stage with a dark background. Optionally, the constant temperature and humidity environment refers to a temperature of 20±2℃, a relative humidity of 65%±4%, and a set time of 24 hours to ensure that the moisture content and state of the down sample remain stable.

[0022] S200: Use the visible light information acquisition module to acquire visible light images of the down sample to obtain an initial image; at the same time, use the spectral information acquisition module to acquire spectral data of the down sample to obtain initial spectral data; and perform noise reduction and impurity removal processing on the initial image and initial spectral data respectively to obtain the down image and down spectral data. The initial image is used to record the overall morphological structure of the fiber assembly, and the initial spectral data is used to reflect the optical response characteristics of the down fibers and their surface attachments.

[0023] In some embodiments, step S200 specifically includes image denoising and spectral noise removal. Image denoising involves sequentially converting the initial image to grayscale, binarizing it, and removing noise to obtain a down image. Spectral noise removal involves smoothing, correcting, and normalizing the initial spectral data to obtain down spectral data.

[0024] Image denoising methods such as grayscale conversion, binarization, and noise removal are common image processing techniques aimed at removing interfering and invalid data from the initial image and improving the accuracy of subsequent processing. Similarly, smoothing, correction, and normalization in spectral de-pollination are also common spectral processing methods, aiming to eliminate the influence of interfering factors and improve the accuracy of spectral feature identification.

[0025] S300. Perform morphological feature analysis on the down images to filter out abnormal areas, slightly suspected areas, and moderately suspected areas, and mark the remaining areas as normal areas.

[0026] In some embodiments, such as Figure 2 As shown, step S300 specifically includes the following steps: S310. Perform connected component feature analysis on all regions of the down image, extract and statistically analyze the structural features of each connected region, and screen out several abnormal regions.

[0027] The connected component feature analysis specifically includes the following three steps: First, identify each connected component in the down feather image and record the pixel area of ​​each connected component. Filter out the area of ​​the largest connected component Calculate the total area of ​​all connected components. and the percentage of the largest connected region area By setting a threshold for the area proportion, when the area of ​​the maximum connected region exceeds the threshold, it is determined that there is a risk of abnormal adhesion in that area. The principle is based on the fact that due to abnormal adhesion of fibers, glue-like fibers often form connected blocks with a large proportion. Therefore, when the connected blocks are too large, it may be glue-like fibers.

[0028] Second, the number N of connected regions and the coefficient of variation of their area distribution are statistically analyzed. By setting thresholds for the number of connected regions and the coefficient of variation, if the number of connected regions is lower than the threshold and the coefficient of variation is higher than the threshold, then the region is determined to have a risk of abnormal adhesion. Where: These represent the standard deviation and average area of ​​the connected regions, respectively. The principle is that normal down feathers typically present a large number of connected regions with a relatively uniform area distribution, while glued down, due to its adhesion, will result in a decrease in the number of connected regions and an increase in the dispersion of the area distribution.

[0029] Third, calculate the solidity of the convex hull for the connected regions. If the solidity exceeds the threshold, then the area is considered to have a risk of abnormal adhesion, where: This represents the convex hull area of ​​the connected region. The principle is that the glued down region exhibits a denser, blocky structure because the fiber gaps are filled with adhesive material, resulting in a significantly higher core density than normal down.

[0030] In summary, if the risk of abnormal adhesion is identified in a region twice or more, the region will be marked as an abnormal region, and its spatial location will be output for subsequent analysis.

[0031] S320. Perform skeleton branch feature analysis on each abnormal area to screen out several areas with mild suspicion. Among them, skeleton branch feature analysis involves extracting the fiber skeleton structure and counting the number of branch points based on the neighborhood topology of skeleton pixels. Number of endpoints and total length of skeleton Euclidean distance between endpoints The ratio represents the degree of skeletal tortuosity. If the number of branch points Number of endpoints and tortuosity If two of the three are below the threshold, it is marked as a slightly suspected area, and its spatial location is output for subsequent analysis. The principle is that normal down fibers usually exhibit a natural radial branching pattern with many branch points and free ends, and a high degree of skeletal tortuosity. However, glued down, due to the action of adhesive, makes the fiber structure rigid and causes adhesion, resulting in a reduction in the number of branch points and ends, a tendency for the skeleton to become straight, and a decrease in tortuosity.

