Waste cloth classification recycling method based on machine vision

Abstract

The invention relates to the technical field of waste cloth classification and recovery, in particular to a waste cloth classification and recovery method based on machine vision. The method comprises the following steps: acquiring hyperspectral image data, three-dimensional point cloud data and polarized light image data of waste cloth to be classified, respectively constructing a hyperspectral feature map, a three-dimensional geometric feature map and a polarized optical feature map based on the data, and respectively inputting the hyperspectral feature map, the three-dimensional geometric feature map and the polarized optical feature map into a classification network to generate multi-modal discrimination features after weight fusion, and outputting a classification result of the waste cloth through a full-connection classification layer at the tail end of the classification network. According to the scheme provided by the invention, comprehensive information representation can be performed on the waste cloth from multiple dimensions such as material components, physical wrinkles and textures, yarn weaving and glossiness and the like, and internal association among materials, geometric and optical attributes is effectively excavated; therefore, the comprehensive classification accuracy and reliability of the material components, the fabric structure and the recoverable grade of the waste cloth are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of waste cloth classification and recycling, and more particularly to a waste cloth classification and recycling method based on machine vision. Background Art

[0002] With the continued growth of global textile consumption and the prevalence of fast fashion, the recycling and reuse of waste textiles has become a critical issue in environmental protection and resource recycling. Traditional waste fabric sorting relies primarily on manual labor, which is labor-intensive, inefficient, subjective, and inaccurate. This makes it difficult to meet the demands of large-scale, refined recycling and processing. Therefore, the development of automated and intelligent waste fabric sorting technologies is an inevitable trend in the industry. In recent years, machine vision technology has been introduced to the field of fabric sorting. For example, using visible light (RGB) images combined with deep learning networks can classify fabric color and pattern, but it struggles to accurately identify material composition. Near-infrared (NIR) spectroscopy or hyperspectral imaging can distinguish different fiber materials (such as cotton, linen, polyester, and wool) based on their spectral absorption characteristics in specific wavelengths, achieving significant progress in material identification. However, relying solely on spectral information is susceptible to surface stains, dyes, moisture, and wrinkles on the fabric, and cannot effectively capture physical properties such as fabric structure and texture, limiting the precision and reliability of classification.

[0003] To overcome the limitations of single-sensor technology, researchers have begun exploring multimodal information fusion. For example, visible light images are combined with near-infrared spectral information to simultaneously obtain color and material information. However, existing multimodal fusion strategies mostly remain at the level of simple feature concatenation or decision-level fusion. Feature concatenation (early fusion) directly concatenates feature vectors from different modalities, failing to fully consider the inherent differences and complementarities between modal data. This can easily affect overall performance due to information redundancy or noise in a particular modality. Decision-level fusion (late fusion) processes and classifies each modal data independently before combining the results. This approach ignores the deep-level interactions between different modal features, resulting in the loss of a significant amount of valuable cross-modal information. This is particularly true for complex items such as waste fabrics, where properties such as material composition, yarn weave (e.g., plain weave, twill), surface gloss, physical wrinkles, and texture are intertwined and collectively determine their ultimate recyclability and reuse value. Therefore, existing technologies still have obvious technical bottlenecks in how to deeply and adaptively integrate multi-source heterogeneous information such as spectrum, three-dimensional geometry and surface optics to achieve a comprehensive and accurate assessment of the material, structure and quality of waste fabrics. Summary of the Invention

[0004] The purpose of the present invention is to propose a method for sorting and recycling waste fabrics based on machine vision to solve the problem of poor classification accuracy and reliability of waste fabrics in the prior art during sorting and recycling. To this end, the present invention provides a solution in the following aspect.

