A textile state detection and analysis system

Abstract

The present application provides a textile state detection and analysis system, which relates to the field of state detection. First, it acquires the surface image of the textile to be inspected and performs image preprocessing. Then, it synchronously constructs a dual-branch feature extraction network to capture the local edge features of microscopic defects and the global distribution law of the macroscopic structure respectively. Further, through cross-scale feature fusion, it correlates local details with global semantics to generate multi-scale image coding features of the surface of the textile to be inspected. Finally, it combines a pre-trained deep learning model to achieve defect localization, classification and confidence evaluation for the fused features, and automatically determines whether to send an alarm signal according to the detection results. In this way, it can make the defect detection maintain high sensitivity in a complex light environment, effectively suppress false detections and missed detections in shadow or overexposed areas, thereby significantly improving the accuracy and reliability of defect detection.

Description

Technical Field

[0001] This application relates to the field of state detection, and more particularly, to a textile state detection and analysis system in the embodiments of this application. Background Art

[0002] In the fields of textile manufacturing and quality control, textile state detection is a crucial link to ensure product quality. During the production process, textiles may be affected by various factors, such as mechanical damage, chemical contamination, or human operation errors, which may all lead to defects on the textile surface. Therefore, timely and accurate detection and identification of these defects are crucial for maintaining brand reputation and meeting market demands.

[0003] However, the dynamically changing lighting conditions in the actual production environment (including natural light fluctuations, uneven lighting from equipment, local shadows caused by fabric wrinkles, etc.) significantly interfere with the quality of the captured images, making it difficult for traditional image processing algorithms to stably extract effective features. This lighting sensitivity defect is specifically manifested as: overexposure artifacts are easily generated in strong reflection areas, details are lost in weak lighting areas, and feature confusion occurs between shadow areas and real defects. Existing technologies usually adopt fixed-threshold lighting compensation or single-scale feature extraction strategies, which can neither effectively eliminate the gradient distortion caused by non-uniform lighting nor have the ability to jointly analyze micro-texture and macro-structure features, resulting in low accuracy of defect detection in complex light environments.

[0004] Therefore, an optimized textile state detection and analysis solution is desired. Summary of the Invention

[0005] To solve the above technical problems, this application is proposed. The embodiments of this application provide a textile state detection and analysis system, which first acquires the surface image of the textile to be inspected and performs image preprocessing, and then synchronously constructs a dual-branch feature extraction network to respectively capture the local edge features of micro-defects and the global distribution law of the macro-structure; further, cross-scale feature fusion is used to associate local details with global semantics to generate multi-scale image coding features of the surface of the textile to be inspected. Finally, a pre-trained deep learning model is combined to achieve defect localization, classification, and confidence evaluation for the fused features, and automatically determine whether to send an alarm signal according to the detection result. In this way, it can make the defect detection still maintain high sensitivity in complex light environments, effectively suppress false detections and missed detections in shadow or overexposed areas, thereby significantly improving the accuracy and reliability of defect detection.

[0006] According to one aspect of this application, a textile state detection and analysis system is provided, which includes:

[0007] A textile surface image acquisition module for acquiring the surface image of the textile to be inspected collected by a camera;

[0008] A textile surface image enhancement module for performing image preprocessing on the surface image of the textile to be inspected to obtain an enhanced surface image of the textile to be inspected;

[0009] A multi-scale feature encoding module for performing multi-scale image feature encoding on the enhanced surface image of the textile to be inspected to obtain multi-scale image encoding features of the surface of the textile to be inspected. The multi-scale feature encoding module includes: a multi-scale feature extraction unit for respectively extracting local detail information and global abstract information from the enhanced surface image of the textile to be inspected to obtain local detail encoding features of the surface of the textile to be inspected and global structure encoding features of the surface of the textile to be inspected; a multi-scale feature fusion unit for dynamically fusing the local detail encoding features of the surface of the textile to be inspected and the global structure encoding features of the surface of the textile to be inspected at multiple scales to obtain the multi-scale image encoding features of the surface of the textile to be inspected;

[0010] An alarm module for obtaining a detection result based on the multi-scale image encoding features of the surface of the textile to be inspected and determining whether to issue an alarm signal.

[0011] Compared with the prior art, a textile state detection and analysis system provided by the present application first obtains the surface image of the textile to be inspected and performs image preprocessing, and then synchronously constructs a dual-branch feature extraction network to respectively capture the local edge features of microscopic defects and the global distribution law of the macroscopic structure; further, cross-scale feature fusion is used to associate local details with global semantics to generate multi-scale image encoding features of the surface of the textile to be inspected. Finally, a pre-trained deep learning model is combined to perform defect localization, classification and confidence evaluation on the fused features, and it is automatically determined whether to issue an alarm signal according to the detection result. In this way, it can be ensured that the defect detection still maintains high sensitivity in a complex light environment, effectively suppressing false detections and missed detections in shadow or overexposed areas, thereby significantly improving the accuracy and reliability of defect detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] By describing the embodiments of the present application in more detail in conjunction with the drawings, the above and other objects, features and advantages of the present application will become more obvious. The drawings are used to provide a further understanding of the embodiments of the present application, and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation to the present application. In the drawings, the same reference numerals generally represent the same components or steps.

[0013] Figure 1 FIG. is a system block diagram of a textile state detection and analysis system according to an embodiment of the present application.

[0014] Figure 2 FIG. is a schematic diagram of data flow of a textile state detection and analysis system according to an embodiment of the present application.

[0015] Figure 3 It is a block diagram of a multi-scale feature encoding module in a textile state detection and analysis system according to an embodiment of the present application.

[0016] Figure 4 It is a block diagram of a multi-scale feature fusion unit in a textile state detection and analysis system according to an embodiment of the present application.

[0017] Figure 5 It is a block diagram of a multi-scale image feature alignment and fusion sub-unit in a textile state detection and analysis system according to an embodiment of the present application.

[0018] Figure 6 It is a block diagram of a secondary sub-unit for calculating attention weights in a textile state detection and analysis system according to an embodiment of the present application. Detailed implementation manners

[0019] Various exemplary embodiments, features and aspects of the present application will be described in detail below with reference to the accompanying drawings. The same reference numerals in the drawings denote elements having the same or similar functions. Although various aspects of the embodiments are shown in the drawings, the drawings do not have to be drawn to scale unless otherwise specified.