[0032] Preferably, the fractal dimension is first obtained to quantify the complexity and naturalness of the fiber branching structure and capture the complexity differences of the fiber texture. The fractal dimension is then normalized to obtain the normalized fractal dimension, where: This is a scale parameter representing the size of the box used to cover the skeleton. In the box-counting method, This usually refers to the side length of the grid or the size of the box; In scale The minimum number of boxes required to cover the skeleton is calculated by superimposing the skeleton of the abnormal area onto the mesh and counting the number of non-empty meshes. Normal down has a high fractal dimension, while glued down has a lower fractal dimension due to the homogenization of the structure. Generally speaking, normal down fiber skeletons have rich and irregular branches, so the fractal dimension has a high complexity, between 1.6 and 1.8. Glued down has a relatively low complexity, between 1.2 and 1.4, because the adhesive causes the fibers to stick together and the structure to become rigid, resulting in a smoother skeleton.

[0033] Next, the branch point density is calculated based on the ratio between the number of branch points and the total length of the skeleton. The maximum and minimum values ​​of the branch point density are then normalized to eliminate the influence of fiber scale, resulting in a normalized branch point density. At the same time, the tortuosity is normalized to obtain a normalized tortuosity, such that a normalized tortuosity of 0 represents a straight line (corresponding to glued pile) and a normalized tortuosity of 1 represents a bend (corresponding to normal pile).

[0034] Finally, the weighted sum of the normalized fractal dimension, normalized branch density, and normalized tortuosity is used as the skeleton branch anomaly score. If the skeleton branch anomaly score is less than the score threshold, it can be identified as a mildly suspected region.

[0035] Specifically, normalized branch density reflects the richness of fiber branches, normalized tortuosity reflects the natural bending of fibers, and normalized fractal dimension characterizes the complexity of branch texture. After weighted fusion of the three, the higher the score, the closer the fiber structure is to the natural radial shape of natural down, while the lower the score, the more rigid and simple the structure. For natural down, the branches are dense, the bending is natural, and the texture is complex, so the abnormal score of the skeleton branch is relatively high. Glue-bonded down, due to the adhesive, causes fiber adhesion, branch disappearance, skeleton rigidity, and homogenization of structure, and the corresponding abnormal score of the skeleton branch is relatively low.

[0036] In this way, by quantifying the complexity of fiber structure and integrating three core features—branch density, skeleton curvature, and fractal complexity—it is possible to effectively distinguish between glued down and normal down, improve the sensitivity to subtle morphological changes, and is especially suitable for situations where glued down is highly concealed. This improves the accuracy, efficiency, and generalization ability of the overall detection model, and provides a standardized judgment basis for the intelligent and automated detection of glued down.

[0037] S330. Perform local morphological feature analysis on the slightly suspected areas to screen out several moderately suspected areas; Among them, the local morphological feature analysis examines the gray-level statistical and textural features of mildly suspected areas, specifically including local gray-level variance. Number of edge pixels With area The ratio represents the local edge density The region is identified by three local morphological features: texture contrast, which reflects the intensity of grayscale changes. If two of these three features are below a threshold, the region is marked as a moderately suspected region, and its spatial location is output for subsequent steps to retrieve the corresponding spectral information for further determination. In the formula, M represents the total number of pixels in a mildly suspected region. Let j be the grayscale value of the j-th pixel within the slightly suspected region. This represents the average grayscale value of all pixels within the slightly suspected area.

[0038] S340. Mark abnormal areas, slightly suspected areas, and moderately suspected areas as abnormal areas, and mark the remaining areas as normal areas.