[0005] The present invention provides a method for sorting and recycling waste fabrics based on machine vision, comprising: Acquire hyperspectral image data, three-dimensional point cloud data and polarized light image data of waste fabrics to be classified; based on the hyperspectral image data, perform wavelet packet transform on the spectral response curve of each pixel, extract the energy and entropy value within the preset fiber material characteristic response frequency band, and construct a pixel-level hyperspectral feature map; based on the three-dimensional point cloud data, calculate the normal vector and curvature of each data point, and project them to generate a three-dimensional geometric feature map representing the wrinkles and texture of the fabric; based on the polarized light image data, calculate the Stokes parameters at different polarization angles, and derive the linear polarization degree and curvature of each pixel point. Polarization angle, construct a polarization optical feature map characterizing the weave and gloss of the fabric yarn; the hyperspectral feature map, three-dimensional geometric feature map and polarization optical feature map are respectively input into a classification network including a parallel convolution backbone and a cross-modal attention fusion module, and the cross-modal attention fusion module performs interactive attention calculation on the deep feature maps extracted by each backbone to generate weighted fused multimodal discriminant features; using the multimodal discriminant features, the classification results of the material composition, fabric structure and recyclability grade of the waste cloth are output through the fully connected classification layer at the end of the classification network.

[0006] Compared with existing technologies, the machine vision-based waste fabric classification and recycling method provided by this invention simultaneously acquires and fuses three heterogeneous data types: hyperspectral, three-dimensional point cloud, and polarized light. This method can comprehensively characterize waste fabrics from multiple dimensions, including material composition, physical wrinkles and texture, yarn weave, and gloss. This overcomes the shortcomings of existing technologies that rely on a single sensor, resulting in insufficient information dimensions and inability to accurately assess fabric structure and quality. By employing a specific network structure that includes a cross-modal attention fusion module, deep features from different modalities are interactively calculated and weightedly fused, effectively exploring the inherent connections between material, geometric, and optical properties, generating more discriminative multimodal features. This improves the accuracy and reliability of the comprehensive classification of waste fabric material composition, fabric structure, and recyclability, providing solid technical support for achieving refined, high-value recycling and reuse.

[0007] Preferably, the step of constructing a pixel-level hyperspectral feature map includes: decomposing the spectral response curve of each pixel using wavelet packet transform; for each preset fiber material type, extracting its energy and entropy value in one or more characteristic response frequency bands as characteristic indicators; combining the characteristic indicators extracted for all preset fiber material types to construct a multi-channel hyperspectral feature map.

[0008] Extracting the energy and entropy values ​​within a specific characteristic response frequency band for each preset fiber material type can ensure that the characteristic indicators are highly correlated with the material properties, thereby improving the distinctiveness and pertinence of the features.

[0009] Preferably, the step of generating a three-dimensional geometric feature map characterizing the wrinkles and texture of the fabric includes: for each data point in the point cloud, searching for its K nearest neighbor point sets, fitting the nearest neighbor point sets using the principal component analysis algorithm PCA, and calculating the covariance matrix; the eigenvector corresponding to the minimum eigenvalue of the covariance matrix is ​​the normal vector of the data point; based on the three eigenvalues ​​of the covariance matrix, the surface curvature of the data point is calculated; the normal vector XYZ components and curvature values ​​of all data points are used to generate a normal vector map and a curvature map aligned with the hyperspectral image by orthographic projection, which together constitute a three-dimensional geometric feature map.

[0010] Preferably, the step of generating a three-dimensional geometric feature map characterizing the wrinkles and texture of the fabric includes: for each data point in the point cloud, selecting all its neighboring points within a radius of 5 mm, and calculating the normal vector of the point by a principal component analysis method; based on the positional relationship between the normal vector and the neighboring points, calculating the average curvature of the point; and ortho-projecting the z component and average curvature value of the normal vector of all data points onto a two-dimensional grid of a preset resolution to form a dual-channel three-dimensional geometric feature map.

[0011] By calculating the normal vector for each data point through principal component analysis and calculating the mean curvature based on the positional relationships of neighboring points, the local geometric features of the fabric surface can be accurately captured. Furthermore, by selecting neighboring points within a 5mm radius for analysis, macroscopic textures reflecting the yarn level can be captured, balancing computational efficiency with the integrity of local geometric information. The combination of the normal vector and mean curvature effectively characterizes differences in fabric structure and surface topography, enhancing the classification network's ability to distinguish between different fabric types and recycling grades.

