Multi-sensor fused cloth online flaw identification method and system

Abstract

The invention provides a multi-sensor fusion cloth online flaw identification method and system, and relates to the technical field of cloth flaw identification. The method comprises the following steps: acquiring specified maps of the front and back sides of current detected cloth; performing space-time alignment on the data in the specified atlas to obtain aligned multi-source data; key information is extracted from the multi-source data; according to the cloth type parameter of the current detected cloth, determining the fusion weight of each data in the key information; performing weighted fusion on data in the key information according to the fusion weight to generate a fusion feature vector; and comparing the fusion feature vector with the fusion feature of the normal cloth, and determining and positioning a chromatic aberration area. The method provided by the invention aims at solving the technical problem of performing color difference defect detection on the cloth which has double-sided property and light transmission and is subjected to an after-finishing process on a cloth manufacturing production line, and realizes more accurate identification and positioning of the color difference defect of the complex cloth.

Description

Technical Field

[0001] This invention relates to the field of fabric defect recognition technology, and more specifically, to a method and system for online fabric defect recognition using multi-sensor fusion. Background Technology

[0002] In fabric manufacturing enterprises, automated quality inspection is a crucial step in ensuring product compliance with standards after the dyeing and finishing processes. Existing automated inspection systems are typically deployed on fabric production lines to inspect fabrics that continuously pass through inspection areas. These systems are often equipped with visible light image sensors and spectral sensors to collect visual and spectral information from the fabric surface, with the primary goal of identifying color differences and defects.

[0003] However, in actual production, companies produce a wide variety of fabrics, some of which are double-sided, such as certain functional fabrics or jacquard fabrics, where the front and back sides may differ in color, texture, or functional coating. To ensure the overall quality of the fabric, both sides need to be inspected simultaneously or in stages. This requires the inspection system to simultaneously acquire image and spectral data from both sides of the fabric. For example, visible light image sensors and spectral sensor arrays can be placed above and below the fabric as it passes through the inspection area, or the fabric can pass through a single-sided inspection station again after passing through a flipping mechanism. Double-sided data acquisition doubles the amount of raw data and requires consideration of the spatial and temporal correlation between the front and back data, which increases the complexity of data processing.

[0004] Furthermore, some fabrics possess a degree of light transmittance, especially thin fabrics such as silk or lightweight synthetic fibers. For these fabrics, color performance depends not only on their surface's light reflection characteristics but also on the internal structure or fiber distribution after light penetrates the fabric, as well as the characteristics of light transmitted from the back of the fabric. Color differences may manifest in reflected light, transmitted light, or both. To comprehensively assess the color quality of such fabrics, the detection system needs to be able to acquire the fabric's transmission spectrum data, as well as image data, possibly combined with transmitted light illumination. This requires placing a light source below the fabric and a transmission spectral sensor and image sensor positioned above or below the fabric. Combining transmission data further increases the complexity of the data source and the difficulty of data fusion.

[0005] After dyeing and finishing, fabrics typically undergo various finishing processes, such as softening, wrinkle-resistant, waterproofing, antistatic, or functional coatings. These finishing processes utilize specific chemical auxiliaries or materials that adhere to the fabric fiber surface or penetrate into the fiber interior. These auxiliaries or materials themselves possess specific optical properties, such as refractive index, absorption spectrum, or scattering characteristics. These optical properties can superimpose or alter the spectral response of the original dyes in the fabric. For example, some waterproof coatings increase the gloss of the fabric surface and change its reflective properties; some UV-resistant finishing agents have strong absorption in the UV band; and optical brighteners emit light in the visible light region under UV excitation. When the finishing process results in localized inhomogeneities on the fabric, such as uneven coating thickness or uneven distribution of auxiliaries, it can lead to slight differences in the optical properties of localized areas compared to surrounding areas. These localized changes in optical properties caused by finishing inhomogeneities may manifest as localized gloss differences, slight color shifts, or brightness variations in visible light images, and as abnormal deviations in the spectral curve of localized areas within a specific wavelength range in spectral data. This optical anomaly may resemble the visual or spectral characteristics of a real color difference defect caused by uneven dye concentration or improper dyeing, or it may be superimposed on the real color difference, making it more difficult to distinguish.

[0006] Traditional methods for detecting color difference in fabrics typically focus only on the reflectance images and spectral data of the fabric surface, and are primarily based on the characteristics of the dyes themselves. These methods are ill-suited for fabrics with double-sidedness, translucency, or those that have undergone complex finishing processes. Simply splicing or fusioning front and back data, reflectance / transmission data, and image / spectral data with fixed weights is insufficient to accurately distinguish true color differences from optical artifacts or interference caused by fabric structure, translucency, or uneven finishing processes. For example, for translucent fabrics, transmission spectra may provide crucial information about internal structure or deep dye distribution, while reflectance spectra focus more on the surface. For finished fabrics, it is necessary to extract features from complex fused data that can isolate the effects of finishing processes and reflect only differences in dye distribution.

[0007] Therefore, innovation is needed in data acquisition, preprocessing, feature extraction, and fusion strategies to fully utilize the complementary information from multi-source heterogeneous data while suppressing the influence of various interference factors. Furthermore, fabrics may experience slight positional shifts or deformations during high-speed movement, posing a challenge to the precise spatial alignment of front and back surfaces, as well as reflective / transmittive data. The key to achieving robust fabric quality inspection lies in how to perform these complex data processing and fusion processes in real time on high-throughput production lines, and provide accurate color difference judgment and positioning. Summary of the Invention

[0008] The purpose of this invention is to provide a multi-sensor fusion method and system for online fabric defect identification. Targeting fabric manufacturing production lines, it aims to solve the technical problem of detecting color difference defects in fabrics that are double-sided, translucent, and have undergone finishing processes, thereby achieving more accurate identification and location of color difference defects in complex fabrics.

[0009] In a first aspect, the present invention provides a method for online fabric defect identification using multi-sensor fusion, comprising the following steps:

[0010] Obtain the specified patterns of both sides of the currently inspected fabric;

[0011] Spatial and temporal alignment are performed on the data in the specified map to obtain aligned multi-source data;

[0012] Extract key information from the multi-source data;

[0013] Based on the fabric type parameters of the currently detected fabric, determine the fusion weight of each data point in the key information;

[0014] The data in the key information are weighted and fused according to the fusion weight to generate a fusion feature vector;

[0015] The fused feature vector is compared with the fused feature of normal fabric. If the comparison result shows a deviation, the color difference region is determined and located by adjusting the color difference judgment threshold based on the intensity value of the post-processing interference feature in the key information.

