Image recognition-based ultra-soft knitted gauze fabric defect detection method and system

Abstract

The application relates to the technical field of image recognition, and particularly discloses a super-soft knitted yarn cloth defect detection method and system based on image recognition, which can be used for solving the problem that low-contrast micro-defects are difficult to perceive and detect in a complex texture background by constructing a three-level collaborative detection framework of "optical enhancement-feature decoupling-cross-modal perception", realizing accurate mapping of defect features from a physical space to a feature space, and breaking through the problem that texture periodic interference caused by a honeycomb-like weaving structure of super-soft yarn cloth and micro-defect sub-pixel level size characteristics are difficult to perceive and detect.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image recognition, and more particularly, to an image recognition-based defect detection method and system for super-soft knitted fabric. BACKGROUND

[0002] Super-soft knitted fabric is a special textile material made of ultra-fine denier fibers (single fiber fineness ≤0.8 dtex) through warp knitting or weft knitting process. Its unique "honeycomb" stitch structure and low weight (usually <30 g / m 2 ) endow the fabric with excellent air permeability, drape and skin touch, and it is widely used in high-end medical dressings, infant clothing and precision instrument packaging fields. In medical scenarios, 0.1 mm level micro-defects (such as broken yarns and hooked yarns) may cause bacterial penetration; in the field of optical instrument packaging, yarn knot defects with a diameter of <0.3 mm may cause mirror scratches, which puts forward the requirement of sub-millimeter level precision for the defect detection system.

[0003] However, traditional defect detection schemes such as gray histogram segmentation method have high false detection rate in super-soft fabric detection, because the gray standard deviation of normal texture and micro-defects has a significant overlap area. At the same time, due to the characteristics of super-soft knitted fabric, the defects themselves may be very small and have low contrast with normal texture, making the detection more difficult. Although subtle defects may not be obvious, they are easy to be ignored or submerged in the background texture, but in some high-quality application scenarios, they still need to be detected. In addition, although traditional frequency domain filtering methods (such as Gabor filter bank) can extract the base texture of the fabric, they cannot distinguish structural defects caused by stitch offset from normal texture fluctuations, and the detection accuracy is difficult to meet the actual needs when the production line speed is high.

[0004] Therefore, an optimized defect detection scheme for super-soft knitted fabric is expected. SUMMARY

[0005] The present application provides an image recognition-based defect detection method and system for super-soft knitted fabric, which can address the periodic texture interference caused by the honeycomb structure of super-soft knitted fabric and the sub-pixel level size characteristics of micro-defects, and realize accurate mapping of defect features from physical space to feature space by constructing a three-level collaborative detection framework of "optical enhancement-feature decoupling-cross-modal perception", thereby breaking through the problem of difficult perception and detection of low-contrast micro-defects in complex texture background. According to one aspect of the present application, an image recognition-based defect detection method for super-soft knitted fabric is provided, comprising: acquiring a super-soft knitted fabric image collected by an image acquisition system;

[0006] performing adaptive image preprocessing and contrast enhancement on the super-soft knitted fabric image to obtain a super-soft knitted fabric enhanced image;

[0007] performing multi-scale feature analysis on the super-soft knitted gauze material enhanced image to obtain super-soft knitted gauze material texture features and super-soft knitted gauze material depth features;

[0008] performing feature alignment joint perception on the super-soft knitted gauze material texture features and the super-soft knitted gauze material depth features to obtain super-soft knitted gauze material feature multi-scale joint perception coding features, including: performing feature decoupling on the super-soft knitted gauze material texture feature coding graph and the super-soft knitted gauze material depth feature coding graph to obtain a set of super-soft knitted gauze material texture feature coding matrices and a set of super-soft knitted gauze material depth local feature coding matrices; and performing phase alignment and attention saliency fusion on the set of super-soft knitted gauze material texture feature coding matrices and the set of super-soft knitted gauze material depth local feature coding matrices to obtain super-soft knitted gauze material feature multi-scale joint perception coding features;

[0009] based on the super-soft knitted gauze material feature multi-scale joint perception coding features, determining whether the super-soft knitted gauze material to be detected has defects.

[0010] In the above image recognition-based super-soft knitted gauze material defect detection method, the image acquisition system includes a high-resolution industrial camera and an illumination system.

[0011] In the above image recognition-based super-soft knitted gauze material defect detection method, the illumination system uses a uniform diffuse reflection light source. In the above image recognition-based super-soft knitted gauze material defect detection method, the super-soft knitted gauze material image is subjected to adaptive image preprocessing and contrast enhancement to obtain a super-soft knitted gauze material enhanced image, including:

[0012] performing adaptive Wiener filtering processing on the super-soft knitted gauze material image to obtain a super-soft knitted gauze material adaptive preprocessing image;

[0013] performing multi-scale contrast enhancement processing on the super-soft knitted gauze material adaptive preprocessing image to obtain a super-soft knitted gauze material enhanced image.

[0014] In the above image recognition-based super-soft knitted gauze material defect detection method, multi-scale feature analysis is performed on the super-soft knitted gauze material enhanced image to obtain super-soft knitted gauze material texture features and super-soft knitted gauze material depth features, including: using a multi-scale convolutional neural network to perform multi-scale image feature extraction on the super-soft knitted gauze material enhanced image to obtain a super-soft knitted gauze material texture feature coding graph as the super-soft knitted gauze material texture features and a super-soft knitted gauze material depth feature coding graph as the super-soft knitted gauze material depth features.

[0015] In the above image recognition-based defect detection method of super-soft knitted yarn fabric, the set of super-soft knitted yarn fabric texture feature encoding matrices and the set of super-soft knitted yarn fabric depth local feature encoding matrices are phase-aligned and attention-saliently fused to obtain super-soft knitted yarn fabric feature multi-scale joint perception encoding features, including:

[0016] Based on the feature phase alignment degree between any two super-soft knitted yarn fabric texture feature encoding matrices and super-soft knitted yarn fabric depth local feature encoding matrices in the set of super-soft knitted yarn fabric texture feature encoding matrices and the set of super-soft knitted yarn fabric depth local feature encoding matrices, the set of super-soft knitted yarn fabric texture feature encoding matrices and the set of super-soft knitted yarn fabric depth local feature encoding matrices are dynamically searched and aligned in feature phase to obtain a set of phase-aligned {super-soft knitted yarn fabric texture feature encoding matrix, super-soft knitted yarn fabric depth local feature encoding matrix} feature pairs; each phase-aligned {super-soft knitted yarn fabric texture feature encoding matrix, super-soft knitted yarn fabric depth local feature encoding matrix} feature pair in the set of phase-aligned {super-soft knitted yarn fabric texture feature encoding matrix, super-soft knitted yarn fabric depth local feature encoding matrix} feature pairs is input into a feature joint perception attention network to obtain a set of super-soft knitted yarn fabric feature joint perception attention weights;

[0017] Based on the set of super-soft knitted yarn fabric feature joint perception attention weights, the set of phase-aligned {super-soft knitted yarn fabric texture feature encoding matrix, super-soft knitted yarn fabric depth local feature encoding matrix} feature pairs is subjected to attention-driven salient aggregation to obtain a super-soft knitted yarn fabric feature multi-scale joint perception encoding map as the super-soft knitted yarn fabric feature multi-scale joint perception encoding feature.

