Textile defect feature library management system

Abstract

The invention provides a textile defect feature library management system, which comprises a central server cluster for storing and managing all defect feature data, and is equipped with a high-performance database management system for supporting multi-user concurrent access and real-time data synchronization; and the distributed image acquisition terminal is deployed at a key quality inspection node of a textile production line, is used for acquiring surface image data of textiles, and is provided with a high-resolution industrial camera and a self-adaptive illumination compensation module to ensure the consistency of image quality. According to the method, deep fusion and intelligent application of multi-source heterogeneous quality inspection data are realized by constructing a unified textile defect feature library management system, and quality inspection personnel can instantly position similar cases from massive historical data by providing a multi-condition combined query and content-based image retrieval function, so that the quality inspection efficiency is improved, and the quality inspection efficiency is improved. And the quality analysis efficiency is greatly improved.

Description

Technical Field

[0001] This invention relates to the field of textile production management technology, and in particular to a textile defect feature database management system. Background Technology

[0002] In the textile production process, various defects on the fabric surface are key factors affecting product quality. Traditional quality inspection mainly relies on manual visual identification, which has many drawbacks such as low efficiency, high labor intensity, strong subjectivity, and easy fatigue leading to missed or false detections. In recent years, automatic inspection equipment based on machine vision has gradually become popular, capable of capturing fabric images in real time. However, in practical applications, these systems often only complete the initial detection and alarm of defects, lacking in-depth mining and effective management of massive historical defect data. The detection results are usually stored in the form of isolated images, failing to be effectively correlated with process parameters, equipment status, environmental data, and raw material batches on the production line, forming data silos. This makes it difficult for quality management personnel to systematically analyze the generation patterns and root causes of defects, and to achieve the transformation from post-processing to pre-prevention, thus restricting further improvement of product quality and refined control of the production process.

[0003] Therefore, in order to address the above problems, a textile defect feature database management system is now being developed. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, this invention provides a textile defect feature library management system.

[0005] The technical solution of this invention is: A textile defect feature database management system includes: A central server cluster is used to store and manage all defect feature data, and is equipped with a high-performance database management system that supports concurrent access by multiple users and real-time data synchronization. Distributed image acquisition terminals are deployed at key quality inspection nodes in textile production lines to collect surface image data of textiles. They are equipped with high-resolution industrial cameras and adaptive illumination compensation modules to ensure consistent image quality. The preprocessing and feature extraction module is used to perform noise reduction, enhancement, and standardization on the acquired images, and to extract the morphology, texture, color, and distribution features of defects through a convolutional neural network algorithm. A multidimensional feature database is used to store extracted defect feature vectors and their metadata, including defect type, severity, location, production batch, equipment parameters, and environmental conditions. The intelligent classification and retrieval engine, based on a machine learning model, automatically classifies defects and matches similarities, supporting users to query through multiple conditions and perform content-based image retrieval. The defect tracing and analysis unit is used to correlate production process data, realize the tracing and analysis of the causes of defects, and generate visual reports; The user interface provides a graphical operation port, supporting quality inspectors to enter, annotate, review and correct defect data, and integrates with production management systems such as ERP and MES; The model optimization and self-learning module continuously updates the feature extraction and classification model based on user feedback and new sample data, improving recognition accuracy and system adaptability. The access control and auditing subsystem enables multi-level user role access control, records all data operation logs, and ensures data security and compliance. The real-time alarm and report generation module triggers defect warnings based on preset thresholds and automatically generates quality statistical reports and trend analysis charts.

[0006] As a preferred embodiment of the present invention, the distributed image acquisition terminal includes a multispectral imaging unit and a linear scan camera, which can acquire high-definition images of textiles under different lighting conditions, perform preliminary image correction and data compression through an embedded processor, and upload the image data to a central server cluster via industrial Ethernet or wireless transmission protocol.