[0020] The special term "exemplary" here means "serving as an example, an embodiment or illustrative". Any embodiment described as "exemplary" here does not have to be construed as superior or better than other embodiments.

[0021] In addition, for a better description of the present application, numerous specific details are given in the following detailed implementation manners. Those skilled in the art should understand that the present application can also be implemented without some specific details. In some instances, methods, means, elements and circuits well-known to those skilled in the art are not described in detail so as to highlight the gist of the present application.

[0022] Furthermore, the terms "first" and "second" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of such features. In the description of the present application, "a plurality" means two or more unless otherwise specifically defined.

[0023] In the process of textile manufacturing, the key to ensuring product quality lies in timely and accurately detecting and identifying surface defects caused by factors such as mechanical damage, chemical contamination, or human operation errors. However, the dynamically changing lighting conditions (such as natural light fluctuations, uneven light projection of equipment, and local shadows caused by fabric wrinkles) seriously affect the quality of image acquisition, making it difficult for traditional algorithms to stably extract effective features, thus reducing the accuracy of defect detection. Existing technologies mostly rely on fixed-threshold lighting compensation and single-scale feature extraction strategies, which cannot effectively cope with the gradient distortion caused by non-uniform lighting and lack the ability to jointly analyze microscopic texture and macroscopic structure features, resulting in poor defect detection effects in complex lighting environments.

[0024] Correspondingly, the technical concept of this application is to use deep learning-based image analysis and coding technologies. First, obtain the surface image of the textile to be inspected collected by the camera. Then, adopt a dynamic lighting correction algorithm to eliminate environmental light interference and enhance the texture details of the surface image of the textile to be inspected based on adaptive noise suppression. After that, synchronously construct a dual-branch feature extraction network to capture the local edge features of microscopic defects and the global distribution law of macroscopic structures respectively; further associate local details with global semantics through cross-scale feature fusion to generate multi-scale image coding features of the surface of the textile to be inspected. Finally, combine the pre-trained deep learning model to achieve defect localization, classification, and confidence evaluation for the fused features, and automatically determine whether to issue an alarm signal according to the detection results. This solution enables high sensitivity in defect detection in complex light environments through dynamic lighting compensation and collaborative analysis of multi-scale features, effectively suppressing false detections and missed detections in shadow or overexposed areas, thus significantly improving the accuracy and reliability of defect detection.

[0025] In view of the above technical problems, this application proposes a textile state detection and analysis system. Figure 1 It is a system block diagram of the textile state detection and analysis system according to an embodiment of this application. Figure 2 It is a schematic diagram of data flow of the textile state detection and analysis system according to an embodiment of this application. As Figure 1 and Figure 2 shown, the textile state detection and analysis system 100 according to an embodiment of this application includes: a textile surface image acquisition module 110 for obtaining the surface image of the textile to be inspected collected by the camera; a textile surface image enhancement module 120 for performing image preprocessing on the surface image of the textile to be inspected to obtain an enhanced surface image of the textile to be inspected; a multi-scale feature coding module 130 for performing multi-scale image feature coding on the enhanced surface image of the textile to be inspected to obtain multi-scale image coding features of the surface of the textile to be inspected; and an alarm module 140 for obtaining a detection result based on the multi-scale image coding features of the surface of the textile to be inspected and determining whether to issue an alarm signal.

[0026] In the above-mentioned textile state detection and analysis system 100, the textile surface image acquisition module 110 is used to obtain the surface image of the textile to be inspected collected by the camera. It should be understood that in the technical solution of the present application, to obtain the surface image of the textile to be inspected collected by the camera, a high-resolution industrial camera is used for image acquisition. Such cameras have a high pixel density and color depth, and can finely display the microscopic structural features of the textile surface, such as fiber orientation, yarn gap, and possible fine defects. At the same time, considering that the textile surface usually has a large area, the system is configured with an appropriate lens combination to meet the shooting requirements of different sizes of textiles and ensure that the entire detection area can be clearly captured. It is worth mentioning that in order to further improve the accuracy and stability of image acquisition, this module integrates an automatic focusing function. By real-time monitoring the image clarity, the system can automatically adjust the position of the camera lens according to the specific morphology of the textile surface (such as flatness, thickness difference) to ensure that the best focusing effect can be achieved for each shot. At the same time, for the application requirements on high-speed production lines, the image acquisition module supports a fast continuous shooting mode and can complete the image acquisition tasks of a large number of samples in a short time to meet the rhythm requirements of high-efficiency production.

[0027] In the above textile state detection and analysis system 100, the textile surface image enhancement module 120 is used to perform image preprocessing on the surface image of the textile to be inspected to obtain an enhanced surface image of the textile to be inspected, including: suppressing noise and correcting uneven illumination on the surface image of the textile to be inspected to obtain the enhanced surface image of the textile to be inspected. It should be understood that considering the dynamic light interference in the actual production environment (such as natural light fluctuations, uneven light projection of equipment, and shadow coverage caused by fabric wrinkles), the directly captured original images often contain a large amount of noise (such as sensor noise, motion blur) and non-uniform light distribution. For example, strong reflection areas may mask the texture features of fine defects due to overexposure, weak light areas may cause small defects to be unidentifiable due to loss of details, and shadow areas are prone to feature confusion with real defects. These problems seriously reduce the effectiveness of subsequent feature extraction, making it difficult for traditional detection algorithms to distinguish real defects from artifacts caused by light or noise, ultimately resulting in false detections or missed detections. If a fixed-threshold light compensation method is used, it is difficult to adapt to the spatial heterogeneity of complex light environments, and a single-scale noise suppression strategy may also damage the microscopic texture of the textile surface due to over-smoothing, further weakening the feature analysis ability. Based on this, the present application suppresses noise and corrects uneven illumination on the surface image of the textile to be inspected to obtain an enhanced surface image of the textile to be inspected. In particular, in a specific example of the present application, the light intensity distribution in each region of the image is analyzed in real time through a dynamic light correction algorithm, and an adaptive weight adjustment strategy is used to suppress the brightness of overexposed regions and compensate the gradient of shadow regions, thereby eliminating the gradient distortion caused by non-uniform illumination; at the same time, combined with an adaptive noise suppression technology based on texture perception, while filtering out sensor noise and motion blur, the microscopic structural details of the textile surface (such as yarn gaps, fiber orientations) are retained. This processing process not only overcomes the limitations of the fixed-threshold method but also ensures the consistency of the processing effect under different light conditions by dynamically adjusting parameters. That is, by eliminating uneven illumination and noise interference, the enhanced image can clearly present the microscopic texture (such as broken yarns, fuzz balls) and macroscopic structure (such as wrinkles, color block distribution) of the textile surface. The image quality is significantly improved, laying a reliable foundation for subsequent feature extraction and model inference.