[0018] In the above image recognition-based defect detection method of super-soft knitted yarn fabric, each phase-aligned {super-soft knitted yarn fabric texture feature encoding matrix, super-soft knitted yarn fabric depth local feature encoding matrix} feature pair in the set of phase-aligned {super-soft knitted yarn fabric texture feature encoding matrix, super-soft knitted yarn fabric depth local feature encoding matrix} feature pairs is input into a feature joint perception attention network to obtain a set of super-soft knitted yarn fabric feature joint perception attention weights, including:

[0019] The correlation matrix between each phase-aligned {super-soft knitted yarn fabric texture feature encoding matrix, super-soft knitted yarn fabric depth local feature encoding matrix} feature pair in the set of phase-aligned {super-soft knitted yarn fabric texture feature encoding matrix, super-soft knitted yarn fabric depth local feature encoding matrix} feature pairs is calculated to obtain a set of textile surface pair scale correlation matrices;

[0020] trace measure each surface pair scale correlation matrix in the set of surface pair scale correlation matrices of the textile surface under test to obtain a set of surface trace measure values of the textile surface under test;

[0021] based on the set of surface trace measure values of the textile surface under test, obtain a set of feature joint perception attention weights of the ultra-soft knitted yarn cloth.

[0022] In the above image recognition-based ultra-soft knitted yarn cloth defect detection method, based on the set of surface trace measure values of the textile surface under test, a set of feature joint perception attention weights of the ultra-soft knitted yarn cloth is obtained, including:

[0023] based on the set of surface trace measure values of the textile surface under test, perform matrix manifold optimization based on double-linkage number on each phase-aligned {ultra-soft knitted yarn cloth texture feature encoding matrix, ultra-soft knitted yarn cloth depth local feature encoding matrix} feature pair in the set of phase-aligned {ultra-soft knitted yarn cloth texture feature encoding matrix, ultra-soft knitted yarn cloth depth local feature encoding matrix} feature pairs to obtain a set of phase-optimized aligned {ultra-soft knitted yarn cloth texture feature encoding matrix, ultra-soft knitted yarn cloth depth local feature encoding matrix} feature pairs;

[0024] based on the set of phase-optimized aligned {ultra-soft knitted yarn cloth texture feature encoding matrix, ultra-soft knitted yarn cloth depth local feature encoding matrix} feature pairs, obtain a set of feature joint perception attention weights of the ultra-soft knitted yarn cloth.

[0025] According to another aspect of the present application, an image recognition-based ultra-soft knitted yarn cloth defect detection system is provided, comprising:

[0026] an ultra-soft knitted yarn cloth image acquisition module for acquiring an ultra-soft knitted yarn cloth image collected by an image collection system; an ultra-soft knitted yarn cloth image enhancement module for performing adaptive image preprocessing and contrast enhancement on the ultra-soft knitted yarn cloth image to obtain an ultra-soft knitted yarn cloth enhanced image;

[0027] an image multi-scale feature analysis module for performing multi-scale feature analysis on the ultra-soft knitted yarn cloth enhanced image to obtain ultra-soft knitted yarn cloth texture features and ultra-soft knitted yarn cloth depth features;

[0028] The image feature alignment joint perception module is used for feature alignment joint perception of the super-soft knitted gauze fabric texture feature and the super-soft knitted gauze fabric depth feature to obtain the super-soft knitted gauze fabric feature multi-scale joint perception coding feature. The image feature alignment joint perception module comprises: an image feature decoupling module, which is used for feature decoupling of the super-soft knitted gauze fabric texture feature coding graph and the super-soft knitted gauze fabric depth feature coding graph to obtain a set of super-soft knitted gauze fabric texture feature coding matrices and a set of super-soft knitted gauze fabric depth local feature coding matrices; and an image feature fusion coding module, which is used for phase alignment and attention significant fusion of the set of super-soft knitted gauze fabric texture feature coding matrices and the set of super-soft knitted gauze fabric depth local feature coding matrices to obtain the super-soft knitted gauze fabric feature multi-scale joint perception coding feature.

[0029] The super-soft knitted gauze fabric detection result determination module is used for determining whether the super-soft knitted gauze fabric to be detected has defects based on the super-soft knitted gauze fabric feature multi-scale joint perception coding feature.

[0030] In the above-mentioned image recognition-based super-soft knitted gauze fabric defect detection system, the image acquisition system comprises a high-resolution industrial camera and an illumination system.

[0031] The application provides an image recognition-based super-soft knitted gauze fabric defect detection method and system, which can be used for the texture periodic interference caused by the honeycomb-like weaving structure of the super-soft gauze and the sub-pixel level size feature of the micro-defects, and can realize accurate mapping of the defect feature from the physical space to the feature space by constructing a three-level collaborative detection framework of “optical enhancement-feature decoupling-cross-modal perception”, thereby breaking through the problem of difficult perception and detection of low-contrast micro-defects in a complex texture background. BRIEF DESCRIPTION OF DRAWINGS

[0032] In order to more clearly illustrate the technical solutions of the embodiments of the application, the drawings of the embodiments of the application will be briefly introduced below. Obviously, the drawings described below only relate to some embodiments of the application, rather than limiting the application.

[0033] Figure 1 The schematic flowchart of the image recognition-based super-soft knitted gauze fabric defect detection method of the embodiments of the application.

[0034] Figure 2 The data flow schematic diagram of the image recognition-based super-soft knitted gauze fabric defect detection method of the embodiments of the application. Figure 3 The schematic flowchart of S2 in the image recognition-based super-soft knitted gauze fabric defect detection method of the embodiments of the application.

[0035] Figure 4A schematic flowchart of S4 in the image recognition-based defect detection method for super-soft knitted yarn fabric of the embodiment of the present application.

[0036] Figure 5 A schematic flowchart of S42 in the image recognition-based defect detection method for super-soft knitted yarn fabric of the embodiment of the present application.

[0037] Figure 6 A schematic block diagram of the image recognition-based defect detection system for super-soft knitted yarn fabric of the embodiment of the present application. DETAILED DESCRIPTION

[0038] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor also fall within the scope of protection of the present application.

[0039] To solve the above technical problems, in the technical solutions of the present application, an image recognition-based defect detection method for super-soft knitted yarn fabric is proposed, which can be used to solve the problems of difficult perception and detection of low-contrast micro-defects in complex texture background by constructing a three-level collaborative detection framework of “optical enhancement-feature decoupling-cross-modal perception”, realizing accurate mapping of defect features from physical space to feature space, and thus breaking through the problems of texture periodic interference caused by honeycomb-like weaving structure of super-soft yarn fabric and micro-defect sub-pixel level size characteristics.