[0012] Preferably, the step of constructing a polarization optical characteristic diagram characterizing the weave and glossiness of the fabric yarn comprises: obtaining four polarized light images I0, I1 at the angles of 0°, 45°, 90°, and 135° of the polarizer. 45 , I 90 , I 135 According to the formula , , , calculate the Stokes parameters S0, S1, S2 of each pixel; use the formula Calculate the degree of linear polarization DoLP using the formula The polarization angle AoLP is calculated; the calculated linear polarization degree and polarization angle are used as two channels to form a polarization optical characteristic map.

[0013] Preferably, the step of constructing a polarization optical characteristic diagram characterizing the weave and glossiness of the fabric yarn comprises: obtaining three polarized light images I0, I1 at the angles of 0°, 45°, and 90° of the polarizer. 45 , I 90 According to the formula , , , calculate the Stokes parameters S0, S1, S2 of each pixel; according to the formula Calculate the degree of linear polarization DoLP; according to the formula Calculate the polarization angle AoLP; combine the DoLP value and the AoLP value of each pixel into a dual-channel image as the polarization optical characteristic map.

[0014] Preferably, the classification network includes: setting three parallel, weight-unshared convolutional neural networks as backbones, respectively receiving the hyperspectral feature map, three-dimensional geometric feature map and polarization optical feature map; each backbone independently performs deep feature extraction on the input feature map, and outputs a deep feature map from the end of each backbone.

[0015] The non-shared weight design ensures that each backbone network independently learns a feature extraction model appropriate for its input modality (feature map data), avoiding interference between data from different modalities and improving the pertinence and accuracy of feature extraction. Furthermore, the fusion of features from different modalities through parallel convolutional neural networks and cross-modal attention enhances the classification network's adaptability and discriminative capabilities for complex scenarios.

[0016] Preferably, the convolutional neural network adopts the ResNet-18 architecture.

[0017] Preferably, the steps of the cross-modal attention fusion module include: combining the deep feature maps of the three modalities of hyperspectral, three-dimensional geometry, and polarization optics in pairs, calculating the cross-attention weights respectively, and obtaining an enhanced feature map of each modality after being enhanced by the information of the other two modalities; splicing or summing the original three deep feature maps with the corresponding three enhanced feature maps to generate the weighted fused multimodal discriminant features.

[0018] By calculating cross-attention weights in pairwise combinations, the deep feature maps of hyperspectral, 3D geometric, and polarization optical modalities can be mutually enhanced, fully utilizing the complementary information of each modality and improving the comprehensive expressive power of features. The original deep feature map is concatenated or summed with the enhanced feature map to generate multimodal discriminant features, preserving the original information while incorporating cross-modal interaction results, thereby enhancing feature integrity and robustness.

[0019] Preferably, the step of outputting the classification results of the material composition, fabric structure and recyclable grade of the waste cloth includes: inputting the multimodal discriminant features into a fully connected classification layer with three parallel output heads, each output head corresponding to the classification tasks of material composition, fabric structure and recyclable grade; the material composition classification head outputs the probability that the cloth belongs to cotton, polyester or wool; the fabric structure classification head outputs the probability that the cloth belongs to plain weave, twill or satin; the recyclable grade classification head outputs the probability that the cloth belongs to grade one, grade two or grade three.

[0020] The beneficial effects of the present invention are as follows: Compared with the existing technology, the machine vision-based waste fabric classification and recycling method provided by the present invention can comprehensively characterize waste fabrics from multiple dimensions, such as material composition, physical wrinkles and texture, yarn weave and gloss, by simultaneously acquiring and fusing three heterogeneous data types: hyperspectral, three-dimensional point cloud, and polarized light. This overcomes the defects of the existing technology that rely on only a single sensor, resulting in insufficient information dimensions and inability to accurately evaluate fabric structure and quality. By adopting a specific network structure including a cross-modal attention fusion module, deep features of different modalities are interactively calculated and weightedly fused, effectively exploring the intrinsic correlations between material, geometric, and optical properties, and generating more discriminative multimodal features, thereby improving the comprehensive classification accuracy and reliability of waste fabric material composition, fabric structure, and recyclable grade, and providing solid technical support for achieving refined and high-value recycling and reuse. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 The flowchart schematically shows the steps of the waste cloth classification and recycling method based on machine vision in this embodiment. DETAILED DESCRIPTION

[0022] The technical solutions in the embodiments of the present invention will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present invention.