[0028] In the above textile state detection and analysis system 100, the multi-scale feature encoding module 130 is configured to perform multi-scale image feature encoding on the enhanced surface image of the textile to be inspected to obtain the multi-scale image encoding features of the surface of the textile to be inspected. It should be understood that considering the variety of defects on the textile surface, there are both local defects such as broken yarns and fuzz balls at the microscopic level, and structural abnormalities such as wrinkles and uneven color block distribution at the macroscopic level. Single-scale feature extraction methods often have difficulty in comprehensively describing these complex defect types and their spatial distribution laws. By performing multi-scale image feature encoding on the enhanced surface image of the textile to be inspected to obtain the multi-scale image encoding features of the surface of the textile to be inspected, multi-level information in the image can be captured, improving the accuracy and robustness of defect detection.

[0029] Figure 3 It is a block diagram of the multi-scale feature encoding module in the textile state detection and analysis system according to an embodiment of the present application. As Figure 3 shown, the multi-scale feature encoding module 130 includes: a multi-scale feature extraction unit 131, configured to extract local detail information and global abstract information from the enhanced surface image of the textile to be inspected respectively to obtain the local detail encoding features of the surface of the textile to be inspected and the global structure encoding features of the surface of the textile to be inspected; a multi-scale feature fusion unit 132, configured to perform dynamic feature structure multi-scale fusion on the local detail encoding features of the surface of the textile to be inspected and the global structure encoding features of the surface of the textile to be inspected to obtain the multi-scale image encoding features of the surface of the textile to be inspected.

[0030] In the embodiments of the present application, the multi-scale feature extraction unit 131 is configured to: use a multi-scale feature extractor based on the atrous spatial pyramid pooling model to extract local detail information and global abstract information from the enhanced surface image of the textile to be inspected respectively, so as to obtain a local detail encoded feature map of the textile surface to be inspected as the local detail encoded feature of the textile surface to be inspected and a global structure encoded feature map of the textile surface to be inspected as the global structure encoded feature of the textile surface to be inspected. It should be understood that considering that textile surface defects (such as broken yarns, lint balls, and fine stains) usually have the characteristics of small size and complex texture, their edge and morphological features may only occupy the width of several pixels in the image and are easily visually confused with normal textile textures (such as yarn interweaving structures and fiber orientations). Especially in the enhanced image after dynamic light correction, although the problems of noise and uneven illumination are alleviated, the local features of small defects still need to be accurately characterized by high-sensitivity feature extraction techniques. At the same time, considering that textile surface defects not only include microscopic local defects such as broken yarns and lint balls, but may also have macroscopic structural abnormalities such as large-scale wrinkles, uneven color block distribution, or pattern misalignment, the identification of these problems depends on the understanding of the overall semantics of the image. For example, large-area color difference defects need to be judged in combination with the color distribution law of the entire image. Therefore, the present application extracts local detail information and global abstract information from the enhanced surface image of the textile to be inspected respectively to obtain the local detail encoded feature of the textile surface to be inspected and the global structure encoded feature of the textile surface to be inspected. In particular, in a specific example of the present application, a multi-scale feature extractor based on the atrous spatial pyramid pooling model is used to extract local detail information and global abstract information from the enhanced surface image of the textile to be inspected respectively, so as to obtain a local detail encoded feature map of the textile surface to be inspected as the local detail encoded feature of the textile surface to be inspected and a global structure encoded feature map of the textile surface to be inspected as the global structure encoded feature of the textile surface to be inspected. Specifically, the model synchronously captures the high-frequency edge features of small defects (such as using small dilation rate convolutions to focus on pixel-level gradient changes) and the context-related features of large-scale defects (such as large dilation rate convolutions covering a wider area to identify continuous defects) without reducing the spatial resolution by deploying multiple atrous convolution layers with different dilation rates on the shared input feature map. For example, for the common fine lint ball defects on the textile surface, small dilation rate convolutions can accurately capture their local clump-like gradient distributions; while for long strip-like scratches spanning multiple yarn cycles, large dilation rate convolutions can effectively integrate their linear extension patterns, so as to improve the sensitivity and characterization ability for multi-scale defects.

[0031] Specifically, the multi-scale feature fusion unit 132 is configured to perform dynamic feature structure multi-scale fusion on the local detail encoded feature of the surface of the textile to be inspected and the global structure encoded feature of the surface of the textile to be inspected to obtain the multi-scale image encoded feature of the surface of the textile to be inspected. It should be understood that in the textile state detection and analysis system, the defect detection on the surface of the textile needs to simultaneously consider the collaborative analysis of microscopic local anomalies (such as broken yarns, fuzz balls) and macroscopic structural distortions (such as wrinkles, color differences). However, due to the misalignment of the spatial geometric structures of local features and global features (such as local texture deformation caused by fabric wrinkles) and the disconnection of semantic information (such as the disconnection between local color differences and global color distribution rules) in traditional methods, false judgments are likely to occur. For example, local detail features may accurately capture the sharp edges of tiny broken yarns, but it is impossible to determine whether they are artifacts under the folds; although global features can identify the wavy distortion of large-scale folds, it is difficult to locate the exact position of tiny defects. The traditional simple feature splicing or weighted average strategy can neither solve the problem of spatial alignment of cross-scale features nor has the ability to model the complex interaction relationships between features, resulting in redundant feature expressions after fusion and sensitivity to light interference. Based on this, the present application performs dynamic feature structure multi-scale fusion on the local detail encoded feature of the surface of the textile to be inspected and the global structure encoded feature of the surface of the textile to be inspected to obtain the multi-scale image encoded feature of the surface of the textile to be inspected. In particular, this method decouples the local and global feature maps into multiple groups of local encoding matrices, uses the feature phase similarity metric to guide the dynamic search alignment process, and finally generates the multi-scale image encoded feature through attention-driven significant aggregation.