[0007] As a preferred embodiment of the present invention, the preprocessing and feature extraction module uses a combination of wavelet transform and morphological operations to suppress image noise and enhance defective regions, and utilizes a deep residual network to extract multi-level features to form a high-dimensional feature descriptor.

[0008] As a preferred embodiment of the present invention, the multidimensional feature database adopts a hybrid architecture of time-series database and relational database, which supports efficient access to structured and unstructured data, and establishes an association index between defect features and production batches, machine numbers, and process parameters.

[0009] As a preferred embodiment of the present invention, the intelligent classification and retrieval engine integrates support vector machines, random forests and deep learning classifiers, and provides an approximate nearest neighbor search algorithm based on hash coding to achieve efficient matching and retrieval of large-scale feature data.

[0010] As a preferred embodiment of the present invention, the defect tracing and analysis unit uses an association analysis algorithm to mine the potential relationship between defect occurrence and equipment status, raw material properties, and environmental parameters, and uses knowledge graph technology to construct a defect cause reasoning model.

[0011] As a preferred embodiment of the present invention, the user interface provides a defect sample annotation tool that supports rectangular, polygonal and free-drawing annotation methods, and can manually correct and retrain the automatic classification results to form a human-machine collaborative optimization closed loop.

[0012] As a preferred embodiment of the present invention, the model optimization and self-learning module adopts an incremental learning mechanism, which periodically fine-tunes the existing model using newly labeled data, and evaluates the performance of different models through A / B testing, and updates the online recognition model based on the best results.

[0013] As a preferred embodiment of the present invention, the permission management and auditing subsystem implements user permission allocation based on the RBAC model and records the entire process of data query, modification and deletion operations in a log, supporting operation traceability and security auditing.

[0014] As a preferred embodiment of the present invention, the real-time alarm and report generation module can automatically trigger alarm information according to the threshold set by the user, and push early warning notifications through the web, email or mobile terminal, while regularly generating quality analysis reports for decision-making reference.

[0015] By adopting the above technical solution, the present invention has the following advantages: 1. This invention, by constructing a unified textile defect feature database management system, realizes the deep integration and intelligent application of multi-source heterogeneous quality inspection data. The multi-condition combination query and content-based image retrieval functions provided enable quality inspectors to instantly locate similar cases from massive historical data, greatly improving the efficiency of quality analysis.

[0016] 2. This invention deeply explores the potential correlation between the causes of defects and the production process through a defect source analysis unit, providing strong data support for quickly locating the root cause of the problem and optimizing process parameters, thereby significantly shortening the response and handling cycle of quality problems and effectively reducing the recurrence rate of similar defects.

[0017] 3. This invention utilizes a self-learning optimization mechanism to enable the identification model to continuously adapt to new products and new defect types, exhibiting excellent evolvability and adaptability. Its integration with the upper-level production management system achieves efficient flow and sharing of quality information within the enterprise, providing comprehensive, real-time, and visualized data support for management's scientific decision-making. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the first partial structure of the present invention.

[0019] Figure 2 This is a schematic diagram of the second partial structure of the present invention. Detailed Implementation

[0020] References to embodiments herein mean that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0021] A textile defect feature database management system, such as Figure 1 and Figure 2 As shown, it includes: The central server cluster stores and manages all defect feature data and is equipped with a high-performance database management system. Supporting concurrent access by multiple users and real-time data synchronization, it serves as the core processing and storage unit of the system. Utilizing a load-balanced architecture composed of multiple high-performance servers, it runs a distributed database system to store massive amounts of defect feature data, production batch information, equipment status logs, and user operation records. This cluster supports concurrent access from multiple clients under the TCP / IP protocol, enabling real-time data synchronization and redundant backup to ensure high system availability and data consistency.