[0039] Among them, the threshold settings in the above connected component feature analysis, skeleton branch feature analysis and local morphological feature analysis can be obtained by referring to the following method: first select pure natural down, take pictures of pure natural down through the visible light information acquisition module, and statistically analyze the relevant feature data of pure natural down as the thresholds in the above feature analysis process.

[0040] S400: Retrieve down spectral data corresponding to the spatial location of the moderately suspected area and the down spectral data corresponding to the normal area, extract the spectral features of the two, and conduct comparative analysis through a multi-feature fusion judgment strategy to screen out highly suspected areas.

[0041] In some embodiments, such as Figure 3 As shown, step S400 specifically includes the following steps: S410. Extract down spectral data from areas with moderate suspected down and from normal areas. S420. Extract spectral features from the down spectral data corresponding to the moderately suspected area and the down spectral data of the normal area to obtain suspected spectral data and normal spectral data containing spectral features; specifically, these include the reflection intensity at the characteristic band, the position and peak intensity of the characteristic absorption peak, the overall morphological parameters of the spectral curve, and the spectral similarity features characterizing the degree of spectral similarity.

[0042] S430. A multi-feature fusion judgment strategy is adopted to conduct a comprehensive analysis of moderately suspected areas and select several highly suspected areas.

[0043] In some embodiments, such as Figure 4 As shown, the multi-feature fusion determination strategy in step S430 specifically includes the following steps: S431. Compare the spectral features of suspected spectral data with the spectral features of normal spectral data. If one of the spectral features of suspected spectral data deviates from the normal range, then it is determined that the spectral feature of that region is abnormal. S432. If the number of spectral features within the same region exceeds a threshold, the region is determined to be a highly suspected region. Spectral features include reflection intensity at characteristic bands, the position and peak intensity of characteristic absorption peaks, overall morphological parameters of the spectral curve, and spectral similarity features characterizing the degree of spectral similarity.

[0044] By employing a multi-feature fusion judgment strategy, the risk of misjudgment caused by fluctuations in a single spectral feature is reduced, thereby improving the stability and reliability of the glue-down identification results. In addition, the threshold settings for the reflection intensity at the aforementioned feature bands, the position and peak intensity of the feature absorption peaks, the overall morphological parameters of the spectral curve, and the spectral similarity features characterizing the degree of spectral similarity can be obtained by referring to the following method: first, select pure natural down, photograph pure natural down through the spectral information acquisition module, and statistically analyze the relevant spectral feature data of pure natural down, which will be used as the thresholds in the above feature analysis process.

[0045] S500: The morphological and spectral features of highly suspected areas are combined to form a fused feature vector. The glue-textured wool recognition model is used to judge the fused feature vector and output the judgment result.

[0046] Specifically, a feature normalization stitching method is used to form a fused feature vector from the morphological and spectral features of highly suspected regions. The feature normalization stitching method includes the following steps: In step S300, all morphological feature parameters corresponding to the highly suspected region are extracted, normalized to a unified dimension, and then concatenated in a fixed order to form a morphological feature sub-vector corresponding to the highly suspected region. The unified dimension normalization process eliminates the influence of differences in units and magnitudes of different features, making all features on the same numerical scale and comparable. Common methods include Z-score standardization and Max-Min normalization. Optionally, Z-score standardization is used to convert each feature value to a mean of 0 and a standard deviation of 1. After multiple features are converted, they are arranged in order to form a morphological feature sub-vector corresponding to the highly suspected region.

[0047] Similarly, all spectral feature parameters corresponding to the highly suspected region in step S400 are extracted, normalized, and then spliced ​​in a preset order to form a spectral feature sub-vector corresponding to the highly suspected region; the normalization process is the same as above.

[0048] Finally, the morphological feature vectors and spectral feature vectors are concatenated and fused to form a fused feature vector corresponding to the same highly suspected region. Optionally, the concatenation and fusion can be vector splicing or weighted splicing to preserve the integrity of the expression of morphological and spectral information in the feature space.