[0016] The multi-sensor fusion-based online fabric defect identification method provided by this invention, in the context of fabric manufacturing production lines, faces stringent conditions such as high-speed operation, exponentially increased data volume, complex data sources, and various optical interferences (e.g., optical changes caused by structure, light transmittance, and uneven finishing processes) when dealing with fabrics that are double-sided, translucent, and have undergone finishing processes. It efficiently collects and accurately aligns the reflection and transmission images and spectral data of the front and back sides of the fabric. Through multi-source data layered feature extraction and fabric type-adaptive weighted fusion analysis, it effectively removes or suppresses optical interference introduced by non-dye factors, accurately identifies and locates true color difference defects caused by dyes, and meets the real-time processing requirements of the production line.

[0017] Secondly, the present invention provides a multi-sensor fusion-based online fabric defect identification system, comprising:

[0018] The acquisition module is used to acquire specified patterns from both sides of the currently inspected fabric.

[0019] The alignment module is used to perform spatial and temporal alignment on the data in the specified map to obtain aligned multi-source data;

[0020] The extraction module is used to extract key information from the multi-source data;

[0021] The determination module is used to determine the fusion weight of each data in the key information based on the fabric type parameter of the currently detected fabric;

[0022] The generation module is used to perform weighted fusion of the data in the key information according to the fusion weight to generate a fusion feature vector;

[0023] The comparison module is used to compare the fused feature vector with the fused features of normal fabric. If the comparison result shows a deviation, the color difference area is determined and located by adjusting the color difference judgment threshold based on the intensity value of the post-processing interference feature in the key information.

[0024] As can be seen from the above, the multi-sensor fusion-based online fabric defect identification method provided by this invention can effectively solve the challenges encountered by existing fabric color difference detection systems when handling complex fabrics with double-sidedness, translucency, and post-processing, especially in high-throughput production line environments. Through multi-source data acquisition, precise alignment, hierarchical feature extraction (including post-processing interference features), and weighted fusion and judgment logic based on fabric type adaptation, it achieves more accurate identification and location of complex fabric color difference defects, improves the robustness and reliability of fabric quality inspection, and meets the high-efficiency requirements of real-time processing on production lines.

[0025] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing embodiments of the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings. Attached Figure Description

[0026] Figure 1 This is a flowchart of a multi-sensor fusion method for online fabric defect identification provided in an embodiment of the present invention.

[0027] Figure 2 This is a schematic diagram of a multi-sensor fusion online fabric defect recognition system provided in an embodiment of the present invention.

[0028] Label Explanation:

[0029] 100. Acquisition Module; 200. Alignment Module; 300. Extraction Module; 400. Confirmation Module; 500. Generation Module; 600. Comparison Module. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0031] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0032] Reference Appendix Figure 1 This invention provides a multi-sensor fusion method for online fabric defect identification, comprising the following steps:

[0033] Obtain specified spectra from both sides of the currently inspected fabric; the specified spectra include reflectance image, reflectance spectrum, transmission image, and transmission spectrum;

[0034] By using image registration and time synchronization techniques, spatial and temporal alignment is performed on the data in a specified map to obtain aligned multi-source data.

[0035] Extract key information from multi-source data; key information includes color space components, local texture features, local brightness features, specific wavelength intensity ratio, spectral curve slope, spectral curve curvature, absorption peak position, absorption peak intensity, correlation features between reflectance and transmission spectra, difference features between reflectance and transmission spectra, and post-processing interference features.

[0036] Based on the fabric type parameters of the current tested fabric, determine the fusion weight of each data point in the key information;

[0037] The data in the key information are weighted and fused according to the fusion weight to generate a fusion feature vector;

[0038] The fused feature vector is compared with the fused feature of normal fabric. If the comparison result shows a deviation, the color difference area is determined and located by adjusting the color difference judgment threshold based on the intensity value of the post-processing interference feature in the key information.

[0039] A specified spectrum refers to a multimodal data set used to describe the optical properties of the fabric being tested. It can include reflectance images, reflectance spectra, transmission images, and transmission spectra, primarily to obtain comprehensive information about both sides of the fabric under different lighting conditions. Image registration and time synchronization techniques are used to correct spatial and temporal discrepancies in data collected by different sensors. Spatial alignment can be achieved using feature-matching-based, frequency-domain analysis-based, or machine learning-based image registration algorithms, while temporal alignment can be achieved using timestamp alignment or hardware-triggered synchronization. The main purpose is to ensure that data from different sensors accurately correspond to the same physical area and time on the fabric. Key information refers to the set of features extracted from aligned multi-source data to characterize fabric condition and potential defects. This includes color space components, local texture features, local brightness features, specific wavelength intensity ratios, spectral curve slope, spectral curve curvature, absorption peak position, absorption peak intensity, correlation features between reflectance and transmission spectra, differences between reflectance and transmission spectra, and post-processing interference features. Its main purpose is to extract discriminative information related to fabric defects, structural characteristics, and post-processing effects from multi-source data. Fabric type parameters describe the inherent properties of the fabric being tested, including fiber type, yarn density, weaving method, and post-processing techniques. These parameters primarily provide prior information about the fabric to guide subsequent feature processing and judgment. Fusion weights are relative importance coefficients assigned to different key information data during multi-source feature fusion. Their main purpose is to dynamically adjust the contribution of each feature in the final judgment based on fabric type and interference. Post-finishing interference features refer to characteristics used to quantify the impact of fabric finishing processes on optical properties. Their main purpose is to identify and evaluate optical artifacts caused by uneven finishing processes, enabling differentiation during defect assessment. Adjusting the color difference judgment threshold refers to dynamically changing the threshold used to determine the presence of color difference based on specific conditions. This is primarily to adaptively adjust the strictness of color difference judgment according to the intensity of post-finishing interference, avoiding misjudgments.

[0040] The working principle of this invention is as follows: First, sensors and light sources located above and below the fabric rapidly and synchronously acquire multi-source data, including reflection images, reflection spectra, transmission images, and transmission spectra from both sides of the fabric. Next, the acquired data undergoes preprocessing, including precise spatial alignment using image registration and temporal synchronization to ensure that data from different sources correspond to the same area of ​​the fabric and at the same time. Then, a layered feature extraction stage is initiated, extracting features at different levels from the raw data. This includes basic mid-level features such as optical intensity, color, texture, and spectral shape, as well as specifically identified optical interference features related to uneven finishing. In the feature fusion and judgment stage, preset parameters are consulted based on the currently detected fabric type, and the weights of different data sources (front / back, reflection / transmission, image / spectrum) and different feature types (color, texture, spectrum, finishing interference) are dynamically adjusted. These features are then weighted and fused to form a comprehensive high-level feature vector. Finally, based on this fused feature vector, it is compared with the feature patterns of normal fabric to determine whether color difference defects exist. During the judgment process, the system references the extracted post-processing interference features. If the interference features are significant, the judgment logic or threshold is adjusted to distinguish between genuine color differences caused by dyes and optical artifacts caused by uneven post-processing, fabric structure, or light transmission. This enables accurate identification and location of color difference defects and reduces false judgments. The entire process is handled in parallel in an assembly line manner to meet the real-time inspection needs of the production line.