[0022] A distributed image acquisition terminal, deployed at key quality inspection nodes in textile production lines, is used to collect surface image data of textiles. Equipped with a high-resolution industrial camera and an adaptive illumination compensation module, it ensures consistent image quality. The distributed image acquisition terminal includes a multispectral imaging unit and a linear scan camera, capable of acquiring high-definition images of textiles under different lighting conditions. Preliminary image correction and data compression are performed by an embedded processor, and the image data is uploaded to a central server cluster via industrial Ethernet or wireless transmission protocols. Specifically, it is deployed at key workstations such as fabric inspection machines and looms in textile production lines. It includes a high-resolution CCD / CMOS industrial camera, an LED linear light source or dome light source, and an embedded image processing unit. The terminal communicates with the central server via gigabit Ethernet or a 5G wireless module. When acquiring surface images of textiles, it automatically triggers light source compensation to eliminate ambient light interference and ensure image clarity and color consistency. For example, for defects such as weft skew and skipped patterns that may occur in satin fabrics, the camera captures images using line scanning or area scanning, and transmits them to the server after preliminary compression.

[0023] The preprocessing and feature extraction module is used to perform noise reduction, enhancement, and standardization on the acquired images. It also extracts the morphology, texture, color, and distribution features of defects using a convolutional neural network algorithm. The preprocessing and feature extraction module uses a combination of wavelet transform and morphological operations to suppress image noise and enhance defect areas. It also uses a deep residual network to extract multi-level features to form a high-dimensional feature descriptor. After receiving the original image, Gaussian filtering or median filtering is first performed to suppress noise. Then, histogram equalization or homomorphic filtering is used to enhance image contrast. In the feature extraction stage, a convolutional neural network (CNN, such as YOLOv4 or ResNet-50 architecture) is used to automatically learn defect features. For example, for holes, the CNN extracts typical features such as sharp edges and internal light transmission through multi-level convolution operations and outputs a feature vector. A multidimensional feature database is used to store extracted defect feature vectors and their metadata, including defect type, severity, location, production batch, equipment parameters, and environmental conditions. The multidimensional feature database adopts a hybrid architecture of time-series and relational databases, supporting efficient access to structured and unstructured data. It establishes an association index between defect features and production batches, machine numbers, and process parameters. The time-series database stores real-time collected production data (such as machine speed and tension), while the relational database stores defect metadata (such as type labels, severity levels, and occurrence time). The database establishes multi-table association indexes, for example, by querying and associating defect records with production batches via SQL.

[0024] The intelligent classification and retrieval engine, based on a machine learning model, automatically classifies defects and matches similarities. It supports users to query through multiple conditions and to retrieve images based on content. The intelligent classification and retrieval engine integrates support vector machines, random forests, and deep learning classifiers, and provides an approximate nearest neighbor search algorithm based on hash coding to achieve efficient matching and retrieval of large-scale feature data. Specifically, Support Vector Machine (SVM) classifiers are suitable for classification scenarios with small samples and high dimensionality. Their core lies in mapping the input feature vectors to a high-dimensional feature space using a kernel function, and constructing the optimal separating hyperplane within this space to classify defect categories. For linearly inseparable defect feature data, the radial basis function (RBF) kernel function is used, and its expression is: , where x i and x j Let be the feature vectors of two defective samples. γ These are kernel function parameters; The classification decision function is: ,in, αi For Lagrange multipliers, yiUsing the sample labels, the optimal solution is obtained by solving a convex quadratic programming problem, thereby classifying the defect types (such as weft skew, warp breakage, and oil stains); Random forest classifiers, as an ensemble learning method, classify data by constructing multiple decision trees and combining their voting results, effectively avoiding overfitting and exhibiting strong robustness to noisy data. During the training phase, this classifier generates multiple training subsets through Bootstrap sampling and constructs a decision tree for each subset. In the classification process, each tree independently judges the input feature vector, and the defect category is ultimately determined by majority voting. The engine integrates a deep learning classifier, primarily employing a deep convolutional neural network model. This model automatically learns hierarchical feature representations of defects through multi-layer nonlinear transformations. The network typically consists of convolutional layers, pooling layers, fully connected layers, and a classification layer. During forward propagation, feature vectors are input into the network and pass through multiple convolutional layers (using activation functions: f ( x )=max(0, x The feature transformation is performed, followed by downsampling through a max pooling layer, and finally mapped to the class probability space through a fully connected layer. The output layer uses a function to calculate the probability of belonging to each defect category, as shown in the formula: The model training minimizes the cross-entropy loss function through the backpropagation algorithm and uses an optimizer to update the network weights; To support users' multi-condition queries, the engine provides a structured query interface. Users can specify query conditions through a graphical interface, such as: defect type "hole", severity "severe", machine of occurrence "loom-5", and time range "2023-10-01 to 2023-10-31". The engine converts the query conditions into an SQL query statement, performs a joint search in the multi-dimensional feature database, and returns a set of records that match exactly. To achieve content-based image retrieval, the engine employs an approximate nearest neighbor search algorithm to address the efficiency challenges of high-dimensional feature vector similarity matching. Specifically, it utilizes Locality-Sensitive Hashing (LSH) technology. The core idea is to use a set of hash functions to map similar vectors in high-dimensional space to the same hash bucket with high probability. For a given query feature vector q, the algorithm first uses a family of hash functions... h 1, h 2,..., hL Calculate its hash value, where each hash function is a random projective hash function, in the form: .