[0032] Figure 4 It is a block diagram of a multi-scale feature fusion unit in a textile state detection and analysis system according to an embodiment of the present application. As Figure 4 shown, in the embodiment of the present application, the multi-scale feature fusion unit 132 includes: a feature decoupling sub-unit 1321 of the textile to be inspected, configured to perform feature decoupling on the local detail encoded feature map of the surface of the textile to be inspected and the global structure encoded feature map of the surface of the textile to be inspected to obtain a set of local detail feature encoding matrices of the surface of the textile to be inspected and a set of global structure local feature encoding matrices of the surface of the textile to be inspected; a multi-scale image feature alignment and fusion sub-unit 1322, configured to perform phase alignment and attention significant fusion on the set of local detail feature encoding matrices of the surface of the textile to be inspected and the set of global structure local feature encoding matrices of the surface of the textile to be inspected to obtain a multi-scale image encoded feature map of the surface of the textile to be inspected as the multi-scale image encoded feature of the surface of the textile to be inspected.

[0033] Specifically, the textile to be inspected feature decoupling subunit 1321 is used to decouple the local detail coding feature map on the surface of the textile to be inspected and the global structure coding feature map on the surface of the textile to be inspected to obtain a set of local detail feature coding matrices on the surface of the textile to be inspected and a set of global structure local feature coding matrices on the surface of the textile to be inspected, which is expressed by the textile to be inspected feature decoupling formula as:

[0034]

[0035]

[0036] Among them, is the local detail coding feature map on the surface of the textile to be inspected, is the global structure coding feature map on the surface of the textile to be inspected, is for the feature decoupling operation, , , and are respectively the 1st, 2nd, th, and th local detail feature coding matrices in the set of local detail feature coding matrices on the surface of the textile to be inspected, , , and are respectively the 1st, 2nd, th, and A local feature encoding matrix for the global structure of the surface of the textile to be inspected. It should be understood that considering the inspection of the textile surface requires simultaneous analysis of microscopic local defects (such as broken yarns, fuzz balls) and macroscopic structural anomalies (such as wrinkles, color differences), but there are significant differences in the spatial geometric structure between local detail features (focusing on pixel-level high-frequency information) and global structural features (representing long-range semantic associations): local features may cause spatial deformation due to fabric wrinkles (such as the distortion of yarn orientation), while global features may lose the precise position of tiny defects due to downsampling. If directly processing high-dimensional features, problems such as high computational complexity, semantic confusion caused by the misalignment of local-global feature spaces (such as misjudging the local break in the wrinkle shadow as a defect), and difficulty in adapting to non-uniform deformations caused by dynamic lighting due to the overly strong integrity of the feature map will be faced. Based on this, feature decoupling is performed on the local detail encoding feature map of the surface of the textile to be inspected and the global structure encoding feature map of the surface of the textile to be inspected. In particular, feature decoupling uses a structured decomposition technique to transform global features into locally semantic units that can be independently operated, which not only reduces the computational complexity of high-dimensional features but also provides an operation basis with granularity adaptation for subsequent cross-scale feature alignment by retaining the spatial correlation of local-global features. This decoupling is not a simple segmentation. Substantially, it constructs an intermediate representation form of local and global features, enabling the subsequent fusion stage to more accurately associate the local anomalies of microscopic defects with the overall semantics of the macroscopic structure, thereby enhancing the robustness of the detection system under complex working conditions.

[0037] Figure 5 As shown in the block diagram of the multi-scale image feature alignment and fusion sub-unit in the textile state detection and analysis system according to an embodiment of the present application. Figure 5As shown, in the embodiment of the present application, the multi-scale image feature alignment and fusion subunit 1322 includes: a feature phase dynamic search and alignment secondary subunit 1322-1, configured to perform feature phase dynamic search and alignment on the set of local detail feature encoding matrices of the surface of the textile to be inspected and the set of local global structure local feature encoding matrices of the surface of the textile to be inspected based on the feature phase alignment degree between any two local detail feature encoding matrices of the surface of the textile to be inspected and the local global structure local feature encoding matrix of the surface of the textile to be inspected in the set of local detail feature encoding matrices of the surface of the textile to be inspected and the set of local global structure local feature encoding matrices of the surface of the textile to be inspected, so as to obtain a set of phase-aligned {local detail feature encoding matrix of the surface of the textile to be inspected, local global structure local feature encoding matrix of the surface of the textile to be inspected} feature pairs; an attention weight calculation secondary subunit 1322-2, configured to input each phase-aligned {local detail feature encoding matrix of the surface of the textile to be inspected, local global structure local feature encoding matrix of the surface of the textile to be inspected} feature pair in the set of phase-aligned {local detail feature encoding matrix of the surface of the textile to be inspected, local global structure local feature encoding matrix of the surface of the textile to be inspected} feature pairs into a feature joint perception attention network to obtain a set of attention weights for joint perception of the surface features of the textile to be inspected; a multi-scale saliency aggregation secondary subunit 1322-3 of the surface of the textile to be inspected, configured to perform attention-driven saliency aggregation on the set of phase-aligned {local detail feature encoding matrix of the surface of the textile to be inspected, local global structure local feature encoding matrix of the surface of the textile to be inspected} feature pairs based on the set of attention weights for joint perception of the surface features of the textile to be inspected to obtain the multi-scale image encoding feature map of the surface of the textile to be inspected.

[0038] Specifically, the feature phase dynamic search and alignment secondary subunit 1322-1 is configured to perform feature phase dynamic search and alignment on the set of local detail feature encoding matrices of the surface of the textile to be inspected and the set of local global structure local feature encoding matrices of the surface of the textile to be inspected based on the feature phase alignment degree between any two local detail feature encoding matrices of the surface of the textile to be inspected and the local global structure local feature encoding matrix of the surface of the textile to be inspected in the set of local detail feature encoding matrices of the surface of the textile to be inspected and the set of local global structure local feature encoding matrices of the surface of the textile to be inspected, so as to obtain a set of phase-aligned {local detail feature encoding matrix of the surface of the textile to be inspected, local global structure local feature encoding matrix of the surface of the textile to be inspected} feature pairs, which is represented by the phase dynamic search and alignment formula as:

[0039]

[0040]