[0023] like Figure 1 As shown, the intelligent security image recognition method based on deep learning in this embodiment includes the following steps: Step S1: Acquire hyperspectral image data, three-dimensional point cloud data, and polarized light image data of waste fabrics to be classified.

[0024] Specifically, the waste fabrics to be sorted are placed on a uniform-speed conveyor belt and sequentially pass through an integrated multi-sensor acquisition system. A push-broom hyperspectral camera is positioned vertically above the system, scanning line by line to acquire hyperspectral cube data in the 400 to 1700 nanometer band. A line laser 3D profilometer based on structured light technology is deployed to acquire 3D point cloud data of the fabric surface. Finally, an industrial camera with a rotating linear polarizer is deployed. By rotating the polarizer to 0, 45, 90, and 135 degrees and exposing them separately, four grayscale images with different polarization angles are acquired as polarized light image data.

[0025] Step S2: Based on the hyperspectral image data, wavelet packet transform is performed on the spectral response curve of each pixel to extract the energy and entropy values ​​within the preset fiber material characteristic response frequency band to construct a pixel-level hyperspectral feature map.

[0026] Specifically, wavelet packet transform is used to decompose the spectral response curve of each pixel; for each preset fiber material type, its energy and entropy values ​​in one or more characteristic response frequency bands are extracted as characteristic indicators; the characteristic indicators extracted for all preset fiber material types are combined to construct a multi-channel hyperspectral feature map.

[0027] The wavelet packet transform decomposes the continuous spectral signal from visible light to near-infrared light captured by each pixel into distinct sub-bands, revealing more precisely the differences in the responses of different materials within specific bands. For example, the spectral curve data of a single pixel was decomposed into eight bands. Research has found that the energy response of pure cotton in the third band is significantly higher than that of polyester, while polyester has an even stronger energy response in the fifth band. Entropy measures the complexity or uncertainty of a spectrum within a specific band. For example, the spectral entropy of blended fabrics is generally higher than that of pure fabrics.

[0028] Assume that three fiber materials are preset: cotton, polyester, and wool. For each pixel, the energy and entropy of its spectrum within the eight characteristic frequency bands that are most closely correlated with the standard spectra of cotton, polyester, and wool are calculated. This results in a six-dimensional feature vector for each pixel, with the six dimensions representing cotton energy, cotton entropy, polyester energy, polyester entropy, wool energy, and wool entropy, respectively. Combining the six-dimensional feature vectors of all pixels in the entire image forms a six-channel hyperspectral feature map, where each channel represents the strength of the pixel's association with a specific material.

[0029] In step S3, based on the three-dimensional point cloud data, the normal vector and curvature of each data point are calculated, and the normal vector and curvature are projected to generate a three-dimensional geometric feature map representing the wrinkles and texture of the fabric.

[0030] Specifically, for each point in the three-dimensional point cloud, its K nearest neighbor point set is searched, for example, K is 20, and the principal component analysis algorithm PCA is used to fit the neighbor point set to calculate the covariance matrix; the eigenvector corresponding to the minimum eigenvalue of the covariance matrix is ​​the normal vector of the data point; based on the three eigenvalues ​​of the covariance matrix, the surface curvature of the data point is calculated; the normal vector XYZ components and curvature values ​​of all data points are used to generate a normal vector map and a curvature map aligned with the hyperspectral image through orthographic projection, which together constitute a three-dimensional geometric feature map.