[0040] Specifically, the technical concept of the present application is: at the level of spatial domain, through the collaborative design of the optical imaging system and the adaptive preprocessing module, a physical separation basis of defect features and background texture is established. The uniform diffuse reflection light source is combined with a specific incident angle configuration, the anisotropic reflection characteristics of the microstructure of the super-soft gauze surface are utilized to convert the geometric distortion of the defect into the differential response of the light intensity distribution, and the physical enhancement effect of the defect is formed. The adaptive filtering algorithm dynamically adjusts the noise reduction strength according to the local texture spectrum, suppresses the high-frequency noise, and at the same time, constructs the gradient protection domain of the defect edge, avoiding the loss of details caused by traditional global filtering. At the level of frequency domain feature modeling, a heterogeneous multi-scale convolutional network architecture is adopted to realize the dual-channel decoupling of the base texture and the defect features. The shallow network branch captures the microscopic topological features of the yarn interlacing nodes through dense convolution kernels, and constructs the phase spectrum of the periodic texture; the deep branch extracts the long-range correlation pattern across the yarn scale by means of dilated convolution, and forms the semantic encoding of the abnormal features. Through the dynamic phase alignment mechanism, the texture phase spectrum and the deep semantic encoding are sub-pixel level feature registration in the frequency domain space, eliminating the feature misregistration interference caused by the elastic deformation of the fabric, establishing a cross-scale joint perception model of the defect response, and improving the signal-to-noise separation ability of sub-millimeter level defects in the dynamic detection scene, providing a basis for the defect detection of super-soft knitted gauze.

[0041] Specifically, Figure 1 The schematic flow chart of the image recognition based defect detection method of the super-soft knitted gauze of the embodiment of the present application is shown in FIG. 1. Figure 2 The data flow schematic diagram of the image recognition based defect detection method of the super-soft knitted gauze of the embodiment of the present application is shown in FIG. 2. Figure 1 and Figure 2 As shown in FIG. 1, the image recognition based defect detection method of the super-soft knitted gauze includes the following steps: S1, acquiring the super-soft knitted gauze image collected by the image acquisition system; S2, performing adaptive image preprocessing and contrast enhancement on the super-soft knitted gauze image to obtain a super-soft knitted gauze enhanced image; S3, performing multi-scale feature analysis on the super-soft knitted gauze enhanced image to obtain super-soft knitted gauze texture features and super-soft knitted gauze depth features; S4, performing feature alignment joint perception on the super-soft knitted gauze texture features and the super-soft knitted gauze depth features to obtain super-soft knitted gauze feature multi-scale joint perception coding features; S5, determining whether the super-soft knitted gauze to be detected has defects based on the super-soft knitted gauze feature multi-scale joint perception coding features.

[0042] Exemplarily, in step S1, an image of the super-soft knitted gauze fabric is acquired by an image acquisition system. In an embodiment, the image acquisition system comprises a high-resolution industrial camera and an illumination system. The illumination system adopts a uniform diffuse reflection light source. That is, in the aspect of optical imaging, the uniform diffuse reflection light source is adopted in cooperation with the high-resolution industrial camera to enhance the stereoscopic characterization ability of the surface micro-topography by controlling the incident light angle (recommended 15°±2° low-angle illumination), so as to produce a directional shadow effect of linear defects such as hooking and broken yarn, and to improve the gray contrast of the defects and the base texture.

[0043] In an embodiment, as shown in Figure 3 S2, adaptive image preprocessing and contrast enhancement are performed on the image of the super-soft knitted gauze fabric to obtain a super-soft knitted gauze fabric enhanced image, comprising: S21, adaptive Wiener filtering processing is performed on the image of the super-soft knitted gauze fabric to obtain a super-soft knitted gauze fabric adaptive preprocessing image; and S22, multi-scale contrast enhancement processing is performed on the super-soft knitted gauze fabric adaptive preprocessing image to obtain a super-soft knitted gauze fabric enhanced image. It can be understood that, since the super-soft knitted gauze fabric base texture has a periodic high-frequency component, and the micro-defects (such as 0.1 mm broken yarn) only occupy a 5-15 pixel area in the image, the traditional global filtering algorithm will cause the gradient of the defect edge to be diffused, and thus the frequency domain aliasing of the defect features and the background texture is aggravated. The adaptive Wiener filtering dynamically perceives the spectral characteristics of the local texture (using a sliding window of 7×7 to 15×15 pixels), constructs a real-time estimation model of the noise variance and the signal power in the frequency domain space, realizes the dynamic balance of high-frequency noise suppression (σ<0.3) and defect edge preservation (sharpness loss rate <8%), and overcomes the artifact interference caused by the traditional fixed parameter filter in the yarn texture mutation area. In addition, the multi-scale contrast enhancement processing is carried out according to the difference in the frequency energy distribution of the defects and the background: the histogram equalization is adopted in the low-frequency component (scale factor λ=8) to enhance the overall contrast of the base texture, and the nonlinear gain amplification is implemented in the high-frequency component (scale factor λ=2) to emphasize the local gradient response of the micro-defects. This strategy enhances the edge gradient intensity of the defect area through the brightness channel reconstruction of the HSV color space, while compresses the background texture fluctuation amplitude, effectively solving the texture artifact problem caused by excessive enhancement in the traditional algorithm in the yarn detection.

[0044] Exemplarily, in step S3, the super-soft knitted gauze material enhanced image is subjected to multi-scale feature analysis to obtain super-soft knitted gauze material texture features and super-soft knitted gauze material depth features. It can be understood that, due to the periodic high-frequency texture of the honeycomb-like weaving structure of the super-soft gauze in the image, and the micro-defects (such as 0.3 mm yarn nodes) only occupy an area of about 15x15 pixels under high-resolution imaging, the fixed receptive field of the traditional single-scale convolutional network is difficult to capture both the microscopic deformation of the yarn nodes and the long-range structural abnormalities across the yarns.

[0045] In one embodiment, the multi-scale feature analysis of the super-soft knitted gauze material enhanced image to obtain super-soft knitted gauze material texture features and super-soft knitted gauze material depth features comprises: using a multi-scale convolutional neural network to perform multi-scale image feature extraction on the super-soft knitted gauze material enhanced image to obtain a super-soft knitted gauze material texture feature encoding graph as the super-soft knitted gauze material texture features and a super-soft knitted gauze material depth feature encoding graph as the super-soft knitted gauze material depth features. In the technical solution of the present application, the multi-scale convolutional neural network is further used to perform multi-scale image feature extraction on the super-soft knitted gauze material enhanced image to obtain the super-soft knitted gauze material texture feature encoding graph and the super-soft knitted gauze material depth feature encoding graph. In the texture feature encoding branch, the yarn base texture is subjected to pixel-level phase analysis through densely stacked small convolutional kernels, and the topological continuity interruption features caused by yarn breakage can be accurately captured in view of the abnormality of the local phase distribution of the super-soft gauze stitch offset defects. In the depth feature encoding branch, an expanded convolution is used to construct a semantic perception field across the yarn scale, and the yarn tension imbalance features caused by the hooking defects are extracted.