[0049] In some embodiments, such as Figure 5 as well as Figure 6 As shown, in step S500, the method for creating the glue-based lint recognition model specifically includes the following steps: S510. Select several samples of glued down and several samples of pure natural down as down samples. S520, Execute steps S100 to S400 to obtain morphological and spectral feature data of each suspected area, as well as morphological and spectral feature data of the normal areas of glue-bonded down samples and pure natural down samples. S530. Merge the morphological feature data and spectral feature data of the same region to form a fused feature vector, and divide the multiple fused feature vectors into training set and test set; the specific fusion method can refer to the feature normalization and splicing method mentioned above, and will not be repeated here. S540. Based on the training set data, a random forest classification model is used to construct a model for recognizing glue fibers. In some embodiments, the following steps are further included after step S540: S541. Through manual inspection, each fusion feature vector obtained in step S530 is labeled. The labeling results include two categories: "glue samples" and "non-glue samples". "Glue samples" are labeled as positive and "non-glue samples" are labeled as negative. S542. A random forest classification model is selected as the basis for the glue-down recognition model. Several labeled training sets are fed into the random forest classification model, and the glue-down recognition model is generated after training.

[0050] The random forest classification model consists of multiple decision tree base classifiers. During training, each decision tree base classifier performs bootstrap sampling from the fused feature vectors of the training set to obtain different training subsets. At each node split, it randomly selects some features from the feature dimension for splitting to form a set of base classifiers with differences.

[0051] During the training process, by training each decision tree base classifier and integrating the output results of each base classifier, a mapping relationship between the fused feature vector and the glue-down judgment result is established, which is the initial glue-down recognition model. The specific training process is a common method for random forest classification models, which will not be elaborated here.

[0052] The judgment results of the glue fluff recognition model include: (1) the glue fluff classification of the input fused feature vector; (2) the corresponding glue fluff confidence score, wherein the confidence score is the statistical result of the output probability of each decision tree in the random forest for the glue fluff category, which is used for subsequent comprehensive judgment or classification processing.

[0053] S550. Use the glue-slip recognition model to identify and judge the fused feature vector of the test set, and verify the performance of the glue-slip recognition model through evaluation indicators. If the verification result is qualified, proceed to step S560; otherwise, proceed to step S570. S560. Based on the successful verification in step S550, different types of down samples, adhesive types, and testing conditions are changed to form multiple combinations. Step S520 is repeated. The morphological feature data and spectral feature data obtained each time are fused to form a fused feature vector. The fused feature vector is used as a test set and returned to step S560 until all combinations are verified. Then the glue-down recognition model is verified. S570. Perform error analysis on the glue lint recognition model that failed verification in step S550, take targeted optimization measures, and return to step S540 to reconstruct and generate the glue lint recognition model.

[0054] The glue-down identification model is evaluated using evaluation metrics. If it fails to meet the criteria, it is optimized and improved until it passes the evaluation. After passing the evaluation, different types of down samples, adhesive types, and testing conditions are used for evaluation. This process is repeated until all common down samples, adhesive types, and testing conditions are evaluated and meet the evaluation metrics. This ensures that the glue-down identification model can identify most down and adhesives under various testing conditions, guaranteeing its wide applicability and high accuracy.

[0055] The error analysis in step S570 specifically includes three points: 1. When misjudged samples are concentrated in a certain type of morphological or spectral features, that is, when a certain feature of a down sample belongs to glued down, but the glued down recognition model cannot recognize the feature, the following optimization measures are taken: (1.1) Adjust the judgment threshold or normalization method of the corresponding feature; (1.2) Add new morphological or spectral features as the basis for glued down recognition; (1.3) Adjust the weight ratio of morphological feature sub-vector and spectral feature sub-vector in the fused feature vector, thereby increasing the weight of the unrecognizable feature so that the glued down recognition model can recognize the feature.