[0041] The core innovation of this application lies in the comprehensive use of images and spectral data from both sides of the fabric, including reflection and transmission, to extract multi-dimensional key information, including post-processing interference features. The feature fusion weights and color difference judgment thresholds are dynamically adjusted according to the fabric type and the degree of post-processing interference, thereby effectively distinguishing between real color differences and optical interference caused by fabric structure, light transmittance, or uneven post-processing, thus improving the accuracy and robustness of fabric defect identification.

[0042] Specifically, this method first acquires specified spectra of both sides of the fabric being tested. These spectra include reflectance images, reflectance spectra, transmission images, and transmission spectra, providing comprehensive raw data. Next, image registration and time synchronization techniques are used to precisely align these multi-source data both spatially and temporally, ensuring that subsequent processing is based on accurately corresponding data. Then, key information is extracted from the aligned multi-source data. This information encompasses various features of the images and spectra, particularly features reflecting the relationship between reflectance and transmission data, as well as post-processing interference features. Based on this, the fusion weights of each data point in the key information are determined according to the fabric type parameter of the fabric being tested, enabling the feature fusion process to adapt to the characteristics of different fabrics. Subsequently, the extracted key information is weighted and fused according to the determined fusion weights to generate a comprehensive fusion feature vector. Finally, this fusion feature vector is compared with the fusion features of normal fabric to determine if there is a deviation. If a deviation exists, the color difference judgment threshold is dynamically adjusted based on the intensity value of the post-processing interference features in the key information, thereby determining and locating the color difference region. The entire process, through multi-source data acquisition, precise alignment, multi-dimensional feature extraction, adaptive fusion based on fabric type, and dynamic judgment based on post-processing interference, forms a complete defect identification system capable of handling complex fabric types and interference factors.

[0043] As a preferred embodiment, the solution of this application is implemented as follows: Visible light cameras, hyperspectral cameras, and transmission and reflection light sources positioned above and below the fabric can be used to acquire reflection images, reflection spectra, transmission images, and transmission spectra of both sides of the fabric. Spatial alignment can be achieved using a phase-correlation-based image registration algorithm. Deformation parameters are determined and corrected by calculating the phase correlation between the image to be registered and the standard image in the frequency domain. Time synchronization can be achieved by attaching precise timestamps to each sensor data point and matching them according to the timestamps during processing. Extraction of key information can be achieved using image processing algorithms (such as color space conversion, local binary mode (LBP), and grayscale statistics) and spectral analysis algorithms (such as spectral normalization, derivative calculation, and peak detection). The correlation / difference characteristics between the reflection and transmission spectra can be obtained by calculating their correlation coefficient, Euclidean distance, or spectral difference. Post-processing interference characteristics can be extracted by analyzing spectral anomalies or changes in image gloss within a specific wavelength range. Fabric type parameters can be pre-stored in the system and input by the operator or automatically identified by the sensors at the start of detection. The determination of fusion weights can be achieved by establishing a lookup table mapping fabric type to initial weights, and adjusting the initial weights using a piecewise linear or nonlinear function based on the intensity of post-processing interference features. The fusion feature vector can be generated using a weighted average or weighted concatenation method. The comparison step involves calculating the Euclidean or Mahalanobis distance between the fusion feature vector and the normal sample feature vector, and comparing it to a set threshold. When the comparison result exceeds the threshold, the final color difference judgment threshold is dynamically calculated based on the intensity of the post-processing interference features, by referring to a preset threshold adjustment curve or formula, and the color difference region is judged and located based on this threshold.

[0044] Through the above-described scheme, this application can effectively utilize the image and spectral information of both sides of the fabric, as well as its reflection and transmission, to comprehensively capture the optical characteristics of the fabric. Precise data alignment ensures the effective fusion of multi-source data. Extracting multi-dimensional key information, including post-processing interference features, provides a basis for distinguishing different types of optical changes. Dynamically adjusting the fusion weights and judgment thresholds based on fabric type and post-processing interference improves adaptability to different fabric types and complex interference environments, thereby significantly enhancing the accuracy and robustness of fabric defect identification. Especially when dealing with fabrics that are double-sided, translucent, or have undergone complex post-processing, it can effectively distinguish between true color differences and optical artifacts caused by structure or post-processing, reducing the false judgment rate and achieving accurate judgment and location of defects.

[0045] In some embodiments, the step of spatially aligning data in a specified atlas includes:

[0046] Based on the reflection or transmission image in the specified spectrum, the frequency domain features of the image are extracted by Fourier transform to obtain the frequency domain feature map;

[0047] Based on the preset deformation parameter range, multiple candidate deformation fields are generated; the deformation parameter range includes the rotation angle range, scaling ratio range, and translation range.

[0048] For each candidate deformation field, the frequency domain feature map is corrected using the deformation field to obtain the corrected frequency domain feature map, and the cross-correlation coefficient between the corrected frequency domain feature map and the standard frequency domain feature map is calculated.

[0049] The candidate deformation field with the largest cross-correlation coefficient is selected as the optimal deformation field, and the optimal deformation field is used to perform spatial correction on the reflection and transmission images in the specified spectrum to obtain spatially aligned multi-source data.

[0050] The Fourier transform is a mathematical tool that transforms an image from the spatial domain to the frequency domain. In the frequency domain, the global characteristics of an image, such as periodicity, directionality, and edge information, are represented by specific frequency components. By extracting frequency domain feature maps using the Fourier transform, image deformations (such as translation, rotation, and scaling) can be converted into corresponding changes in the frequency domain, facilitating subsequent analysis and correction. A preset deformation parameter range defines the boundaries of the possible deformation types and their degrees that need to be searched and considered during spatial alignment. It limits the possible values ​​of rotation angles, scaling ratios, and translations. Setting this range can narrow the search space and improve alignment efficiency. A deformation field is a mathematical model describing the transformation of an image from one state to another. Based on the preset deformation parameter range, a series of specific deformation parameter combinations can be generated, each corresponding to a candidate deformation field. These candidate deformation fields represent various deformation scenarios that the image may undergo. The cross-correlation coefficient is a measure of the similarity between two signals or images. Here, it is used to evaluate the degree of matching between the frequency domain feature map of an image corrected by a candidate deformation field and a pre-defined standard frequency domain feature map. The standard frequency domain feature map can represent the frequency domain features of the image under ideal, deformation-free conditions. The larger the cross-correlation coefficient, the more similar the corrected image is to the standard image, and the better the compensation effect of the candidate deformation field on the current image deformation. Among all generated candidate deformation fields, the deformation field that maximizes the cross-correlation coefficient between the corrected frequency domain feature map and the standard frequency domain feature map is selected as the optimal deformation field. This optimal deformation field is considered to most accurately describe the actual deformation of the current image to be aligned relative to the standard state.