[0025] Here, 'a' is a random Gaussian vector. bLet q be a uniformly distributed random offset. By computing multiple independent hash tables, retrieve the set of candidate vectors that fall into the same or similar hash buckets as the query vector q, and then precisely calculate the cosine similarity within this set. .

[0026] It returns the Top-K most similar defect samples and their metadata (such as the original image and production information). This method reduces the retrieval time complexity from linear to sublinear, achieving millisecond-level response for a defect feature database of tens of millions of samples.

[0027] The defect tracing and analysis unit is used to correlate production process data, realize the tracing and analysis of the causes of defects, and generate a visual report. The defect tracing and analysis unit uses correlation analysis algorithm to explore the potential relationship between defect occurrence and equipment status, raw material properties, and environmental parameters, and uses knowledge graph technology to construct a defect cause reasoning model. The user interface provides a graphical operation port, supporting quality inspectors to enter, annotate, review and correct defect data, and integrates with production management systems such as ERP and MES. The user interface provides defect sample annotation tools, supporting rectangular, polygonal and free drawing annotation methods, and can manually correct and retrain the automatic classification results to form a human-machine collaborative optimization closed loop. The model optimization and self-learning module continuously updates the feature extraction and classification model based on user feedback and new sample data to improve recognition accuracy and system adaptability. The model optimization and self-learning module adopts an incremental learning mechanism, regularly fine-tunes the existing model using new labeled data, and evaluates the performance of different models through A / B testing, selecting the best to update the online recognition model. The access control and auditing subsystem implements multi-level user role access control, records all data operation logs, and ensures data security and compliance. The access control and auditing subsystem is based on the RBAC model to allocate user permissions and records the entire process of data query, modification, and deletion operations, supporting operation traceability and security auditing. The real-time alarm and report generation module triggers defect warnings based on preset thresholds and automatically generates quality statistical reports and trend analysis charts. The real-time alarm and report generation module can automatically trigger alarm information according to user-defined thresholds and push warning notifications through the web, email or mobile terminal. At the same time, it regularly generates quality analysis reports for decision-making reference.