[0056] 2. If the glue down recognition model has performance shortcomings in terms of precision or recall, the following optimization measures should be taken: (2.1) Adjust the hyperparameter settings of the glue down recognition model. Too high hyperparameters will cause missed judgments, that is, the recall rate is too high, while too low hyperparameters will cause incomplete training and misjudgment of normal down. Therefore, it is necessary to adjust the hyperparameters according to the precision and recall rate; (2.2) Replace or combine different types of classification models; change the decision of the random forest model to achieve the purpose of adjusting the mapping relationship, and finally achieve accurate model recognition and judgment; (2.3) Resample or adjust the sample weights of the training samples. This is because the sample data is unbalanced, which makes it easier for the glue down recognition model to recognize one or several types of glue down, but not easy to recognize other types of glue down.

[0057] 3. When misjudged samples are concentrated in specific types of flocking, adhesives, or detection conditions, the following optimization measures should be taken: Expand the training samples of the corresponding types. This is because there are few training samples for this specific combination, which leads to a low bias in the model training process. Therefore, it is necessary to expand the corresponding samples.

[0058] In some embodiments, in step S550, the evaluation metrics specifically include detection accuracy and F1 score. Detection accuracy is the overall correctness of the glue-textured wool recognition model for all samples, and F1 score is the weighted harmonic average of precision and recall. Precision is the proportion of samples that are identified as glue fluff but are actually glue fluff, while recall is the proportion of samples that are actually glue fluff that are correctly detected by the glue fluff identification model.

[0059] Among them, the detection accuracy rate ( ): This measures the overall accuracy of the model's judgments across all samples, reflecting the model's comprehensive recognition ability. The calculation formula is:

[0060] in, (True positive) is the number of samples that are actually glued fleece and are correctly identified as glued fleece by the model; (True negative) is the number of samples that are actually normal down and were correctly identified as normal down by the model; (False positives) are the number of samples that are actually normal down but were misclassified as glued down by the model; (False negatives) are the number of samples that were actually glued down but were misclassified as normal down by the model.

[0061] Precision rate ( This measures the proportion of samples that the model classifies as "glue down" but are actually glue down, focusing on controlling the risk of "misclassifying normal down as glue down." The calculation formula is as follows:

[0062] Recall rate ( This measures the proportion of samples that are actually glue-filled fibers that are correctly detected by the model, focusing on controlling the risk of "missed detection of glue-filled fibers." The calculation formula is as follows:

[0063] F1 score: The weighted harmonic average of precision and recall, balancing the model's "precision" and "recall," and serving as a core objective for model parameter optimization. The calculation formula is as follows:

[0064] The trained model is validated using a test set. If the model performance does not meet the preset requirements (accuracy ≥ 95%, F1 score ≥ 94%), features are re-selected or the model structure is adjusted, and step S540 is executed again for training until the performance requirements are met.