[0051] The working principle of this method is as follows: First, the input reflected or transmitted image is transformed into the frequency domain. This is because certain spatial transformations (such as translation, rotation, and scaling) have specific correspondences in the frequency domain, facilitating analysis and correction. Next, based on the estimated possible deformation of the fabric, a search range for deformation parameters is set, and a series of possible deformation models, i.e., candidate deformation fields, are generated within this range. Then, for each candidate deformation field, it is applied to the frequency domain feature map of the image for correction, simulating compensation for the impact of the deformation. By calculating the cross-correlation coefficient between the corrected frequency domain feature map and a standard frequency domain feature map, the alignment effect of the deformation field can be quantified. Cross-correlation coefficients, especially those based on phase correlation, have good robustness to brightness variations and complex textures in images, making them suitable for fabric image alignment. By comparing the cross-correlation coefficients of all candidate deformation fields, the optimal deformation field is selected, which is considered to most accurately reflect the actual deformation of the image. Finally, this optimal deformation field is used to perform a spatial transformation on the original reflected and transmitted images to achieve precise alignment. Applying this precise spatial alignment method to a multi-sensor fusion fabric defect identification process ensures accurate spatial correspondence between data from different sensors (such as reflective and transmissive image sensors) and the front and back of the fabric. This accurate correspondence forms the basis for subsequent extraction of key information from multi-source data and weighted fusion, effectively avoiding misjudgments or missed detections due to alignment errors. This is particularly beneficial when dealing with fabrics with complex textures, high deformability, or translucency, significantly improving the accuracy and reliability of defect identification, especially color difference detection.

[0052] As a specific implementation method, the Fourier transform can employ the Fast Fourier Transform (FFT) algorithm, which can be efficiently implemented on a Digital Signal Processor (DSP) or Graphics Processing Unit (GPU) to meet the speed requirements of online detection. The preset deformation parameter range can be empirically set based on the fabric type and motion state on the actual production line. For example, the rotation angle range can be set to ± a few degrees, the scaling ratio range can be set to a certain interval close to 1, and the translation range can be set according to the field of view of image acquisition and the degree of fabric offset. Generating multiple candidate deformation fields can be achieved by discrete sampling within these parameter ranges. Applying the candidate deformation fields to the correction of the frequency domain feature map can involve corresponding phase adjustment, scaling, or rotation operations in the frequency domain. The cross-correlation coefficient can be calculated using phase correlation techniques. By calculating the inverse Fourier transform of the normalized cross-power spectrum of two frequency domain feature maps, the peak position indicates the translation amount, and the peak size reflects the similarity. For rotation and scaling, the frequency domain amplitude spectrum can be correlated in logarithmic polar coordinates. The standard frequency domain feature map can be pre-obtained by processing a large number of normal fabric sample images and averaging them. The selection of the optimal deformation field involves finding the parameter combination that corresponds to the maximum cross-correlation coefficient. Spatial correction of the original image using the optimal deformation field can be performed using standard image affine transformation or perspective transformation algorithms, resampling image pixels based on the parameters of the optimal deformation field.

[0053] By employing a spatial alignment method based on Fourier transform and deformation field search, this method effectively compensates for deformations such as rotation, scaling, and translation of fabric during high-speed movement, ensuring accurate spatial correspondence between reflected and transmitted images. This method utilizes the frequency domain features of the images for alignment, exhibiting better robustness to complex textures or patterns on fabric surfaces compared to traditional methods relying on spatial domain feature points, enabling accurate alignment under challenging conditions. This provides an accurate spatial foundation for subsequent feature extraction and fusion of multi-source data (reflected images, transmitted images, etc.), significantly improving the accuracy and reliability of online fabric defect identification, particularly color difference detection.

[0054] In some embodiments, the step of calculating the cross-correlation coefficient between the corrected frequency domain feature map and the standard frequency domain feature map includes:

[0055] The phase correlation between the corrected frequency domain feature map and the standard frequency domain feature map is calculated to obtain the cross-correlation coefficient. If the fabric surface is determined to have periodic texture according to the preset texture judgment algorithm, the cross-correlation coefficient is weighted to reduce the influence of periodic texture on the cross-correlation coefficient, and the weighted cross-correlation coefficient is used as the final cross-correlation coefficient.

[0056] Phase correlation refers to an image similarity measurement method based on Fourier transform. It determines the relative translation between two images by calculating the cross-correlation between the phase spectra of their Fourier transforms. This can be achieved by calculating the peak position and intensity of the inverse Fourier transform of the conjugate product of the Fourier transforms of the two images.

[0057] The preset texture judgment algorithm refers to an algorithm used to analyze image content and identify whether there are repeating or periodic structures. It can be implemented using methods based on frequency domain analysis or spatial domain analysis.

[0058] Weighting the cross-correlation coefficients refers to the process of correcting or adjusting the calculated cross-correlation values ​​according to specific conditions or rules. This can be achieved by multiplying by a coefficient less than 1 or by adjusting the weight distribution based on the intensity or direction information of the periodic texture.

[0059] Based on the aforementioned technical features, the working principle of the cross-correlation coefficient calculation method in this application is as follows: When spatially aligning data in a specified spectrum, it is necessary to calculate the cross-correlation coefficient between the corrected frequency domain feature map and the standard frequency domain feature map to evaluate the merits of different candidate deformation fields. This scheme, based on this, first uses phase correlation calculation to obtain preliminary cross-correlation values. Utilizing the robustness of phase correlation to image translation, rotation, and other deformations, the accuracy of the preliminary evaluation is improved. Furthermore, considering the potential interference of periodic textures in the fabric with the frequency domain feature map, after calculating the preliminary cross-correlation coefficient, a preset texture judgment algorithm is used to detect whether periodic textures exist on the fabric surface. If periodic textures are determined to exist, the preliminary cross-correlation coefficient is weighted. This weighting aims to weaken the excessive influence of strong peaks generated by periodic textures in the frequency domain on the cross-correlation coefficient, making the cross-correlation coefficient more accurately reflect the alignment degree of non-periodic content (i.e., the deformation of the fabric itself). In this way, even in the presence of periodic textures, the cross-correlation coefficient can more accurately evaluate the spatial correction effect. Finally, the cross-correlation coefficient, either weighted (if periodic textures exist) or unprocessed (if periodic textures do not exist), is used as the final cross-correlation coefficient to select the candidate deformation field with the highest cross-correlation coefficient as the optimal deformation field. Applying this improved cross-correlation coefficient calculation method to the spatial alignment process can more accurately determine the optimal deformation field, thereby improving the accuracy of spatial alignment of fabric image data. Especially when processing fabrics with periodic textures, it effectively avoids misjudgments caused by texture interference, providing more reliable alignment data for subsequent defect identification.