[0028] It should be noted that when the inspection machine or loom on the textile production line is running, the high-resolution industrial cameras deployed at its key quality inspection nodes are triggered in linear or area scan mode. Adaptive illumination compensation, using LED linear light sources or dome light sources, is employed to eliminate ambient light interference and ensure consistent image quality. After preliminary correction and compression by the embedded processing unit, the image data is uploaded in real-time to the central server cluster via gigabit industrial Ethernet or 5G wireless transmission modules. The central server cluster receives image stream data from each acquisition terminal and distributes it to the preprocessing and feature extraction module. This module first applies Gaussian or median filtering to the original image to suppress noise, followed by histogram equalization or homomorphic filtering. The wavelet algorithm enhances image contrast and detail visibility. Then, a pre-trained deep convolutional neural network model is used for feature extraction. This network automatically learns the morphology, texture, and spectral features of defects through multi-layer convolution and pooling operations, and outputs high-dimensional feature vectors. For example, for a hole defect, the network extracts its edge gradient, internal pixel distribution, and differences from the surrounding fabric to form a structured feature descriptor. The extracted feature vectors and related metadata (including image acquisition time, machine number, production line speed, tension parameters, ambient temperature and humidity, etc.) are synchronously stored in a multi-dimensional feature database. This database adopts a hybrid architecture: a time-series database stores time-correlated production parameter stream data, and a relational database stores defect features. The system collects feature vectors, type labels, severity levels, and manually labeled information, and establishes a complete index chain between defect records, production batches, and process configurations through foreign key associations. When a user initiates a query or analysis request, the intelligent classification and retrieval engine activates a multi-model processing pipeline. For automatic classification tasks, the engine calls support vector machines, random forests, and deep learning classifiers in parallel. The support vector machine uses a radial basis function kernel to map feature vectors to a high-dimensional space and solves for the optimal splitting hyperplane. The random forest constructs multiple decision trees through sampling and outputs classification results through voting. The deep learning classifier calculates the probability distribution of input features belonging to each category through forward propagation. The outputs of the three models are weighted and fused to form the final classification conclusion. For the similarity retrieval task, users can submit historical defect images or manually set multiple condition combinations (such as defect type, severity, machine, and time range). The engine first feeds the query image into the same feature extraction network to obtain its feature vector. Then, it uses the Locality Sensitive Hash algorithm to perform an approximate nearest neighbor search. By using a random projection hash function, the high-dimensional feature vector is mapped to a low-dimensional hash code, quickly retrieving similar candidate samples in the hash bucket. Then, it accurately calculates the cosine similarity between the candidate sample and the query vector, returning the Top-K results with the highest similarity and their complete metadata. Simultaneously, the defect tracing and analysis unit monitors the classification and retrieval results in real time, and analyzes the potential causal relationship between defect occurrence and production parameters based on association rule mining algorithms.For example, when the system detects a significant increase in the weft shrinkage defect rate of a certain type of loom at speeds exceeding 1200 rpm, it automatically generates knowledge graph nodes to visually display the association path between equipment parameters, raw material batches, and defect types. The user interface provides quality inspectors with a graphical entry point, supporting manual annotation of defective samples, review and correction of classification results, flexible configuration of search conditions, and a visual preview of analysis reports. All user operations are monitored in real-time through a permission management and auditing subsystem. This subsystem uses a role-based access control model to restrict data access and operation permissions for different users and records complete operation logs for auditing purposes. The model optimization and self-learning module regularly collects manually labeled and corrected results and newly collected defect samples. It updates the weight parameters of the deep learning classifier using incremental learning and minimizes the cross-entropy loss function through backpropagation to continuously optimize model classification accuracy. The real-time alarm and report generation module automatically triggers alarms based on preset threshold rules (such as consecutive occurrences of the same type of defect or defect area exceeding limits), sends notifications to designated users via API, and regularly generates multi-dimensional quality statistical reports. Ultimately, this forms a closed-loop control system from image acquisition, feature extraction, intelligent analysis to feedback optimization, comprehensively improving the automation and intelligence level of textile quality inspection.

[0029] The above embodiments are provided for those skilled in the art to implement or use the present invention. Those skilled in the art can make various modifications or changes to the above embodiments without departing from the spirit of the present invention. Therefore, the scope of protection of the present invention is not limited to the above embodiments, but should be the maximum scope that conforms to the innovative features mentioned in the claims.