[0046] Exemplarily, in step S4, the texture feature and the depth feature of the super-soft knitted gauze are subjected to feature alignment joint perception to obtain a super-soft knitted gauze feature multi-scale joint perception coding feature. It can be understood that, due to the elastic deformation of the honeycomb weaving structure of the super-soft gauze in the dynamic detection process, a sub-pixel level spatial misalignment exists between the texture feature coding graph and the depth feature coding graph. The traditional feature fusion method (such as channel splicing or weighted average) cannot eliminate the feature response deviation caused by such geometric distortion, therefore, in the technical solution of the present application, the texture feature coding graph and the depth feature coding graph of the super-soft knitted gauze are further subjected to feature alignment joint perception to obtain a super-soft knitted gauze feature multi-scale joint perception coding graph. Specifically, by decoupling the texture feature coding graph (focusing on the yarn node phase information) and the depth feature coding graph (representing the cross-yarn semantic association) of the super-soft knitted gauze into a local feature coding matrix set, the feature phase alignment degree (distribution graph smoothness calculated based on the number of double connections) is used to dynamically search for the optimal matching pair, breaking through the dependence of the traditional spatial alignment method on rigid transformation assumption. In the phase alignment process, the embedding manifold tightening technology reduces the geometric correlation error of the feature pair at the pixel level by optimizing the trace metric operation of the discrete manifold representation, significantly improving the matching robustness under long-range deformation. In particular, the processing process of the feature alignment joint perception models the interaction relationship of the phase alignment feature pair, realizes the significant enhancement of the defect response area and the synergistic suppression of the background texture interference, while reducing the structural defect misjudgment rate caused by stitch deviation, breaking through the "high false negative-high false positive" dilemma caused by feature misalignment of the traditional method.

[0047] In one embodiment, as shown in Figure 4 S4 includes: S41, decoupling the texture feature coding graph and the depth feature coding graph of the super-soft knitted gauze to obtain a set of super-soft knitted gauze texture feature coding matrices and a set of super-soft knitted gauze depth local feature coding matrices; S42, performing phase alignment and attention significant fusion on the set of super-soft knitted gauze texture feature coding matrices and the set of super-soft knitted gauze depth local feature coding matrices to obtain super-soft knitted gauze feature multi-scale joint perception coding features.

[0048] Specifically, in step S41, the texture feature coding graph and the depth feature coding graph of the super-soft knitted gauze are decoupled to obtain a set of super-soft knitted gauze texture feature coding matrices and a set of super-soft knitted gauze depth local feature coding matrices, and specifically, the process can be represented by the formula:

[0049] Decouple(F1)={M 11 ,M 12 ,...,M1i ..., M 1n}

[0050] Decouple(F2) = {M 21 , M 22 ..., M 2j ..., M 2n}

[0051] where F1 is the texture feature encoding map of the super-soft knitted fabric, F2 is the depth feature encoding map of the super-soft knitted fabric, Decouple(·) is the feature decoupling operation, M 11 , M 12 , M 1i and M 1n are the 1st, 2nd, i-th and n-th super-soft knitted fabric texture feature encoding matrix in the set of super-soft knitted fabric texture feature encoding matrices, M 21 , M 22 , M 2j and M 2n are the 1st, 2nd, j-th and n-th super-soft knitted fabric depth local feature encoding matrix in the set of super-soft knitted fabric depth local feature encoding matrices.

[0052] It can be understood that by decomposing the texture feature encoding map (mainly focusing on the micro-topological features of yarn nodes) and the depth feature encoding map (emphasizing long-range correlation patterns across yarn scales) of the super-soft knitted fabric into multiple local feature encoding matrices, a more detailed analysis can be conducted against the complex periodic texture background of the super-soft knitted fabric. This decomposition allows each local feature encoding matrix to focus on describing subtle structural variations within specific regions, such as the specific manifestations of stitch deviation, hooking defects, etc., thereby improving the ability to capture these detailed features. For example, by decoupling the texture feature encoding map of the super-soft knitted fabric, a local feature encoding matrix reflecting the topological continuity interruption features caused by yarn breakage can be obtained; similarly, the local feature encoding matrix extracted from the depth feature encoding map can accurately capture the yarn tension imbalance features caused by hooking defects.

[0053] In one embodiment, as Figure 5The set of super-soft knitted yarn fabric texture feature coding matrices and the set of super-soft knitted yarn fabric depth local feature coding matrices are phase-aligned and attention-significantly fused to obtain super-soft knitted yarn fabric feature multi-scale joint perception coding features, including: S421, based on the feature phase alignment degree between any two super-soft knitted yarn fabric texture feature coding matrices and super-soft knitted yarn fabric depth local feature coding matrices in the set of super-soft knitted yarn fabric texture feature coding matrices and the set of super-soft knitted yarn fabric depth local feature coding matrices, the set of super-soft knitted yarn fabric texture feature coding matrices and the set of super-soft knitted yarn fabric depth local feature coding matrices are dynamically searched and aligned in feature phase to obtain a set of phase-aligned {super-soft knitted yarn fabric texture feature coding matrix, super-soft knitted yarn fabric depth local feature coding matrix} feature pairs; S422, input each phase-aligned {super-soft knitted yarn fabric texture feature coding matrix, super-soft knitted yarn fabric depth local feature coding matrix} feature pair in the set of phase-aligned {super-soft knitted yarn fabric texture feature coding matrix, super-soft knitted yarn fabric depth local feature coding matrix} feature pairs into a feature joint perception attention network to obtain a set of super-soft knitted yarn fabric feature joint perception attention weights; S423, based on the set of super-soft knitted yarn fabric feature joint perception attention weights, the set of phase-aligned {super-soft knitted yarn fabric texture feature coding matrix, super-soft knitted yarn fabric depth local feature coding matrix} feature pairs are significantly aggregated based on attention driving to obtain a super-soft knitted yarn fabric feature multi-scale joint perception coding graph as the super-soft knitted yarn fabric feature multi-scale joint perception coding feature.

[0054] Specifically, in step S421, the process of feature phase dynamic search alignment can be represented by the formula:

[0055]

[0056] P ik ={M 1i ,M 2k} pair

[0057] Where T is the transpose operation, ‖·‖ F is the calculation of the Frobenius norm, d P (M 1i ,M 2j ) is the feature phase alignment degree between M 1i and M 2j , is to return the j value corresponding to the maximum value, k is the position of searching the maximum approximate matching value in the set of super-soft knitted yarn fabric depth local feature coding matrices, {M 1i ′,M 2k ′} pairis a phase-optimized alignment pair of {texture feature encoding matrix of super-soft knitted fabric, depth local feature encoding matrix of super-soft knitted fabric} features, P ik represents a feature pair.

[0058] It should be understood that the feature phase dynamic search alignment emphasizes finding the feature pair with the highest phase alignment degree in the feature set, rather than simply relying on the static spatial position correspondence. This enables the model to flexibly cope with the elastic deformation that may occur during the dynamic detection of super-soft knitted fabric within a certain range, thereby ensuring accurate matching even in the presence of sub-pixel level spatial misalignment. By optimizing the trace metric operation of the discrete manifold representation, the geometric correlation error is further reduced, and the matching robustness under long-range deformation is improved. This precise feature alignment provides high-quality basic data for subsequent feature joint perception, which helps to improve the quality of the generated joint perception feature map. In addition, the feature phase dynamic search alignment not only improves the accuracy of feature alignment, but also enhances the model's ability to recognize complex backgrounds and minor defects. In the detection of super-soft knitted fabric, even slight deformation can affect the final detection results, so this strategy is crucial for ensuring high-precision defect detection. Through effective measurement of feature phase alignment degree and dynamic search strategy, the model can efficiently find the best matching pair in the set of local feature encoding matrices, overcoming problems such as spatial misalignment, deformation, or occlusion that may exist in the feature map, and ultimately generating a set of phase-aligned feature pairs. These feature pairs, after being processed by the subsequent attention mechanism, can adaptively learn the importance weights of different feature pairs, further optimizing the feature aggregation process, highlighting important feature contributions, suppressing noise interference, and enhancing the model's interpretability and generalization ability.