[0060] To more clearly illustrate the technical solution of this application, a specific embodiment is described in detail below. In a specific embodiment, calculating the phase correlation between the corrected frequency domain feature map and the standard frequency domain feature map can be achieved by first performing inverse Fourier transforms on the two frequency domain feature maps respectively to obtain the corresponding spatial domain images. Then, the product of the Fourier transform of one image and the conjugate of the Fourier transform of the other image is calculated, and the result is then subjected to an inverse Fourier transform to obtain the phase correlation plane. The peak positions in this phase correlation plane indicate the relative translation of the image, and the peak intensity can be used as the cross-correlation coefficient. Simultaneously, a preset texture judgment algorithm based on frequency domain analysis can be used to determine whether there is periodic texture on the fabric surface. For example, the power spectrum of the original image can be calculated, and the presence of obvious, concentrated peaks far from the origin can be analyzed. These peaks correspond to periodic structures in the image. If these peaks are detected, it is determined that there is periodic texture on the fabric surface. If periodic texture is determined to exist, the calculated cross-correlation coefficient is weighted. For example, the weighting coefficients can be determined based on the intensity of the periodic peaks detected in the power spectrum. The stronger the periodic peaks, the greater the interference of periodic textures on the frequency domain feature map. In this case, a smaller weighting coefficient can be used to multiply the initially calculated cross-correlation coefficient, or a nonlinear function can be used to adjust the cross-correlation coefficient based on the peak intensity. This weighting process reduces the influence of periodic textures on the cross-correlation coefficient, making the weighted cross-correlation coefficient more accurately reflect the alignment degree of non-periodic content. Finally, the weighted cross-correlation coefficient is used as the basis for evaluating the spatial correction effect.

[0061] By adopting the above technical solution, this application achieves the following technical effects: It utilizes the robustness to deformation by calculating phase correlation as the basis for cross-correlation coefficients. Furthermore, by judging periodic textures and performing weighted processing, the interference of periodic textures on the calculation of cross-correlation coefficients is effectively reduced. This allows the cross-correlation coefficients to more accurately reflect the degree of deformation alignment of the fabric itself, improving the accuracy and robustness of spatial alignment on fabrics with periodic textures. Accurate spatial alignment provides a reliable foundation for subsequent multi-sensor data fusion and defect identification, helping to improve the accuracy of defect identification and reduce false positives and false negatives.

[0062] In some embodiments, if it is determined that the fabric surface has periodic textures, the step of weighting the cross-correlation coefficients includes:

[0063] Obtain the direction and periodicity information of the periodic texture;

[0064] Based on the direction information, determine the frequency components perpendicular to the direction of the periodic texture;

[0065] Calculate the energy value of the frequency component based on the periodic information;

[0066] The weighting coefficients are determined based on the energy values. The larger the energy value, the smaller the weighting coefficient. The initial cross-correlation coefficients are then weighted using the weighting coefficients to obtain the weighted cross-correlation coefficients.

[0067] The directional and periodic information of periodic textures refers to parameters describing the orientation and repetition interval of the periodic structure on the fabric surface. These can be obtained by performing a two-dimensional Fourier transform on the image and analyzing the location of energy concentration regions in the frequency domain. Frequency components perpendicular to the periodic texture direction refer to frequency components orthogonal to the main direction of the periodic texture in the frequency domain. These frequency components typically correspond to variations in the texture in the vertical direction and can be selected in the frequency domain based on the obtained texture direction through coordinate transformation or sector filtering. The energy value of a frequency component refers to the sum of the squares of the intensity or amplitude represented by the image in the frequency domain within a specific frequency range. This value reflects the strength of the corresponding frequency component and can be calculated by squaring and summing the frequency domain coefficients corresponding to the selected frequency components. Weighting coefficients are multiplicative factors used to adjust the initial cross-correlation values. The magnitude of these factors is inversely related to the calculated energy values ​​of the frequency components and can be determined based on the energy values ​​using a preset mapping function or lookup table. The initial cross-correlation coefficient refers to the cross-correlation value calculated through phase correlation before performing periodic texture weighting processing. The weighted cross-correlation coefficient refers to the final cross-correlation value obtained by multiplying the initial cross-correlation coefficient by a determined weighting coefficient.

[0068] This application's solution is based on an in-depth analysis of the frequency domain characteristics of periodic textures. By acquiring the direction and periodicity information of the periodic texture, it can accurately pinpoint the frequency components most relevant to texture interference in the frequency domain. These frequency components are perpendicular to the texture direction, and their frequencies are closely related to the texture period, representing key areas of texture energy concentration. By calculating the energy values ​​of these specific frequency components, the intensity of the periodic texture can be quantified. Higher energy values ​​indicate more significant periodic textures and greater interference with cross-correlation calculations. Therefore, weighting coefficients are determined based on energy values, with larger energy values ​​resulting in smaller weighting coefficients. This inverse relationship ensures a greater suppression effect on cross-correlation coefficients affected by strongly periodic textures. Using these determined weighting coefficients to weight the initial cross-correlation coefficients can specifically weaken the weight of periodic textures in cross-correlation calculations, making the weighted cross-correlation coefficients more accurately reflect the alignment degree of the image's non-periodic features. This differs from methods that only perform uniform weighting, which cannot distinguish the characteristics and intensity of textures and may over- or under-suppress interference. By employing this adaptive weighting based on texture characteristics and intensity, this scheme can more effectively reduce the impact of periodic textures on spatial alignment accuracy, thereby improving the reliability of optimal deformation field selection and the final spatial correction accuracy. Combined with the fundamental method of spatial alignment using Fourier transform to extract frequency domain features and calculate cross-correlation coefficients, this scheme can significantly improve overall spatial alignment performance in complex fabric image processing scenarios with periodic textures.