[0031] In an optional embodiment, the step of constructing a three-dimensional geometric feature map includes: for each data point in the point cloud, selecting all its neighboring points within a radius of 5 mm, and calculating the normal vector of the point by a principal component analysis method; based on the positional relationship between the normal vector and the neighboring points, calculating the average curvature of the point; and ortho-projecting the z component and average curvature value of the normal vector of all data points onto a two-dimensional grid of a preset resolution to form a dual-channel three-dimensional geometric feature map.

[0032] Specifically, a neighborhood with a radius of 5 mm was selected to capture macrotexture that reflects the yarn level. Principal component analysis analyzes the three-dimensional coordinate distribution of the nearest points around a point to find the direction with the smallest variance. This direction is the normal direction of the microscopic surface on which the point lies. For a flat plain weave fabric, the normal vectors of most data points will be approximately perpendicular to the fabric plane, with a z-component value close to 1. However, for twill fabric with an undulating structure, the normal vectors will be deflected at the yarn interlacing points, and the z-component value will decrease accordingly.

[0033] Specifically, the mean curvature describes the degree of curvature of the local surface at a point. For example, on a flat yarn surface, the mean curvature of a point is close to zero; however, at the edges or bumps where yarns interweave, the surface is more curved and the mean curvature value is correspondingly higher. Finally, the two scalar values—the z component of the normal vector and the mean curvature of all data points—are projected onto a two-dimensional grid of, for example, 512 by 512 pixels. This results in a two-channel image: the first channel reflects the flatness and orientation of the fabric surface, while the second channel highlights the texture and roughness of the fabric, together forming a detailed description of the fabric's three-dimensional structure.

[0034] Step S4: Based on the polarized light image data, the Stokes parameters at different polarization angles are calculated, and the linear polarization degree and polarization angle of each pixel are derived to construct a polarization optical characteristic map that characterizes the weave and gloss of the fabric yarn.

[0035] Specifically, four polarized light images I0, I1 are obtained at the polarizer angles of 0°, 45°, 90°, and 135°. 45 , I90 , I 135 According to the formula , , , calculate the Stokes parameters S0, S1, S2 of each pixel; use the formula Calculate the degree of linear polarization DoLP using the formula The polarization angle AoLP is calculated. Based on the linear polarization degree and polarization angle information of each pixel, a linear polarization degree map and a polarization angle map are constructed respectively, and they are fused as two channels to construct a polarization optical feature map.

[0036] In an optional embodiment, the step of constructing a polarization optical characteristic diagram comprises: obtaining three polarized light images I0, I1 at the angles of 0°, 45°, and 90° of the polarizer; 45 , I 90 According to the formula , , , calculate the Stokes parameters S0, S1, S2 of each pixel; according to the formula Calculate the degree of linear polarization DoLP; according to the formula Calculate the polarization angle AoLP; combine the DoLP value and the AoLP value of each pixel into a dual-channel image as the polarization optical characteristic map.

[0037] Specifically, the different reflection characteristics of polarized light on the surfaces of different materials are utilized. For example, the reflected light of smooth artificial fibers such as polyester has strong polarization characteristics, while the reflected light of rough and porous natural fibers such as cotton tends to be non-polarized diffuse reflection. By rotating the polarizer to obtain images at three angles of 0 degrees, 45 degrees, and 90 degrees, the complete information of the change of light intensity with polarization direction can be captured and mathematically expressed by the Stokes parameter. S0 represents the total light intensity, and S1 and S2 describe the differences in polarization components in the horizontal and vertical directions and the 45-degree direction, respectively.

[0038] The degree of linear polarization DoLP is a value between 0 and 1 that directly quantifies the proportion of polarization components in the light. For example, the DoLP value of a polyester fiber pixel may be as high as 0.8, while the DoLP value of a cotton fiber pixel may be only 0.2. The angle of polarization AoLP indicates the main vibration direction of polarized light, which is closely related to the direction of the yarn on the surface of the fabric. For example, for plain fabric with distinct warp and weft, its AoLP image will show two regular, nearly vertical angle distributions. Using the DoLP and AoLP values ​​of each pixel as two channels, a polarization optical characteristic map is synthesized. The resulting polarization optical characteristic map can effectively distinguish the smoothness of the fiber and the microscopic directionality of the fabric.