[0041] Wherein, is the transpose operation, To calculate the Frobenius norm, is and the characteristic phase alignment degree between To return the value corresponding to the maximum value, is to find the position of the maximum approximate matching value in the set of local feature encoding matrices of the global structure of the textile to be inspected. It should be understood that due to problems such as wrinkle deformation and uneven dynamic illumination during the processing of textiles, there are significant non-rigid deformations and phase offsets between local detail features (such as the sharp edges of broken yarns) and global structure features (such as the wavy distribution of wrinkles) at the spatial geometry level. Traditional feature alignment methods based on physical position matching are vulnerable to such deformations, resulting in the failure of cross-scale feature correlation and thus causing false detections or missed detections. For example, local broken yarn features may show offsets in the global structure due to local stretching of the fabric. If coordinate alignment is directly relied on, their semantic association will be severed. Therefore, by quantifying the structural similarity in the feature space (rather than physical position correspondence), a sliding window is used to traverse the set of local detail feature encoding matrices of the textile to be inspected and the set of local feature encoding matrices of the global structure of the textile to be inspected, and the F-norm maximization matching mechanism is used to dynamically search for the best alignment relationship, establishing a high-confidence cross-scale semantic association. This similarity measure based on matrix norm can capture the essential association between local and global features in the texture gradient distribution, rather than simply relying on pixel-level position matching, thus effectively overcoming the spatial misalignment caused by fabric deformation and illumination interference. In this way, a high-confidence cross-scale semantic association is established through the dynamic search mechanism to adaptively match the best feature pairs. Moreover, through the global sensitivity of the F-norm to the global feature distribution, the dynamic search mechanism driven by the maximum value of the F-norm enhances the robustness of the system to local deformations and illumination changes, overcomes the spatial misalignment caused by fabric deformation and illumination interference, and provides a high-quality multi-scale feature alignment basis for the subsequent fusion stage.

[0042] Figure 6 is a block diagram of the attention weight calculation secondary subunit in the textile state detection and analysis system according to an embodiment of the present application. As Figure 6As shown, in the embodiment of the present application, the attention weight calculation secondary subunit 1322-2 includes: a to-be-inspected textile surface pair-scale correlation calculation tertiary subunit 1322-21, configured to calculate the correlation matrix between each phase-aligned {local detail feature encoding matrix of the to-be-inspected textile surface, local feature encoding matrix of the global structure of the to-be-inspected textile surface} feature pair in the set of phase-aligned {local detail feature encoding matrix of the to-be-inspected textile surface, local feature encoding matrix of the global structure of the to-be-inspected textile surface} feature pairs to obtain a set of to-be-inspected textile surface pair-scale correlation matrices; a to-be-inspected textile surface trace metric tertiary subunit 1322-22, configured to perform a trace metric on each to-be-inspected textile surface pair-scale correlation matrix in the set of to-be-inspected textile surface pair-scale correlation matrices to obtain a set of to-be-inspected textile surface trace metric values; and a trace metric value weight normalization tertiary subunit 1322-23, configured to obtain a set of attention weights for joint perception of the to-be-inspected textile surface features based on the set of to-be-inspected textile surface trace metric values. It should be understood that although the semantic association between local details and global structure features is established through the phase dynamic search alignment mechanism, there are significant differences in the contribution degrees of different feature pairs in the actual scenario. For example, some feature pairs may contain complementary information of key defects (such as the high-frequency gradient of broken yarns and the association with the global yarn alignment offset), while others may introduce misleading signals due to light residue noise or wrinkle textures (such as the local-global feature pairs in the wrinkled shadow area). Traditional methods use fixed weights or simple averaging strategies, which are difficult to distinguish effective features from noise, resulting in the fusion result being sensitive to environmental interference. In the technical solution of the present application, by introducing a joint feature perception attention network, dynamic weight allocation is performed on the feature pairs after phase alignment to solve the problem of redundant interference still existing after cross-scale feature alignment. That is, the present application deeply mines the complex interaction patterns between phase-aligned feature pairs through the joint feature perception attention network, learns to dynamically allocate weights, suppresses the pseudo-feature pairs generated by interference such as shadows and wrinkles, and highlights the cross-scale feature interaction patterns of real defects. Its essence is to learn the complex interaction patterns between feature pairs through a deep learning model, identify which cross-scale feature associations can effectively represent real defects (such as the strong association between local gradient aggregation in the fuzz ball area and abnormal global yarn density), and which associations only reflect environmental noise (such as the artifacts in the wrinkled shadow area).

[0043] Specifically, the processing processes of the to-be-inspected textile surface pair-scale correlation calculation tertiary subunit 1322-21 and the to-be-inspected textile surface trace metric tertiary subunit 1322-22 are expressed by the following formula:

[0044]

[0045] Wherein, is the trace metric value of the matrix, is a phase alignment {local detail feature coding matrix of the surface of the textile to be inspected, local feature coding matrix of the global structure of the surface of the textile to be inspected} feature pair, is and the trace metric value of the surface of the textile to be inspected between.

[0046] Specifically, the trace metric value weight normalization three-level subunit 1322-23 is used for: based on the set of trace metric values of the surface of the textile to be inspected, performing matrix manifold optimization based on the number of biconnected components on each phase alignment {local detail feature coding matrix of the surface of the textile to be inspected, local feature coding matrix of the global structure of the surface of the textile to be inspected} feature pair in the set of phase alignment {local detail feature coding matrix of the surface of the textile to be inspected, local feature coding matrix of the global structure of the surface of the textile to be inspected} feature pairs to obtain a set of phase-optimized alignment {local detail feature coding matrix of the surface of the textile to be inspected, local feature coding matrix of the global structure of the surface of the textile to be inspected} feature pairs; based on the set of phase-optimized alignment {local detail feature coding matrix of the surface of the textile to be inspected, local feature coding matrix of the global structure of the surface of the textile to be inspected} feature pairs, obtaining a set of joint perception attention weights of the surface features of the textile to be inspected.