[0069] In one embodiment, obtaining the direction and periodicity information of a periodic texture can be achieved by performing a two-dimensional Fourier transform on the input cloth image to obtain its power spectrum. In the power spectrum, the periodic texture will appear as discrete energy peaks far from the origin. By detecting the positions of these energy peaks, the main direction and period of the texture can be determined. For example, the angle of the peak relative to the origin corresponds to the texture direction, and the reciprocal of the distance from the peak to the origin is related to the texture period. Based on the obtained texture direction information, the frequency range perpendicular to that direction can be determined. For example, if the texture direction is close to horizontal, the frequency components in the frequency domain that are close to the vertical direction are selected. Based on the periodicity information, specific frequency points or frequency intervals corresponding to the texture period can be determined within the selected vertical frequency range. Calculating the energy value of the frequency components can be done by integrating or summing the squares of the amplitudes of the corresponding frequency domain coefficients within these specific frequency points or frequency intervals. Based on the calculated energy value, an inverse proportional function can be used to determine the weighting coefficients. For example, the weighting coefficients can be set to a constant divided by (1 plus a scaling factor multiplied by the energy value), ensuring that the larger the energy value, the smaller the weighting coefficients. Finally, the calculated weighted coefficients are multiplied by the initial cross-correlation coefficients to obtain the weighted cross-correlation coefficients, which are then used for the subsequent selection of the optimal deformation field.

[0070] Using the above method, this application enables targeted weighting of cross-correlation coefficients when periodic textures exist on the fabric surface. This processing method, based on quantitative analysis of the direction, period, and energy of the periodic texture, can more accurately identify and suppress the interference of texture on spatial alignment calculations. This allows for greater reliance on non-periodic features in the image that reflect the true structure or defects when selecting the optimal deformation field, thereby improving the accuracy and robustness of spatial alignment. It is particularly suitable for high-precision online detection of fabrics with complex periodic textures.

[0071] In some embodiments, the step of determining the fusion weights of each data point in the key information based on the fabric type parameter of the currently detected fabric includes:

[0072] Based on the fabric type parameters of the current tested fabric, the corresponding initial fusion weight is retrieved from the pre-established weight lookup table; the fabric type parameters include fiber type, yarn density, weaving method, and finishing process; the weight lookup table contains the mapping relationship between fabric type parameters, key information data types, and fusion weights;

[0073] Based on the intensity value of the post-processing interference features in the key information, a piecewise function is used to dynamically adjust the initial fusion weights, and the adjusted fusion weights are determined as the final fusion weights of each data in the key information.

[0074] Fabric type parameters refer to technical indicators describing the inherent properties and processing status of fabrics. They can be represented by specific classification information such as fiber type, yarn density, weaving method, and finishing processes. A weighted lookup table is a data structure that stores the mapping relationship between fabric type parameters, key information data types, and corresponding initial fusion weights. It can be implemented using a database, configuration file, or in-memory lookup table. The intensity value of finishing interference features in key information refers to a numerical value that quantifies the degree to which the fabric is affected by uneven finishing processes. It can be represented by a value obtained by analyzing and calculating the extracted finishing interference features. A piecewise function is a mathematical function with different expressions in different input intervals. It can be implemented using a series of conditional statements or a lookup table.

[0075] This solution improves the process of determining the fusion weights of key information based on fabric type parameters. It aims to address the problem that in high-speed fabric production lines with numerous and frequently changing fabric types, fixed weights based solely on fabric type parameters cannot adequately adapt to the dynamic changes in optical properties introduced by post-processing techniques, leading to poor fusion results and consequently affecting the accuracy of color difference detection. This solution dynamically determines the fusion weights by combining fabric type parameters and the intensity values ​​of post-processing interference features. This allows for fine-tuning of the initial weights, effectively adapting to the dynamic changes in optical properties introduced by post-processing techniques, improving fusion results, and thus enhancing the accuracy of color difference detection. Specifically, firstly, based on the fabric type parameters of the currently inspected fabric, the corresponding initial fusion weights are retrieved from a pre-established weight lookup table. Fabric type parameters, such as fiber type, yarn density, weaving method, and post-processing techniques, reflect the inherent characteristics of the fabric. These characteristics determine the relative importance of different types of key information (such as image features, spectral features, and reflection / transmission features) in identifying specific defects. The weight lookup table stores the mapping relationship between fabric type parameters, key information data types, and fusion weights. This allows the system to quickly obtain a preliminary weight configuration suitable for that type of fabric based on prior knowledge, providing a foundation for subsequent dynamic adjustments. Secondly, and this is a key improvement of the solution, a piecewise function is used to dynamically adjust the initial fusion weights based on the intensity values ​​of post-processing interference features in the key information. The unevenness of the post-processing process is a significant source of interference affecting the fabric's optical properties and potentially leading to misjudgments. By extracting and quantifying the intensity values ​​of post-processing interference features, the system can perceive the degree of post-processing interference in the current detection area in real time. Using a piecewise function for dynamic adjustment means that the weight adjustment strategy changes according to different intervals of interference intensity, allowing the fusion weights to adaptively respond to actual interference conditions. For example, when post-processing interference is strong, the weights of susceptible features can be reduced, while the weights of features reflecting the interference itself or less susceptible features can be increased. This effectively suppresses the impact of post-processing interference during feature fusion, making the fused feature vector more accurately reflect the true defect information. Finally, the dynamically adjusted fusion weights are determined as the final fusion weights for each data point in the key information. These final weights will be used in subsequent weighted fusion steps to generate more robust and accurate fusion feature vectors, thereby improving the accuracy and reliability of fabric defect identification, especially when faced with complex fabric types and post-processing interference.By combining the prior information provided by the fabric type parameters with the real-time information provided by the intensity of post-processing interference features, and using a piecewise function for fine-tuning, this scheme can generate fusion weights that better reflect the actual fabric condition. This allows for more effective integration of multi-source key information in subsequent weighted fusion steps, highlighting features related to real defects, suppressing interference caused by uneven post-processing, and ultimately improving the discriminative power of the fused feature vectors, thereby enhancing the overall accuracy of defect identification.

[0076] In one embodiment, based on the currently detected fabric, its fabric type parameters can be determined as "cotton, plain weave, waterproof treatment". The system queries a pre-built weight lookup table based on these parameters. This lookup table can be a structure stored in a file or database, recording the initial fusion weights for various key information such as color space components, local texture features, spectral curve slope, and finishing interference features under different combinations of fabric type parameters. For example, for a fabric of type "cotton, plain weave, waterproof treatment", the lookup table might specify an initial weight of 0.4 for the color space components, 0.3 for local texture features, and 0.2 for finishing interference features, with the remaining weights assigned to other features. After extracting the key information, the system obtains the intensity value of the finishing interference features in the currently detected area. This intensity value can be a value between 0 and 1, representing the degree of interference. The system then calls a piecewise function, which takes the initial fusion weights and the intensity value of the finishing interference features as input. This piecewise function can define multiple intensity intervals and set different weight adjustment rules for each interval. For example, when the intensity value is low, the initial weights are fine-tuned; when the intensity value is high, the weights of post-processing interference features are significantly increased, while the weights of other susceptible features are decreased. After calculation using a piecewise function, a set of adjusted fusion weights is obtained. For instance, if the intensity value of post-processing interference is high, the weight of the post-processing interference feature might be adjusted to 0.5, while the weights of color space components and local texture features are correspondingly reduced. Finally, this set of adjusted weights is determined as the final weights used for weighted fusion.