[0059] In one embodiment, in step S422, each phase-aligned {texture feature encoding matrix of super-soft knitted fabric, depth local feature encoding matrix of super-soft knitted fabric} feature pair in the set of phase-aligned {texture feature encoding matrix of super-soft knitted fabric, depth local feature encoding matrix of super-soft knitted fabric} feature pairs is input into the feature joint perception attention network to obtain a set of super-soft knitted fabric feature joint perception attention weights, including: calculating the correlation matrix between each phase-aligned {texture feature encoding matrix of super-soft knitted fabric, depth local feature encoding matrix of super-soft knitted fabric} feature pair in the set of phase-aligned {texture feature encoding matrix of super-soft knitted fabric, depth local feature encoding matrix of super-soft knitted fabric} feature pairs to obtain a set of surface pair scale correlation matrices of the textile to be inspected; and performing trace metric on each surface pair scale correlation matrix in the set of surface pair scale correlation matrices of the textile to be inspected to obtain a set of surface trace metric values of the textile to be inspected, specifically, the process can be represented by the formula:

[0060] ai = Tr({M 1i ,M 2k} pair ) = Tr(M 1i T M 2k )

[0061] wherein Tr(·) is a trace metric value of a matrix, a i is a surface trace metric value between M 1i and M 2k .

[0062] Based on the set of surface trace metric values of the textile to be inspected, a set of feature joint perception attention weights of the super-soft knitted gauze fabric is obtained, and specifically, the process can be represented by the formula:

[0063]

[0064] wherein A net (·) is a feature joint perception attention weight calculation, M 1i T ' is an optimized phase-optimized aligned super-soft knitted gauze fabric texture feature encoding matrix of M 1i , M 2l ' is an optimized phase-optimized aligned super-soft knitted gauze fabric depth local feature encoding matrix of M 2k , a i ' is an optimized surface trace metric value between M 1i and M 2k , sigmoid is a normalization function, and w i is a feature joint perception attention weight of the super-soft knitted gauze fabric between M 1i and M 2k .

[0065] It should be appreciated that, in the context of the complex honeycomb-like weaving structure of super-soft knitted fabric, traditional feature fusion methods often struggle to fully exploit the complementary information between different modalities or levels, and may even introduce noise interference. However, the application of the feature joint perception attention network can effectively address these issues. First, the feature joint perception attention network is not limited to predicting weights, but is more importantly a key component of the joint perception capability. Although the phase alignment step has preliminarily established the correspondence relationship between features with similar structures, the importance of different feature pairs and their complex interaction patterns still need to be further modeled and learned. Through the attention mechanism, the network can assign appropriate importance weights to each phase-aligned feature pair and learn the interaction between the feature pairs. This adaptive learning approach enables the model to selectively integrate complementary feature information while suppressing redundant or irrelevant information during the subsequent feature aggregation process, thereby improving the quality of the final feature representation.

[0066] Secondly, the feature joint perception attention network can significantly enhance the model's ability to recognize complex backgrounds and minor defects. In the detection of super-soft knitted fabric, texture features and depth features provide different information: texture feature maps mainly capture the microscopic topological features of yarn nodes, while depth feature maps focus on long-range correlation patterns across yarn scales. Through the feature joint perception attention network, the weights of these feature pairs can be dynamically adjusted to ensure that important features are fully valued while secondary features or noise are effectively suppressed. For example, when detecting yarn breakage defects caused by stitch deviation, the attention network can assign higher weights to relevant texture features, while increasing the weight of depth features when detecting tension imbalance caused by hooking, thereby improving the accuracy of defect recognition.

[0067] In addition, the feature joint perception attention network also enhances the model's robustness to geometric deformation and occlusion. Since super-soft knitted fabric may experience elastic deformation or have partially occluded regions in practical applications, traditional methods tend to perform poorly in such cases. However, through the attention mechanism, the model can automatically correct deformations to some extent and suppress the influence of occluded regions. Specifically, the feature joint perception attention network can identify and highlight feature pairs that contain key information, ensuring the stability and reliability of the detection results even in the presence of slight deformation or occlusion.

[0068] In one embodiment, based on the set of textile surface trace metric values of the textile to be inspected, a set of joint perception attention weights of the ultra-soft knitted gauze fabric features is obtained, including: based on the set of textile surface trace metric values of the textile to be inspected, performing matrix manifold optimization based on the double incidence number on each phase-aligned {ultra-soft knitted gauze fabric texture feature encoding matrix, ultra-soft knitted gauze fabric depth local feature encoding matrix} feature pair in the set of phase-aligned {ultra-soft knitted gauze fabric texture feature encoding matrix, ultra-soft knitted gauze fabric depth local feature encoding matrix} feature pairs to obtain a set of phase-optimized aligned {ultra-soft knitted gauze fabric texture feature encoding matrix, ultra-soft knitted gauze fabric depth local feature encoding matrix} feature pairs; based on the set of phase-optimized aligned {ultra-soft knitted gauze fabric texture feature encoding matrix, ultra-soft knitted gauze fabric depth local feature encoding matrix} feature pairs, a set of joint perception attention weights of the ultra-soft knitted gauze fabric features is obtained.

[0069] In particular, here, for each M 1i T and M 2k , let M a = M 1i T and M b = M 2k , and m ai ∈ M a , m bi ∈ M b , the double incidence number α and β between the matrices M ai and M bi is obtained by calculating the number of eigenvalues that satisfy the L1 distance d L1 (m ai , m bi )<ε and the L2 distance d L2 (m ai , m bi )<ε between the corresponding m a and m b , that is, the distribution map smoothness between the phase-aligned ultra-soft knitted gauze fabric texture feature encoding matrix and the ultra-soft knitted gauze fabric depth local feature encoding matrix.

[0070] Then, in the discrete manifold representation of the ultra-soft knitted gauze fabric texture feature encoding matrix and the ultra-soft knitted gauze fabric depth local feature encoding matrix, the trace metric operation of the ultra-soft knitted gauze fabric texture feature encoding matrix and the ultra-soft knitted gauze fabric depth local feature encoding matrix is embedded optimized based on the embedding manifold compactification of the double incidence number, so as to improve the calculation accuracy of the joint matrix trace metric on the basis of improving the long-range correlation incidence robustness:

[0071] M1i T ′=exp[-(α-β+a i )⊙M 1i T ]

[0072] M 2k ′=exp[-(β-α+a i )⊙M 2k ]

[0073] Among them, based on the initial M 1i T and M 2k Calculate a i , α and β, substitute a i ,α,β, initial M 1i T and M 2k Optimize and then calculate a i ′=Tr(M 1i T 'M 2k ′) Calculate the optimized trace metric.