[0047] More specifically, based on the set of trace metric values of the surface of the textile to be inspected, performing matrix manifold optimization based on the number of biconnected components on each phase alignment {local detail feature coding matrix of the surface of the textile to be inspected, local feature coding matrix of the global structure of the surface of the textile to be inspected} feature pair in the set of phase alignment {local detail feature coding matrix of the surface of the textile to be inspected, local feature coding matrix of the global structure of the surface of the textile to be inspected} feature pairs to obtain a set of phase-optimized alignment {local detail feature coding matrix of the surface of the textile to be inspected, local feature coding matrix of the global structure of the surface of the textile to be inspected} feature pairs, and the specific processing process is as follows:

[0048] In particular, here, for each and for which matrix trace metric operations are performed, let and , and , , by calculating the number of eigenvalues that satisfy the corresponding and between the distance and distance to obtain the number of biconnected components and between and , that is, the distribution map smoothness measure between the local detail feature coding matrix of the textile to be inspected for phase alignment and the local feature coding matrix of the global structure of the textile to be inspected.

[0049] Then, in the discrete manifold representations of the local detail feature coding matrix of the textile to be inspected and the local feature coding matrix of the global structure of the textile to be inspected, embedding optimization is performed on the trace metric operations of the local detail feature coding matrix of the textile to be inspected and the local feature coding matrix of the global structure of the textile to be inspected through embedding manifold compactification based on the number of biconnectivity, so as to improve the calculation accuracy of the joint matrix trace metric on the basis of enhancing the long-range correlation connectivity robustness:

[0050]

[0051]

[0052] That is, based on the initial and calculate to obtain , and In the case of substituting , , , the initial and for optimization, and then calculate Calculate the optimized trace metric.

[0053] More specifically, based on the set of feature pairs of the phase-optimized alignment {local detail feature coding matrix of the textile to be inspected, local feature coding matrix of the global structure of the textile to be inspected}, the set of joint perception attention weights of the surface features of the textile to be inspected is obtained, which is expressed by the formula:

[0054]

[0055] where is the calculation of the joint perception attention weight, is the feature pair of the phase-optimized alignment {local detail feature coding matrix of the textile to be inspected, local feature coding matrix of the global structure of the textile to be inspected}, is the optimized phase-optimized alignment local detail feature coding matrix of the textile to be inspected, is the optimized phase-optimized alignment local feature coding matrix of the global structure of the textile to be inspected, is and the optimized surface trace metric value of the textile to be inspected is a normalization function is and the joint perception attention weight of the surface features of the textile to be inspected

[0056] Specifically, the multi-scale significant aggregation secondary subunit 1322-3 of the surface of the textile to be inspected is used to perform attention-driven significant aggregation on the set of the {local detail feature encoding matrix of the surface of the textile to be inspected, local feature encoding matrix of the global structure of the surface of the textile to be inspected} feature pairs of the phase alignment based on the set of the joint perception attention weights of the surface features of the textile to be inspected to obtain the multi-scale image encoding feature map of the surface of the textile to be inspected, which is expressed by the significant aggregation formula as:

[0057]

[0058] wherein and are the first joint perception weight matrix and the second joint perception weight matrix is subtraction by position points is addition by position points is and the joint perception matrix of the surface features of the textile to be inspected after combination, that is, the th joint perception matrix of the surface features of the textile to be inspected in the set of the joint perception matrices of the surface features of the textile to be inspected , and are respectively the 1st, 2nd, and the th joint perception matrices of the surface features of the textile to be inspected in the set of the joint perception matrices of the surface features of the textile to be inspected It is the multi-scale image coding feature map of the surface of the textile to be inspected. It should be understood that in the technical solution of this application, by introducing an attention-driven significant aggregation mechanism, the weight distribution of different feature pairs is dynamically adjusted. Its essence is to learn the complex interaction patterns between feature pairs through a deep learning model, identify which cross-scale feature associations can effectively represent real defects (such as the strong association between local gradient aggregation in the fuzz ball area and global yarn density anomaly), and which associations only reflect environmental noise (such as artifacts in the wrinkle shadow area). Specifically, the system multiplies the feature pair {local detail feature coding matrix of the surface of the textile to be inspected, local feature coding matrix of the global structure of the surface of the textile to be inspected} after phase alignment (such as the local gradient matrix representing the broken yarn defect and the global matrix reflecting the overall deviation of the yarn direction) with its corresponding attention weight to strengthen the expression of high-weight feature pairs (such as the cross-scale collaborative signal of real defects) and suppress low-weight feature pairs (such as the local-global artifact association in the wrinkle shadow area). Among them, for high-weight feature pairs (such as the local and global collaborative features of real defects), the collaborative effect is amplified through position-wise addition operation to strengthen the cross-scale representation of defects; for low-weight feature pairs (such as interference signals in the wrinkle area), the negative impact is suppressed through position-wise subtraction operation. This mechanism enables the system to adaptively filter out interference signals caused by uneven illumination or fabric deformation while retaining key defect features, and finally generate a multi-scale image coding feature map of the surface of the textile to be inspected with high discrimination. In this way, the adaptability of the system to complex working conditions is significantly improved. For example, when detecting overexposed areas, the attention network assigns lower weights to the feature pairs in this area, reduces the interference of distorted local features on the fusion result, and at the same time enhances the weights of adjacent unaffected areas, and uses the integrity of the global structure features to compensate for the lack of local information; for real defects, the cross-scale representation of key defects is adaptively highlighted by focusing on the collaborative expression of local anomalies and global semantics through high weights (such as the strong correlation between local gradient aggregation in the fuzz ball area and abnormal global yarn density distribution), so as to maintain high sensitivity and accuracy under complex illumination and deformation conditions.

[0059] In the above textile state detection and analysis system 100, the alarm module 140 is configured to obtain a detection result based on the multi-scale image coding features of the surface of the textile to be inspected and determine whether to issue an alarm signal, and is used for: inputting the multi-scale image coding feature map of the surface of the textile to be inspected into a trained defect detection model to obtain a detection result; and determining whether to issue an alarm signal based on the detection result.