[0077] By obtaining initial fusion weights based on fabric type parameters and dynamically adjusting them using a piecewise function in conjunction with the intensity values ​​of post-processing interference features, this scheme enables the fusion weights of each data point in the key information to more accurately reflect the actual state and degree of interference of the current fabric. This overcomes the limitation of relying solely on fixed weights, which cannot adapt to the dynamic changes in fabric post-processing, improves the effectiveness of weighted fusion of key information, thereby enhancing the ability of the fused feature vector to represent real defects, and ultimately improving the accuracy and robustness of fabric color difference detection.

[0078] In some embodiments, the step of dynamically adjusting the initial fusion weights based on the intensity value of the post-processing interference features in the key information includes:

[0079] If the intensity value is greater than the preset first threshold, the initial fusion weight corresponding to the post-processing interference feature will be adjusted to the maximum value, and the initial fusion weights corresponding to the color space component, local texture feature, and local brightness feature will be reduced accordingly.

[0080] If the intensity value is less than the preset second threshold, the initial fusion weight corresponding to the post-processing interference feature is adjusted to the minimum value, and the initial fusion weights corresponding to the color space component, local texture feature, and local brightness feature are increased accordingly; the second threshold is less than the first threshold.

[0081] If the intensity value is between the first threshold and the second threshold, then based on the relative distance between the intensity value and the first threshold, and the relative distance between the intensity value and the second threshold, the initial fusion weights corresponding to the rearranged interference features, color space components, local texture features, and local brightness features are linearly adjusted.

[0082] The intensity value refers to the quantitative representation of post-processing interference features extracted from multi-source data. It reflects the degree of influence of post-processing unevenness on the optical properties of the fabric. It can be obtained by analyzing the amplitude, energy, or difference between the post-processing interference features and their corresponding features in normal fabric. The preset first and second thresholds are boundary values ​​used to divide the intensity range of post-processing interference, guiding the weight adjustment strategy under different intensities. They can be determined based on statistical analysis of the intensity distribution of post-processing interference features from a large number of fabric samples or expert experience. The maximum and minimum values ​​refer to the upper and lower limits that the initial fusion weights corresponding to the post-processing interference features, color space components, local texture features, and local brightness features can reach after adjustment. They can be set based on the theoretical importance of different features to the final defect judgment and the actual test results. The relative distance refers to the difference or proportional relationship between the currently detected post-processing interference feature intensity value and the first or second threshold. It is used for linear interpolation calculation when the intensity value is between the two thresholds. It can be obtained using simple difference calculation or normalized difference calculation. Linear adjustment refers to using linear interpolation or other linear function relationships to calculate the adjusted fusion weight when the intensity value is between the first threshold and the second threshold, so that the weight presents a smooth linear transition with the change of intensity value. It can be achieved by performing linear interpolation calculation based on the position of the intensity value in the interval [second threshold, first threshold].

[0083] This scheme dynamically adjusts the initial fusion weights based on the intensity of post-processing interference features using a piecewise function, thereby achieving adaptive processing of post-processing interference of varying degrees. Specifically, when the detected intensity of post-processing interference features is high (greater than a preset first threshold), it indicates that the post-processing interference has a significant impact on the fabric's optical properties. In this case, the weight of the post-processing interference feature itself is adjusted to its maximum value, while the weights of color space components, local texture features, and local brightness features, which are significantly affected by post-processing interference, are correspondingly reduced. This is done to focus more on identifying and quantifying the interference itself in a strong interference environment, while reducing the weights of other features affected by the interference, thus avoiding misjudgments caused by feature distortion due to interference. Conversely, when the intensity of post-processing interference features is low (less than a preset second threshold), it indicates that the post-processing interference has a smaller impact on the fabric's optical properties. In this case, the weight of the post-processing interference feature is adjusted to its minimum value, while the weights of color space components, local texture features, and local brightness features are correspondingly increased. This approach aims to leverage features like color, texture, and brightness—which better reflect the fabric's inherent characteristics—for judgment in weak interference environments, reducing the noise impact potentially introduced by minor post-processing interference features. When the intensity of the post-processing interference feature falls between the first and second thresholds, it indicates a moderate level of interference. At this point, based on the relative distance between the intensity value and the two thresholds, linear interpolation is used to smoothly adjust the weights of the post-processing interference feature, color space components, local texture features, and local brightness features. This linear adjustment makes weight changes more continuous and refined, more accurately reflecting the relative importance of each feature under different interference intensities, and avoiding abrupt weight changes near the thresholds. This strategy of dynamically adjusting weights based on the intensity of post-processing interference is a further optimization based on the initial weights determined according to the fabric type. It ensures that the weights consider not only the inherent properties of the fabric but also the actual detected level of local interference. In this way, the system can intelligently adjust the fusion weights of different key information based on the actual detected intensity of post-processing interference, enabling the final fused feature vector to more effectively remove interference from uneven post-processing and more accurately reflect the true defect information of the fabric.

[0084] For example, in one embodiment, a preset first threshold can be set to 0.8, and a preset second threshold can be set to 0.2. The maximum value of the post-processing interference feature weight can be set to 1.0, and the minimum value can be set to 0.0. The sum of the initial weights of the color space component, local texture feature, and local brightness feature can be set to W. When the detected post-processing interference feature intensity value is 0.9, since it is greater than the first threshold of 0.8, the system can adjust the weight of the post-processing interference feature to 1.0, and correspondingly reduce the weights of the color space component, local texture feature, and local brightness feature. When the detected intensity value is 0.1, since it is less than the second threshold of 0.2, the system can adjust the weight of the post-processing interference feature to 0.0, and correspondingly increase the weights of the color space component, local texture feature, and local brightness feature. When the detected intensity value is 0.5, since it is between 0.2 and 0.8, the system can perform linear interpolation based on the relative distance between the intensity value of 0.5 and the thresholds 0.8 and 0.2. For example, the weight of post-processing interference features can be adjusted to 0.0 + (1.0 - 0.0) * (0.5 - 0.2) / (0.8 - 0.2) = 0.5. Then, the remaining weight (W - 0.5) is proportionally allocated to the color space components, local texture features, and local brightness features. Through this specific numerical setting and adjustment logic, fine-grained weight control of post-processing interference of different intensities can be achieved.