Claims

1. A textile defect feature database management system, characterized in that, Including: A central server cluster is used to store and manage all defect feature data, and is equipped with a high-performance database management system that supports concurrent access by multiple users and real-time data synchronization. Distributed image acquisition terminals are deployed at key quality inspection nodes in textile production lines to collect surface image data of textiles. They are equipped with high-resolution industrial cameras and adaptive illumination compensation modules to ensure consistent image quality. The preprocessing and feature extraction module is used to perform noise reduction, enhancement, and standardization on the acquired images, and to extract the morphology, texture, color, and distribution features of defects through a convolutional neural network algorithm. A multidimensional feature database is used to store extracted defect feature vectors and their metadata, including defect type, severity, location, production batch, equipment parameters, and environmental conditions. The intelligent classification and retrieval engine, based on a machine learning model, automatically classifies defects and matches similarities, supporting users to query through multiple conditions and perform content-based image retrieval. The defect tracing and analysis unit is used to correlate production process data, realize the tracing and analysis of the causes of defects, and generate visual reports; The user interface provides a graphical operation port, supporting quality inspectors to enter, annotate, review and correct defect data, and integrates with production management systems such as ERP and MES; The model optimization and self-learning module continuously updates the feature extraction and classification model based on user feedback and new sample data, improving recognition accuracy and system adaptability. The access control and auditing subsystem enables multi-level user role access control, records all data operation logs, and ensures data security and compliance. The real-time alarm and report generation module triggers defect warnings based on preset thresholds and automatically generates quality statistical reports and trend analysis charts.

2. A textile defect feature database management system based on claim 1, characterized in that, The distributed image acquisition terminal includes a multispectral imaging unit and a linear scan camera, which can acquire high-definition images of textiles under different lighting conditions, perform preliminary image correction and data compression through an embedded processor, and upload the image data to the central server cluster via industrial Ethernet or wireless transmission protocol.

3. A textile defect feature database management system based on claim 1, characterized in that, The preprocessing and feature extraction module uses a combination of wavelet transform and morphological operations to suppress image noise and enhance defective regions, and uses a deep residual network to extract multi-level features to form a high-dimensional feature descriptor.

4. A textile defect feature database management system based on claim 1, characterized in that, The multidimensional feature database adopts a hybrid architecture of time-series database and relational database, supports efficient access to structured and unstructured data, and establishes an association index between defect features and production batches, machine numbers, and process parameters.

5. A textile defect feature database management system based on claim 1, characterized in that, The intelligent classification and retrieval engine integrates support vector machines, random forests, and deep learning classifiers, and provides an approximate nearest neighbor search algorithm based on hash coding to achieve efficient matching and retrieval of large-scale feature data.

6. A textile defect feature database management system based on claim 1, characterized in that, The defect tracing and analysis unit uses correlation analysis algorithms to uncover the potential relationships between defect occurrence and equipment status, raw material properties, and environmental parameters, and uses knowledge graph technology to construct a defect cause reasoning model.

7. A textile defect feature database management system based on claim 1, characterized in that, The user interface provides a defect sample annotation tool, which supports rectangular, polygonal and free-drawing annotation methods, and allows for manual correction and retraining of automatic classification results, forming a human-machine collaborative optimization closed loop.

8. A textile defect feature database management system based on claim 1, characterized in that, The model optimization and self-learning module adopts an incremental learning mechanism, which regularly fine-tunes the existing model using newly labeled data, and evaluates the performance of different models through A / B testing, and updates the online recognition model based on the best results.

9. A textile defect feature database management system based on claim 1, characterized in that, The permission management and auditing subsystem is based on the RBAC model to allocate user permissions and logs all data query, modification and deletion operations, supporting operation traceability and security auditing.

10. A textile defect feature database management system based on claim 1, characterized in that, The real-time alarm and report generation module can automatically trigger alarm information based on user-defined thresholds and push early warning notifications via web, email, or mobile terminal. It also regularly generates quality analysis reports for decision-making reference.

Images