[0074] Specifically, in step S423, the process of attention-driven saliency aggregation can be expressed as follows:

[0075]

[0076] F f =[M1′,M2′,...,M i ′,...,M n ′]

[0077] Among them, W1 and W2 are the first joint perception weight matrix and the second joint perception weight matrix, It is subtracted by position point, It is added by position point, M′ i It's M 1i and M 2k The combined super-flexible knitting yarn fabric feature joint perception matrix is ​​the i-th super-flexible knitting yarn fabric feature joint perception matrix in the set of super-flexible knitting yarn fabric feature joint perception matrices, M′1, ′2 and M′ n are the first, second, and nth super-soft knitting yarn fabric feature joint perception matrices in the set of super-soft knitting yarn fabric feature joint perception matrices, F fis a super-soft knitted fabric feature multi-scale joint perception encoding map. Here, it should be known to those skilled in the art that the first joint perception weight matrix and the second joint perception weight matrix and other weight matrices or bias and the like network parameters of the present application are obtained by training. In the initial stage of training, the parameters in these weight matrices are randomly initialized. With the advancement of the training process, the weight values are gradually adjusted according to the feedback information provided by the loss function using optimization strategies such as back propagation algorithm and gradient descent method. The design of the loss function aims to measure the gap between the model prediction results and the actual labels, and the network parameters are optimized by minimizing this gap.

[0078] It should be understood that the attention-driven salient aggregation process aims to highlight important feature contributions rather than simply performing average or concatenation operations. By applying feature joint perception attention weights to the set of phase-aligned feature pairs, it can ensure that those containing key information dominate in the final feature map, while secondary features or noise are suppressed. This selective aggregation approach enables the model to more clearly identify target objects, key regions, or important feature relationships in complex backgrounds, thereby improving the performance of subsequent tasks. For example, in detecting subtle defects such as broken yarn or hook yarn in super-soft knitted fabric, attention-driven salient aggregation can focus on these key areas, providing more accurate and reliable feature representation. Secondly, salient aggregation not only improves the quality of the feature map, but also enhances the model's robustness to geometric deformation and occlusion. Since super-soft knitted fabric may experience elastic deformation or have partially occluded regions in actual applications, traditional methods may experience performance degradation in such cases. Through the attention mechanism, the model can automatically correct the deformation to some extent and suppress the influence of the occluded region. Specifically, in the salient aggregation process, attention weights guide the model to preferentially focus on feature pairs that are not occluded and contain key information, ensuring the stability and reliability of the detection results even in the presence of slight deformation or occlusion. In addition, salient aggregation further optimizes the effect of feature fusion, achieving effective integration of multi-scale information. The texture features and depth features of super-soft knitted fabric provide different information: the texture feature encoding map mainly captures the microscopic topological features of yarn nodes, while the depth feature encoding map focuses on long-range correlation patterns across yarn scales. Through salient aggregation, the model can organically combine these different scale information to form a more comprehensive and robust feature representation. For example, in detecting broken yarn defects caused by stitch offset, salient aggregation can consider the relevant information of texture features and depth features to ensure that the final generated feature map can more accurately reflect these subtle changes, thereby improving the accuracy of defect detection.

[0079] Exemplarily, in step S5, based on the multi-scale joint perceptual coding features of the super-flexible knitting yarn fabric characteristics, it is determined whether the super-flexible knitting yarn fabric to be detected has defects. In particular, in a specific example of the present application, the multi-scale joint perceptual coding map of the super-flexible knitting yarn fabric characteristics is detected by a fabric defect detector based on a classifier to determine whether the super-flexible knitting yarn fabric to be detected has defects. It should be understood that the output layer of the classifier generally uses a fully connected layer and a softmax function to map the extracted high-level features to a specific category probability distribution. First, the multi-scale joint perceptual coding map of the super-flexible knitting yarn fabric characteristics is expanded into a multi-scale joint perceptual coding vector of the super-flexible knitting yarn fabric characteristics. The fully connected layer converts the multi-scale joint perceptual coding vector of the super-flexible knitting yarn fabric characteristics into a fixed-length vector representation, and the softmax function maps these vectors to the probability values ​​of each category. In practical applications, two categories can be defined: normal area and defective area. The model compares the probability values ​​of each category and selects the category with the highest probability as the final prediction result.

[0080] In summary, according to the embodiment of the present application, the image recognition-based ultra-flexible knitted yarn fabric defect detection method and system are explained, which can target the texture periodic interference and micro-defect sub-pixel size characteristics caused by the honeycomb weaving structure of ultra-flexible gauze, and realize the accurate mapping of defect characteristics from physical space to feature space by constructing a three-level collaborative detection framework of "optical enhancement-feature decoupling-cross-modal perception", thereby breaking through the problem of low-contrast micro-defects being difficult to perceive and detect in a complex texture background.

[0081] Figure 6 Schematic block diagram of the ultra-soft knitted yarn fabric defect detection system based on image recognition according to an embodiment of the present application. Figure 6 As shown, a super-flexible knitted yarn fabric defect detection system 10 based on image recognition includes: a super-flexible knitted yarn fabric image acquisition module 11, which is used to acquire a super-flexible knitted yarn fabric image acquired by an image acquisition system; a super-flexible knitted yarn fabric image enhancement module 12, which is used to perform adaptive image preprocessing and contrast enhancement on the super-flexible knitted yarn fabric image to obtain a super-flexible knitted yarn fabric enhanced image; an image multi-scale feature analysis module 13, which is used to perform multi-scale feature analysis on the super-flexible knitted yarn fabric enhanced image to obtain super-flexible knitted yarn fabric texture features and super-flexible knitted yarn fabric depth features; an image feature alignment and joint perception module 14, which is used to perform feature alignment and joint perception on the super-flexible knitted yarn fabric texture features and super-flexible knitted yarn fabric depth features to obtain super-flexible knitted yarn fabric feature multi-scale joint perception coding features; and a super-flexible knitted yarn fabric detection result determination module 15, which is used to determine whether the super-flexible knitted yarn fabric to be detected has defects based on the super-flexible knitted yarn fabric feature multi-scale joint perception coding features.

[0082] In one embodiment, the image feature alignment joint perception module comprises: an image feature decoupling module for decoupling the super-soft knitted fabric texture feature coding graph and the super-soft knitted fabric depth feature coding graph to obtain a set of super-soft knitted fabric texture feature coding matrices and a set of super-soft knitted fabric depth local feature coding matrices; and an image feature fusion coding module for performing phase alignment and attention significant fusion on the set of super-soft knitted fabric texture feature coding matrices and the set of super-soft knitted fabric depth local feature coding matrices to obtain super-soft knitted fabric feature multi-scale joint perception coding features.

[0083] In one embodiment, the image acquisition system comprises a high-resolution industrial camera and an illumination system.

[0084] Here, those skilled in the art can understand that the specific operations of each module in the above image recognition-based super-soft knitted fabric defect detection system have been described in detail above with reference to the description of the image recognition-based super-soft knitted fabric defect detection method of Figures 1 to 5 , and therefore, the repeated description thereof will be omitted.