[0085] By dynamically adjusting the initial fusion weights using a piecewise function based on the intensity of post-processing interference features, this scheme can intelligently adjust the fusion weights of different key information according to the actual detected intensity of post-processing interference. This allows the final generated fusion feature vector to more effectively remove interference caused by uneven post-processing and more accurately reflect the true defect information of the fabric, especially improving the accuracy and robustness of color difference judgment.

[0086] Reference Appendix Figure 2 This invention provides a multi-sensor fusion-based online fabric defect identification system, comprising:

[0087] The acquisition module 100 is used to acquire specified patterns on both sides of the currently detected fabric.

[0088] Alignment module 200 is used to perform spatial and temporal alignment on data in a specified map to obtain aligned multi-source data.

[0089] Extraction module 300 is used to extract key information from multi-source data;

[0090] The determination module 400 is used to determine the fusion weight of each data in the key information based on the fabric type parameters of the current detected fabric.

[0091] The generation module 500 is used to perform weighted fusion of the data in the key information according to the fusion weight to generate a fusion feature vector.

[0092] The comparison module 600 is used to compare the fused feature vector with the fused feature of normal fabric. If the comparison result shows a deviation, the color difference area is determined and located by adjusting the color difference judgment threshold based on the intensity value of the post-processing interference feature in the key information.

[0093] In this document, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, without necessarily requiring or implying any such actual relationship or order between these entities or operations.

[0094] The above description is merely an embodiment of the present invention and is not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for online fabric defect identification using multi-sensor fusion, characterized in that, Includes the following steps: Obtain the specified patterns of both sides of the currently inspected fabric; Spatial and temporal alignment are performed on the data in the specified map to obtain aligned multi-source data; Extract key information from the multi-source data; Based on the fabric type parameters of the currently detected fabric, determine the fusion weight of each data point in the key information; The data in the key information are weighted and fused according to the fusion weight to generate a fusion feature vector; The fused feature vector is compared with the fused feature of normal fabric. If the comparison result shows a deviation, the color difference region is determined and located by adjusting the color difference judgment threshold based on the intensity value of the post-processing interference feature in the key information.

2. The multi-sensor fusion method for online fabric defect identification according to claim 1, characterized in that, The specified spectra include reflectance images, reflectance spectra, transmission images, and transmission spectra.

3. The multi-sensor fusion method for online fabric defect identification according to claim 2, characterized in that, The steps for spatially aligning the data in the specified map include: Based on the reflection or transmission image in the specified spectrum, the frequency domain features of the image are extracted by Fourier transform to obtain a frequency domain feature map; Based on the preset deformation parameter range, multiple candidate deformation fields are generated; For each candidate deformation field, the frequency domain feature map is corrected using the deformation field to obtain the corrected frequency domain feature map, and the cross-correlation coefficient between the corrected frequency domain feature map and the standard frequency domain feature map is calculated. The candidate deformation field with the largest cross-correlation coefficient is selected as the optimal deformation field, and the optimal deformation field is used to perform spatial correction on the reflection and transmission images in the specified spectrum to obtain spatially aligned multi-source data.

4. The multi-sensor fusion method for online fabric defect identification according to claim 3, characterized in that, The deformation parameter range includes the rotation angle range, scaling ratio range, and translation range.

5. The multi-sensor fusion method for online fabric defect identification according to claim 3, characterized in that, The steps for calculating the cross-correlation coefficient between the corrected frequency domain feature map and the standard frequency domain feature map include: The phase correlation between the corrected frequency domain feature map and the standard frequency domain feature map is calculated to obtain the cross-correlation coefficient. If it is determined that there is a periodic texture on the fabric surface, the cross-correlation coefficient is weighted and the weighted cross-correlation coefficient is used as the final cross-correlation coefficient.

6. The multi-sensor fusion method for online fabric defect identification according to claim 5, characterized in that, If it is determined that the fabric surface has periodic textures, the step of weighting the cross-correlation coefficients includes: Obtain the direction and periodicity information of the periodic texture; Based on the direction information, determine the frequency components perpendicular to the direction of the periodic texture; Calculate the energy value of the frequency component based on the periodic information; The weighting coefficients are determined based on the energy values, and the initial cross-correlation coefficients are weighted using these coefficients to obtain the weighted cross-correlation coefficients.

7. The multi-sensor fusion method for online fabric defect identification according to claim 2, characterized in that, The key information includes color space components, local texture features, local brightness features, specific wavelength intensity ratio, spectral curve slope, spectral curve curvature, absorption peak position, absorption peak intensity, correlation features between reflection and transmission spectra, difference features between reflection and transmission spectra, and post-processing interference features.

8. The multi-sensor fusion method for online fabric defect identification according to claim 7, characterized in that, The steps for determining the fusion weights of each data point in the key information based on the fabric type parameters of the currently detected fabric include: Based on the fabric type parameters of the current detected fabric, the corresponding initial fusion weight is retrieved from the pre-established weight lookup table; Based on the intensity value of the post-processing interference features in the key information, the initial fusion weights are dynamically adjusted, and the adjusted fusion weights are determined as the final fusion weights of each data in the key information.

9. The multi-sensor fusion method for online fabric defect identification according to claim 8, characterized in that, The step of dynamically adjusting the initial fusion weights based on the intensity value of the post-processing interference features in the key information includes: If the intensity value is greater than the preset first threshold, the initial fusion weight corresponding to the post-processing interference feature is adjusted to the maximum value, and the initial fusion weights corresponding to the color space component, local texture feature, and local brightness feature are reduced accordingly. If the intensity value is less than the preset second threshold, the initial fusion weight corresponding to the post-processing interference feature is adjusted to the minimum value, and the initial fusion weights corresponding to the color space component, local texture feature, and local brightness feature are increased accordingly; the second threshold is less than the first threshold; If the intensity value is between the first threshold and the second threshold, then based on the relative distance between the intensity value and the first threshold, and the relative distance between the intensity value and the second threshold, the initial fusion weights corresponding to the rearranged interference features, color space components, local texture features, and local brightness features are linearly adjusted.

10. A multi-sensor fusion online fabric defect identification system, characterized in that, include: The acquisition module is used to acquire specified patterns from both sides of the currently inspected fabric. The alignment module is used to perform spatial and temporal alignment on the data in the specified map to obtain aligned multi-source data; The extraction module is used to extract key information from the multi-source data; The determination module is used to determine the fusion weight of each data in the key information based on the fabric type parameter of the currently detected fabric; The generation module is used to perform weighted fusion of the data in the key information according to the fusion weight to generate a fusion feature vector; The comparison module is used to compare the fused feature vector with the fused features of normal fabric. If the comparison result shows a deviation, the color difference area is determined and located by adjusting the color difference judgment threshold based on the intensity value of the post-processing interference feature in the key information.

Images