[0039] In step S5, the hyperspectral feature map, the three-dimensional geometric feature map, and the polarization optical feature map are respectively input into a classification network comprising a parallel convolution backbone and a cross-modal attention fusion module. The cross-modal attention fusion module performs interactive attention calculation on the deep feature maps extracted by each backbone to generate weighted fused multimodal discriminant features.

[0040] Specifically, three parallel, weight-unshared convolutional neural networks are set as the backbone, which respectively receive hyperspectral feature maps, three-dimensional geometric feature maps, and polarization optical feature maps; each backbone independently performs deep feature extraction on the input feature map and outputs a deep feature map from the end of each backbone.

[0041] The deep feature maps of the three modalities of hyperspectral, three-dimensional geometry, and polarization optics are combined in pairs, and the cross-attention weights are calculated respectively to obtain the enhanced feature map of each modality after being enhanced by the information of the other two modalities; the original three deep feature maps and the corresponding three enhanced feature maps are spliced ​​or summed to generate the multimodal discriminant features after the weighted fusion.

[0042] The reason for using three independent, non-weighted backbone networks is that the data content and structure of the three feature maps are very different. The multiple channels of the hyperspectral feature map represent chemical composition responses, the dual channels of the 3D geometric feature map represent physical morphology, and the dual channels of the polarization optical feature map represent surface optical properties. Having a dedicated network, such as a ResNet-18 architecture, learn the spectral absorption peak patterns in the hyperspectral data, while having another independent ResNet-18 network learn the texture and curvature patterns in the 3D geometric data, ensures that each network optimally learns the unique features of the corresponding modality, avoiding confusion and interference between the features of different modalities.

[0043] Taking a hyperspectral feature map as an example, its size may be 512 by 512 by 6. After being input into the first backbone network, it undergoes a series of convolution, activation, and pooling operations, gradually reducing its spatial size while increasing the number of channels, ultimately outputting a deep feature map of size 16 by 16 by 512. The output feature map is no longer the raw pixel-level energy and entropy values, but instead contains more abstract and discriminative information such as the material mixing ratio and the presence of chemical additives. Similarly, the other two backbone networks convert the input 3D geometric image and polarization optical image into their own 16 by 16 by 512 deep feature maps, respectively encoding high-level semantic information about the fabric structure and surface finish.

[0044] The cross-attention mechanism achieves information alignment and complementarity between modalities. For example, when analyzing hyperspectral feature maps, the model uses information from the 3D geometric feature map as guidance. If the 3D geometric feature map shows a distinct twill structure in a certain area, the cross-attention mechanism will increase the weight of spectral features associated with common denim dyes in the corresponding area of ​​the hyperspectral feature map. In this way, 3D geometric information helps hyperspectral analysis more accurately focus on key chemical components. By performing cross-attention calculations on the hyperspectral features with the 3D geometric and polarization optical features, a hyperspectral feature enhanced by both geometric and optical information is obtained.

[0045] This process is performed on all three modalities. For example, the original hyperspectral deep feature map A, after cross-attention calculation with the three-dimensional geometric and polarization optical features, generates an enhanced hyperspectral feature map A'. The original feature map A and the enhanced feature map A' are spliced ​​on the channel to form a more comprehensive hyperspectral representation with doubled feature dimensions. The same operation is performed on the three-dimensional geometric and polarization optical modalities to obtain their respective enhanced representations. These three enhanced, doubled-dimensional feature maps are spliced ​​again or fused through a small convolutional network to ultimately form a single multimodal discriminant feature that contains strong correlations between all modalities, providing an extremely rich and reliable basis for subsequent classification tasks.

[0046] Step S6: using the multimodal discriminant features, the fully connected classification layer at the end of the classification network outputs the classification results of the material composition, fabric structure and recyclability of the waste cloth.