[0065] In the description of this specification, the terms "connection," "installation," and "fixing," etc., should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; it can be a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0066] In the description of this specification, the terms "one embodiment," "some embodiments," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0067] The above are merely preferred embodiments of this application and are not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A smart detection method for glue fibers based on morphological features and spectral information, characterized in that, include: S100. After processing, the down sample is placed on the testing platform; S200: Use the visible light information acquisition module to acquire visible light images of the down sample to obtain an initial image; at the same time, use the spectral information acquisition module to acquire spectral data of the down sample to obtain initial spectral data; and perform noise reduction and impurity removal processing on the initial image and initial spectral data respectively to obtain the down image and down spectral data. S300. Perform morphological feature analysis on the down images, filter out abnormal areas, slightly suspected areas, and moderately suspected areas, and mark the remaining areas as normal areas. S400: Retrieve down spectral data corresponding to the spatial location of the moderately suspected area and down spectral data corresponding to the normal area, extract the spectral features of the two, and conduct comparative analysis through a multi-feature fusion judgment strategy to screen out highly suspected areas. S500: The morphological and spectral features of highly suspected areas are combined to form a fused feature vector. The glue-textured wool recognition model is used to judge the fused feature vector and output the judgment result. Step S300 specifically includes the following steps: S310. Perform connected component feature analysis on all regions of the down image, extract and statistically analyze the structural features of each connected region, and screen out several abnormal regions. S320. Perform skeleton branch feature analysis on each abnormal area to screen out several areas with mild suspicion. S330. Perform local morphological feature analysis on the slightly suspected areas to screen out several moderately suspected areas; S340. Mark abnormal areas, slightly suspected areas, and moderately suspected areas as abnormal areas, and mark the remaining areas as normal areas; Step S400 specifically includes the following steps: S410. Extract down spectral data from areas with moderate suspected down and from normal areas. S420. Extract spectral features from the down spectral data corresponding to the moderately suspected area and the down spectral data of the normal area to obtain suspected spectral data and normal spectral data containing spectral features. S430. A multi-feature fusion judgment strategy is used to comprehensively analyze the moderately suspected areas and select several highly suspected areas. In step S500, the method for creating the glue-based lint recognition model specifically includes the following steps: S510. Select several samples of glued down and several samples of pure natural down as down samples. S520, Execute steps S100 to S400 to obtain morphological and spectral feature data of each suspected area, as well as morphological and spectral feature data of the normal areas of glue-bonded down samples and pure natural down samples. S530. Merge morphological feature data and spectral feature data of the same region to form a fused feature vector, and divide multiple fused feature vectors into training set and test set; S540. Based on the training set data, a random forest classification model is used to construct a model for recognizing glue fibers.

2. The intelligent detection method for glue fibers based on morphological features and spectral information according to claim 1, characterized in that, Step S100 specifically involves placing the down sample in a constant temperature and humidity environment for a set time, and then spreading it evenly on a dark-background testing platform.

3. The intelligent detection method for glue fibers based on morphological features and spectral information according to claim 1, characterized in that, In step S200, the noise reduction and noise removal process specifically includes image noise reduction and spectral noise removal. Image noise reduction involves sequentially converting the initial image to grayscale, binarizing it, and removing noise to obtain a down image. Spectral noise removal involves smoothing, correcting, and normalizing the initial spectral data to obtain down spectral data.

4. The intelligent detection method for glue fibers based on morphological features and spectral information according to claim 1, characterized in that, The multi-feature fusion determination strategy in step S430 specifically includes the following steps: S431. Compare the spectral features of suspected spectral data with the spectral features of normal spectral data. If one of the spectral features of suspected spectral data deviates from the normal range, then it is determined that the spectral feature of that region is abnormal. S432. If the number of spectral features in the same region exceeds the threshold, the region is determined to be a highly suspected region.

5. The intelligent detection method for glue fibers based on morphological features and spectral information according to claim 1, characterized in that, Step S540 is followed by the following steps: S550. Use the glue-slip recognition model to identify and judge the fused feature vector of the test set, and verify the performance of the glue-slip recognition model through evaluation indicators. If the verification result is qualified, proceed to step S560; otherwise, proceed to step S570. S560. Based on the successful verification in step S550, different types of down samples, adhesive types, and testing conditions are changed to form multiple combinations. Step S520 is repeated. The morphological feature data and spectral feature data obtained each time are fused to form a fused feature vector. The fused feature vector is used as a test set and returned to step S560 until all combinations are verified. Then the glue-down recognition model is verified. S570. Perform error analysis on the glue lint recognition model that failed verification in step S550, take targeted optimization measures, and return to step S540 to reconstruct and generate the glue lint recognition model.

6. The intelligent detection method for glue fibers based on morphological features and spectral information according to claim 5, characterized in that, In step S550, the evaluation metrics specifically include detection accuracy and F1 score. Detection accuracy is the overall correctness of the glue-textured wool recognition model for all samples, and F1 score is the weighted harmonic average of precision and recall. Precision is the proportion of samples that are identified as glue fluff but are actually glue fluff, while recall is the proportion of samples that are actually glue fluff that are correctly detected by the glue fluff identification model.

Images