[0085] Finally, it should be noted that the above-described embodiments are part of the embodiments of the present application, rather than all the embodiments. The detailed description of the embodiments of the present application is not intended to limit the scope of the claimed application, but only represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

Claims

1. An image recognition-based defect detection method for super-soft knitted fabric, characterized in that, The method comprises the following steps: acquiring an image of a super-soft knitted gauze collected by an image collection system; performing adaptive image preprocessing and contrast enhancement on the image of the super-soft knitted gauze to obtain an enhanced image of the super-soft knitted gauze; performing multi-scale feature analysis on the enhanced image of the super-soft knitted gauze to obtain texture features of the super-soft knitted gauze and depth features of the super-soft knitted gauze; performing feature alignment joint perception on the texture features of the super-soft knitted gauze and the depth features of the super-soft knitted gauze to obtain multi-scale joint perception coding features of the super-soft knitted gauze, comprising: performing feature decoupling on texture feature coding maps and depth feature coding maps of the super-soft knitted gauze to obtain a set of texture feature coding matrices and a set of depth local feature coding matrices of the super-soft knitted gauze; and performing phase alignment and attention saliency fusion on the set of texture feature coding matrices and the set of depth local feature coding matrices to obtain the multi-scale joint perception coding features of the super-soft knitted gauze; wherein the phase alignment and attention saliency fusion on the set of texture feature coding matrices and the set of depth local feature coding matrices to obtain the multi-scale joint perception coding features of the super-soft knitted gauze comprises: based on a feature phase alignment degree between any two of the set of texture feature coding matrices and the set of depth local feature coding matrices, performing feature phase dynamic search alignment on the set of texture feature coding matrices and the set of depth local feature coding matrices to obtain a set of phase alignment {texture feature coding matrix, depth local feature coding matrix} feature pairs; inputting each phase alignment {texture feature coding matrix, depth local feature coding matrix} feature pair in the set of phase alignment {texture feature coding matrix, depth local feature coding matrix} feature pairs into a feature joint perception attention network to obtain a set of super-soft knitted gauze feature joint perception attention weights; based on the set of super-soft knitted gauze feature joint perception attention weights, performing attention-driven saliency aggregation on the set of phase alignment {texture feature coding matrix, depth local feature coding matrix} feature pairs to obtain a multi-scale joint perception coding map of the super-soft knitted gauze as the multi-scale joint perception coding features of the super-soft knitted gauze; based on the multi-scale joint perception coding features of the super-soft knitted gauze, determining whether a defect exists in a to-be-detected super-soft knitted gauze.

2. The image recognition based ultra-soft knitted fabric defect detection method according to claim 1, wherein, The image collection system comprises a high-resolution industrial camera and an illumination system.

3. The image recognition based ultra-soft knitted fabric defect detection method according to claim 2, wherein, The illumination system adopts a uniform diffuse reflection light source.

4. The image recognition based ultra-soft knitted fabric defect detection method according to claim 3, wherein, Adaptive image preprocessing and contrast enhancement are performed on the super-soft knitted gauze material image to obtain a super-soft knitted gauze material enhanced image, including: Adaptive Wiener filtering is performed on the super-soft knitted gauze material image to obtain a super-soft knitted gauze material adaptive preprocessed image; Multi-scale contrast enhancement is performed on the super-soft knitted gauze material adaptive preprocessed image to obtain the super-soft knitted gauze material enhanced image.

5. The image recognition based ultra-soft knitted fabric defect detection method as claimed in claim 4, wherein, Multi-scale feature analysis is performed on the super-soft knitted gauze material enhanced image to obtain super-soft knitted gauze material texture features and super-soft knitted gauze material depth features, including: using a multi-scale convolutional neural network to perform multi-scale image feature extraction on the super-soft knitted gauze material enhanced image to obtain a super-soft knitted gauze material texture feature encoding matrix as the super-soft knitted gauze material texture features and a super-soft knitted gauze material depth feature encoding matrix as the super-soft knitted gauze material depth features.

6. The image recognition based ultra-soft knitted fabric defect detection method as claimed in claim 1, wherein, Each phase-aligned {super-soft knitted gauze material texture feature encoding matrix, super-soft knitted gauze material depth local feature encoding matrix} feature pair in the set of phase-aligned {super-soft knitted gauze material texture feature encoding matrix, super-soft knitted gauze material depth local feature encoding matrix} feature pairs is input into a feature joint perception attention network to obtain a set of super-soft knitted gauze material feature joint perception attention weights, including: A correlation matrix between each phase-aligned {super-soft knitted gauze material texture feature encoding matrix, super-soft knitted gauze material depth local feature encoding matrix} feature pair in the set of phase-aligned {super-soft knitted gauze material texture feature encoding matrix, super-soft knitted gauze material depth local feature encoding matrix} feature pairs is calculated to obtain a set of textile surface pair scale correlation matrices; Trace measurement is performed on each textile surface pair scale correlation matrix in the set of textile surface pair scale correlation matrices to obtain a set of textile surface trace measurement values; Based on the set of textile surface trace measurement values, the set of super-soft knitted gauze material feature joint perception attention weights is obtained.

7. The image recognition based ultra-soft knitted fabric defect detection method according to claim 6, wherein, Based on the set of textile surface trace measurement values, the set of super-soft knitted gauze material feature joint perception attention weights is obtained, including: Based on the set of textile surface trace measurement values, the set of super-soft knitted gauze material feature joint perception attention weights is obtained, including: Based on the set of textile surface trace measurement values, the set of super-soft knitted gauze material feature joint perception attention weights is obtained, including:

8. An image recognition based ultra-soft knitted fabric defect detection system characterized in that, Based on the set of textile surface trace measurement values, the set of super-soft knitted gauze material feature joint perception attention weights is obtained, including: An adaptive image preprocessing and contrast enhancement module is configured to perform adaptive image preprocessing and contrast enhancement on a super-soft knitted gauze material image to obtain a super-soft knitted gauze material enhanced image, including: An adaptive Wiener filtering module is configured to perform adaptive Wiener filtering on the super-soft knitted gauze material image to obtain a super-soft knitted gauze material adaptive preprocessed image; A multi-scale contrast enhancement module is configured to perform multi-scale contrast enhancement on the super-soft knitted gauze material adaptive preprocessed image to obtain the super-soft knitted gauze material enhanced image. A multi-scale feature analysis module is configured to perform multi-scale feature analysis on the super-soft knitted gauze material enhanced image to obtain super-soft knitted gauze material texture features and super-soft knitted gauze material depth features, including: using a multi-scale convolutional neural network to perform multi-scale image feature extraction on the super-soft knitted gauze material enhanced image to obtain a super-soft knitted gauze material texture feature encoding matrix as the super-soft knitted gauze material texture features and a super-soft knitted gauze material depth feature encoding matrix as the super-soft knitted gauze material depth features. A feature joint perception attention network is configured to input each phase-aligned {super-soft knitted gauze material texture feature encoding matrix, super-soft knitted gauze material depth local feature encoding matrix} feature pair in the set of phase-aligned {super-soft knitted gauze material texture feature encoding matrix, super-soft knitted gauze material depth local feature encoding matrix} feature pairs to obtain a set of super-soft knitted gauze material feature joint perception attention weights, including: A correlation matrix between each phase-aligned {super-soft knitted gauze material texture feature encoding matrix, super-soft knitted gauze material depth local feature encoding matrix} feature pair in the set of phase-aligned {super-soft knitted gauze material texture feature encoding matrix, super-soft knitted gauze material depth local feature encoding matrix} feature pairs is calculated to obtain a set of textile surface pair scale correlation matrices; Trace measurement is performed on each textile surface pair scale correlation matrix in the set of textile surface pair scale correlation matrices to obtain a set of textile surface trace measurement values; Based on the set of textile surface trace measurement values, the set of super-soft knitted gauze material feature joint perception attention weights is obtained. Based on the set of textile surface trace measurement values, the set of super-soft knitted gauze material feature joint perception attention weights is obtained, including: Based on the set of textile surface trace measurement values, the set of super-soft knitted gauze material feature joint perception attention weights is obtained, including: Based on the set of textile surface trace measurement values, the set of super-soft knitted gauze material feature joint perception attention weights is obtained, including: Based on the set of textile surface trace measurement values, the set of super-soft knitted gauze material feature joint perception attention weights is obtained, including: An adaptive image preprocessing and contrast enhancement module is configured to perform adaptive image preprocessing and contrast enhancement on a super-soft knitted gauze material image to obtain a super-soft knitted gauze material enhanced image, including: An adaptive Wiener filtering module is configured to perform adaptive Wiener filtering on the super-soft knitted gauze material image to obtain a super-soft knitted gauze material adaptive preprocessed image; A multi-scale contrast enhancement module is configured to perform multi-scale contrast enhancement on the super-soft knitted gauze material adaptive preprocessed image to obtain the super-soft knitted gauze material enhanced image. A multi-scale feature analysis module is configured to perform multi-scale feature analysis on the super-soft knitted gauze material enhanced image to obtain super-soft knitted gauze material texture features and super-soft knitted gauze material depth features, including: using a multi-scale convolutional neural network to perform multi-scale image feature extraction on the super-soft knitted gauze material enhanced image to obtain a super-soft knitted gauze material texture feature encoding matrix as the super-soft knitted gauze material texture features and a super-soft knitted gauze material depth feature encoding matrix as the super-soft knitted gauze material depth features. A feature joint perception attention network is configured to input each phase-aligned {super-soft knitted gauze material texture feature encoding matrix, super-soft knitted gauze material depth local feature encoding matrix} feature pair in the set of phase-aligned {super-soft knitted gauze material texture feature encoding matrix, super-soft knitted gauze material depth local feature encoding matrix} feature pairs to obtain a set of super-soft knitted gauze material feature joint perception attention weights, including: A correlation matrix between each phase-aligned {super-soft knitted gauze material texture feature encoding matrix, super-soft knitted gauze material depth local feature encoding matrix} feature pair in the set of phase-aligned {super-soft knitted gauze material texture feature encoding matrix, super-soft knitted gauze material depth local feature encoding matrix} feature pairs is calculated to obtain a set of textile surface pair scale correlation matrices; Trace measurement is performed on each textile surface pair scale correlation matrix in the set of textile surface pair scale correlation matrices to obtain a set of textile surface trace measurement values; Based on the set of textile surface trace measurement values, the set of super-soft knitted gauze material feature joint perception attention weights is obtained. Based on the set of textile surface trace measurement values, the set of super-soft knitted gauze material feature joint perception attention weights is obtained, including: Based on the set of textile surface trace measurement values, the set of super-soft knitted gauze material feature joint perception attention weights is obtained, including: Based on the set of textile surface trace measurement values, the set of super-soft knitted gauze material feature joint perception attention weights is obtained, including: Based on the set of textile surface trace measurement values, the set of super-soft knitted gauze material feature joint perception attention weights is obtained, including: An adaptive image preprocessing and contrast enhancement module is configured to perform adaptive image preprocessing and contrast enhancement on a super-soft knitted gauze material image to obtain a super-soft knitted gauze material enhanced image, including: An adaptive Wiener filtering module is configured to perform adaptive Wiener filtering on the super-soft knitted gauze material image to obtain a super-soft kn The image enhancement module is used for adaptive image preprocessing and contrast enhancement on the super-soft knitted gauze cloth image to obtain a super-soft knitted gauze cloth enhanced image. The image multi-scale feature analysis module is used for multi-scale feature analysis on the super-soft knitted gauze cloth enhanced image to obtain super-soft knitted gauze cloth texture features and super-soft knitted gauze cloth depth features. The image feature alignment joint perception module is used for feature alignment joint perception on the super-soft knitted gauze cloth texture features and the super-soft knitted gauze cloth depth features to obtain super-soft knitted gauze cloth feature multi-scale joint perception coding features, wherein the image feature alignment joint perception module comprises: an image feature decoupling module, which is used for feature decoupling on super-soft knitted gauze cloth texture feature coding graphs and super-soft knitted gauze cloth depth feature coding graphs to obtain a set of super-soft knitted gauze cloth texture feature coding matrices and a set of super-soft knitted gauze cloth depth local feature coding matrices; and an image feature fusion coding module, which is used for phase alignment and attention saliency fusion on the set of super-soft knitted gauze cloth texture feature coding matrices and the set of super-soft knitted gauze cloth depth local feature coding matrices to obtain super-soft knitted gauze cloth feature multi-scale joint perception coding features. The phase alignment and attention saliency fusion on the set of super-soft knitted gauze cloth texture feature coding matrices and the set of super-soft knitted gauze cloth depth local feature coding matrices to obtain super-soft knitted gauze cloth feature multi-scale joint perception coding features comprises: based on feature phase alignment degrees between any two super-soft knitted gauze cloth texture feature coding matrices and super-soft knitted gauze cloth depth local feature coding matrices in the set of super-soft knitted gauze cloth texture feature coding matrices and the set of super-soft knitted gauze cloth depth local feature coding matrices, performing feature phase dynamic search alignment on the set of super-soft knitted gauze cloth texture feature coding matrices and the set of super-soft knitted gauze cloth depth local feature coding matrices to obtain a set of phase alignment {super-soft knitted gauze cloth texture feature coding matrix, super-soft knitted gauze cloth depth local feature coding matrix} feature pairs; inputting each phase alignment {super-soft knitted gauze cloth texture feature coding matrix, super-soft knitted gauze cloth depth local feature coding matrix} feature pair in the set of phase alignment {super-soft knitted gauze cloth texture feature coding matrix, super-soft knitted gauze cloth depth local feature coding matrix} feature pairs into a feature joint perception attention network to obtain a set of super-soft knitted gauze cloth feature joint perception attention weights; based on the set of super-soft knitted gauze cloth feature joint perception attention weights, performing attention-driven saliency aggregation on the set of phase alignment {super-soft knitted gauze cloth texture feature coding matrix, super-soft knitted gauze cloth depth local feature coding matrix} feature pairs to obtain super-soft knitted gauze cloth feature multi-scale joint perception coding graphs as the super-soft knitted gauze cloth feature multi-scale joint perception coding features. The super-soft knitted gauze fabric detection result determination module is configured to determine whether the to-be-detected super-soft knitted gauze fabric has defects based on the super-soft knitted gauze fabric feature multi-scale joint perception coding feature.

9. The image recognition based ultra-soft knitted fabric defect detection system as claimed in claim 8, wherein, The image acquisition system includes a high-resolution industrial camera and an illumination system.

Images