[0060] Specifically, the multi-scale image coding feature map of the surface of the textile to be inspected is input into the trained defect detection model to obtain the detection result. Among them, in the embodiments of the present application, the detection result includes the defect position, defect type, and confidence level. That is, the multi-scale image coding feature map of the surface of the textile to be inspected, which is obtained by fusing the local detail coding feature map and the global structure coding feature map of the surface of the textile to be inspected, is used for detection to obtain the detection result. In this way, the defect position can be accurate to the pixel level (such as the sub-pixel coordinates of the hole boundary), the defect type can achieve fine-grained classification based on the semantic association of cross-scale features (such as distinguishing broken yarn caused by mechanical damage from color spots formed by chemical pollution), and the confidence level reflects the credibility of the detection result through the probability distribution output by the model (such as a high confidence level indicates a clear defect, and a low confidence level prompts manual re-inspection). It is worth mentioning that in the technical solution of the present application, the trained defect detection model is constructed based on deep learning technology, aiming to achieve efficient and accurate identification of textile surface defects, involving multiple key steps, including data input, model design and training, data augmentation, model optimization, and final data output. Specifically, first, in the data input stage, the feature map data stream that fuses multi-scale features and a high-quality defect annotation data set are integrated. These data sets contain defect samples and defect-free samples under various types, morphologies, and lighting conditions, ensuring that the model can comprehensively cover various situations that may be encountered in actual production. In this way, the model can not only learn the features of different types of defects but also adapt to different lighting and background conditions, thereby improving its robustness in complex environments. Next, in the model design and training stage, the selected object detection model is usually an object detection framework such as YOLOv5 or Faster R-CNN. These frameworks can directly locate and identify the defect position and category in the image, with high detection accuracy and speed. In addition, semantic segmentation models such as U-Net or DeepLabv3+ are also used to accurately segment the pixel-level area of the defect, further refining the detection result to ensure that each defect can be accurately captured. To improve the generalization ability and robustness of the model, data augmentation techniques are further used. By operations such as image rotation, flipping, scaling, cropping, and lighting transformation, a large number of variant images can be generated to expand the training data set. This approach not only increases the amount of training samples for the model but also enables the model to maintain stable performance when facing unknown or changing environments. Next, appropriate loss functions (such as Focal Loss, Dice Loss) and optimizers (such as AdamW) are used, and through model parameter tuning and optimization, the model parameters are continuously adjusted to achieve the best performance. For example, Focal Loss can effectively solve the problem of imbalance between positive and negative samples, while Dice Loss helps to improve the model's detection ability for small targets.Meanwhile, the inference speed of the model is accelerated by hardware such as GPUs to ensure the efficient progress of real-time detection. In this way, a high-performance defect detection model is finally obtained. This model can collect and preprocess the image data stream in real time, input it into the trained model for real-time defect identification and classification. The inference speed of the model is accelerated by hardware such as GPUs to meet the real-time requirements of high-speed production lines. The results output by the model include the position coordinates, category labels, and confidence information of the defects, and this information will be used in subsequent processes.

[0061] Specifically, based on the detection results, it is determined whether to issue an alarm signal. It should be understood that by monitoring the defect detection results in real time, once the position and type of the defect are detected and a certain confidence level is reached, it indicates that there is a problem with the textile, and the system immediately issues an alarm signal. This can remind the on-site operators in the first time, enabling them to quickly check and debug the production equipment, correct the abnormalities in the production process in time, and control the generation of defective textiles within the minimum range. For example, when the system identifies a broken yarn defect located in the central area of the fabric, with specific coordinates from (X1, Y1) to (X2, Y2), and the type is identified as "broken yarn caused by mechanical damage" with a confidence level as high as 95%. Based on the above detection results, the system will comprehensively consider the position, type, and confidence level of the defect to make a final decision on whether to issue an alarm signal. Among them, since this defect is located in the core visual area of the fabric and belongs to a key type that affects product quality, the system decides to trigger the alarm mechanism. The alarm signal is first conveyed to the operator through the audible and visual alarm installed on-site. The alarm emits a harsh sound and a flashing red light, quickly attracting the attention of the production line staff. This immediate visual and auditory prompt can remind the relevant personnel to take immediate action in the first time to prevent the problem from spreading further. At the same time, the system sends detailed alarm information to the mobile device of the production line supervisor through network communication, notifying him to immediately go to the problem area for inspection. The alarm information not only includes the specific position, type, and confidence level of the defect, but also includes a real-time captured image segment for quickly locating and analyzing the root cause of the problem. This intelligent monitoring and early warning system greatly improves the production stability and product quality, and provides a strong guarantee for realizing automated and intelligent quality control.

[0062] In summary, the textile state detection and analysis system 100 based on the embodiments of the present application is elucidated. It first acquires the surface image of the textile to be inspected and performs image preprocessing, and then synchronously constructs a dual-branch feature extraction network to capture the local edge features of microscopic defects and the global distribution law of the macroscopic structure respectively; further, through cross-scale feature fusion to associate local details with global semantics, generates multi-scale image coding features of the surface of the textile to be inspected. Finally, combined with a pre-trained deep learning model, defect localization, classification and confidence evaluation are realized for the fused features, and it is automatically determined whether to issue an alarm signal according to the detection results. In this way, it can make the defect detection still maintain high sensitivity in a complex light environment, effectively suppress false detections and missed detections in shadow or overexposed areas, thereby significantly improving the accuracy and reliability of defect detection.

[0063] As described above, the textile state detection and analysis system 100 according to the embodiments of the present application can be implemented in various terminal devices. In one example, the textile state detection and analysis system 100 can be integrated into the terminal device as a software module and / or a hardware module. For example, the textile state detection and analysis system 100 can be a software module in the operating system of the terminal device, or can be an application program developed for the terminal device; of course, the textile state detection and analysis system 100 can also be one of the many hardware modules of the terminal device.

[0064] Alternatively, in another example, the textile state detection and analysis system 100 and the terminal device can also be separate devices, and the textile state detection and analysis system 100 can be connected to the terminal device through a wired and / or wireless network and transmit interaction information in accordance with a predefined data format.

[0065] In several embodiments provided by the present invention, it should be understood that the disclosed devices, apparatuses and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the modules is only a logical function division, and there may be other division methods in actual implementation.

[0066] The modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical units, that is, they may be located in one place, or may be distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0067] In addition, in each of the embodiments of the present invention, the functional modules can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above integrated unit can be implemented in the form of hardware, or in the form of a combination of hardware and software functional modules.

[0068] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and without departing from the spirit or basic characteristics of the present invention, the present invention can be implemented in other specific forms.

[0069] Therefore, from any point of view, the embodiments should be regarded as exemplary and non-restrictive. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present invention. Any reference signs in the claims should not be regarded as limiting the claims involved.

[0070] In addition, it is obvious that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. The multiple units or devices stated in the system claims can also be implemented by one unit or device through software or hardware. The terms such as "second" are used to denote names and do not denote any particular order.

[0071] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit of the technical solutions of the present invention.

Claims

1. A textile state detection and analysis system, characterized in that Including: A textile surface image acquisition module for acquiring a surface image of the textile to be inspected collected by a camera; A textile surface image enhancement module for performing image preprocessing on the surface image of the textile to be inspected to obtain an enhanced surface image of the textile to be inspected; A multi-scale feature encoding module for performing multi-scale image feature encoding on the enhanced surface image of the textile to be inspected to obtain multi-scale image encoding features of the surface of the textile to be inspected. Among them, the multi-scale feature encoding module includes: a multi-scale feature extraction unit for respectively extracting local detail information and global abstract information from the enhanced surface image of the textile to be inspected to obtain local detail encoding features of the surface of the textile to be inspected and global structure encoding features of the surface of the textile to be inspected; a multi-scale feature fusion unit for performing dynamic feature structure multi-scale fusion on the local detail encoding features of the surface of the textile to be inspected and the global structure encoding features of the surface of the textile to be inspected to obtain the multi-scale image encoding features of the surface of the textile to be inspected; An alarm module for obtaining a detection result based on the multi-scale image encoding features of the surface of the textile to be inspected and determining whether to issue an alarm signal; The multi-scale feature fusion unit includes: A feature decoupling sub-unit of the textile to be inspected for decoupling the local detail encoding feature map of the surface of the textile to be inspected and the global structure encoding feature map of the surface of the textile to be inspected to obtain a set of local detail feature encoding matrices of the surface of the textile to be inspected and a set of global structure local feature encoding matrices of the surface of the textile to be inspected; The multi-scale image feature alignment and fusion subunit includes: a feature phase dynamic search and alignment secondary subunit, which is used to perform feature phase dynamic search and alignment on the set of local detail feature encoding matrices of the surface of the textile to be inspected and the set of local feature encoding matrices of the global structure of the surface of the textile to be inspected based on the feature phase alignment degree between any two local detail feature encoding matrices of the surface of the textile to be inspected and the local feature encoding matrix of the global structure of the surface of the textile to be inspected in the set, so as to obtain a set of phase-aligned {local detail feature encoding matrix of the surface of the textile to be inspected, local feature encoding matrix of the global structure of the surface of the textile to be inspected} feature pairs; an attention weight calculation secondary subunit, which is used to input each phase-aligned {local detail feature encoding matrix of the surface of the textile to be inspected, local feature encoding matrix of the global structure of the surface of the textile to be inspected} feature pair in the set of phase-aligned {local detail feature encoding matrix of the surface of the textile to be inspected, local feature encoding matrix of the global structure of the surface of the textile to be inspected} feature pairs into a feature joint perception attention network to obtain a set of attention weights for the joint perception of the surface features of the textile to be inspected; a multi-scale saliency aggregation secondary subunit for the surface of the textile to be inspected, which is used to perform attention-driven saliency aggregation on the set of phase-aligned {local detail feature encoding matrix of the surface of the textile to be inspected, local feature encoding matrix of the global structure of the surface of the textile to be inspected} feature pairs based on the set of attention weights for the joint perception of the surface features of the textile to be inspected to obtain the multi-scale image encoding feature map of the surface of the textile to be inspected.

2. The textile state detection and analysis system according to claim 1, wherein The textile surface image enhancement module is used to: perform noise suppression and uneven illumination correction on the surface image of the textile to be inspected to obtain the enhanced surface image of the textile to be inspected.

3. The textile state detection and analysis system according to claim 1, wherein The multi-scale feature extraction unit is used to: use a multi-scale feature extractor based on the dilated spatial pyramid pooling model to extract local detail information and global abstract information from the enhanced surface image of the textile to be inspected respectively to obtain the local detail encoding feature map of the surface of the textile to be inspected as the local detail encoding feature of the surface of the textile to be inspected and the global structure encoding feature map of the surface of the textile to be inspected as the global structure encoding feature of the surface of the textile to be inspected.

4. The textile state detection and analysis system according to claim 3, characterized in that, The attention weight calculation secondary subunit includes: A three-level subunit for calculating the pair-scale correlation of the surface of the textile to be inspected, which is used to calculate the correlation matrix between each phase-aligned {local detail feature encoding matrix of the surface of the textile to be inspected, local feature encoding matrix of the global structure of the surface of the textile to be inspected} feature pair in the set of phase-aligned {local detail feature encoding matrix of the surface of the textile to be inspected, local feature encoding matrix of the global structure of the surface of the textile to be inspected} feature pairs to obtain a set of pair-scale correlation matrices of the surface of the textile to be inspected; A three-level subunit for trace metric of the surface of the textile to be inspected, which is used to perform trace metric on each pair-scale correlation matrix in the set of pair-scale correlation matrices of the surface of the textile to be inspected to obtain a set of trace metric values of the surface of the textile to be inspected; The trace metric value weight normalization three - level subunit is used to obtain a set of joint perception attention weights of the surface features of the textile to be inspected based on a set of trace metric values on the surface of the textile to be inspected.

5. The textile state detection and analysis system according to claim 4, wherein The trace metric value weight normalization three - level subunit is used for: Based on a set of trace metric values on the surface of the textile to be inspected, perform matrix manifold optimization based on the number of biconnectivities on each phase - aligned {local detailed feature encoding matrix of the surface of the textile to be inspected, local global - structure local feature encoding matrix of the surface of the textile to be inspected} feature pair in the set of phase - aligned {local detailed feature encoding matrix of the surface of the textile to be inspected, local global - structure local feature encoding matrix of the surface of the textile to be inspected} feature pairs to obtain a set of phase - optimized aligned {local detailed feature encoding matrix of the surface of the textile to be inspected, local global - structure local feature encoding matrix of the surface of the textile to be inspected} feature pairs; Based on the set of phase - optimized aligned {local detailed feature encoding matrix of the surface of the textile to be inspected, local global - structure local feature encoding matrix of the surface of the textile to be inspected} feature pairs, obtain a set of joint perception attention weights of the surface features of the textile to be inspected.

6. The textile state detection and analysis system according to claim 5, characterized in that, The alarm module is used for: Input the multi - scale image encoding feature map of the surface of the textile to be inspected into the trained defect detection model to obtain a detection result; Based on the detection result, determine whether to issue an alarm signal.

7. The textile state detection and analysis system according to claim 6, wherein The detection result includes the position of the defect point, the type of the defect point, and the confidence level.

Images