[0047] Specifically, the multimodal discriminant features are input into a fully connected classification layer with three parallel output heads, each output head corresponding to the classification tasks of material composition, fabric structure and recyclability level; the material composition classification head outputs the probability that the fabric belongs to a preset material such as cotton, polyester, wool, etc.; the fabric structure classification head outputs the probability that the fabric belongs to a preset structure such as plain, twill, satin, etc.; the recyclability level classification head outputs the probability that the fabric belongs to a preset level such as primary, secondary, tertiary, etc.

[0048] Adopting a multi-head output structure is an efficient multi-task learning strategy. The fused multimodal discriminant features contain comprehensive information about various fabric properties. These features are simultaneously fed into three independent classification heads, each consisting of one or two fully connected layers and a softmax activation function. This allows different decision boundaries to be learned for different tasks. For example, the material composition classification head focuses on the combination of hyperspectral and polarization optical features, while the fabric structure classification head prioritizes three-dimensional geometric features. This design enables the network to simultaneously solve three related but distinct classification problems within a unified framework. For example, for a piece of polyester-cotton blended twill fabric submitted for inspection, after the previous processing, the final multimodal discriminant features are fed into the output layer. The material composition classification head might output a probability distribution, such as cotton 0.6, polyester 0.4, and wool 0.0, indicating a cotton-polyester blend. The fabric structure classification head might output plain 0.1, twill 0.9, and satin 0.0, clearly indicating a twill texture. The recyclability classification head considers the blending composition and possible dyeing conditions, and may output a grade of 0.1 for Grade 1, 0.85 for Grade 2, and 0.05 for Grade 3. Ultimately, the system classifies the fabric based on the highest probability for each head: cotton-polyester blend, twill, and Grade 2 recyclable.

[0049] In the description of this specification, “a plurality of” means at least two, for example, two, three or more, etc., unless otherwise clearly defined.

[0050] Although this specification has shown and described several embodiments of the present invention, it is obvious to those skilled in the art that such embodiments are provided by way of example only. Those skilled in the art will conceive of many modifications, changes and substitutions without departing from the idea and spirit of the present invention.

Claims

1. A method for recycling waste fabrics based on machine vision, characterized in that: include: Obtaining hyperspectral image data, three-dimensional point cloud data, and polarized image data of waste fabrics to be sorted; Based on the hyperspectral image data, a wavelet packet transform is performed on the spectral response curve of each pixel to extract the energy and entropy values ​​within a preset fiber material characteristic response frequency band, and a pixel-level hyperspectral feature map is constructed; Based on the three-dimensional point cloud data, the normal vector and curvature of each data point are calculated, and the normal vector and curvature are projected to generate a three-dimensional geometric feature map representing the wrinkles and texture of the fabric; Based on the polarized light image data, the Stokes parameters at different polarization angles are calculated, and the linear polarization degree and polarization angle of each pixel are derived to construct a polarization optical characteristic map representing the weave and gloss of the fabric yarn; The hyperspectral feature map, the three-dimensional geometric feature map, and the polarization optical feature map are respectively input into a classification network comprising a parallel convolution backbone and a cross-modal attention fusion module. The cross-modal attention fusion module performs interactive attention calculation on the deep feature maps extracted by each backbone to generate weighted fused multimodal discriminant features. By utilizing the multimodal discriminant features, a classification result of the material composition, fabric structure and recyclability grade of the waste cloth is output through a fully connected classification layer at the end of the classification network.

2. The method for recycling waste fabrics based on machine vision according to claim 1, characterized in that: The step of constructing a pixel-level hyperspectral feature map includes: Wavelet packet transform is used to decompose the spectral response curve of each pixel; for each preset fiber material type, the energy and entropy values ​​in one or more characteristic response frequency bands are extracted as characteristic indicators; the characteristic indicators extracted for all preset fiber material types are combined to construct a multi-channel hyperspectral feature map.

3. The method for recycling waste fabrics based on machine vision according to claim 1, characterized in that: The step of generating a three-dimensional geometric feature map representing cloth wrinkles and textures includes: For each data point in the point cloud, its K nearest neighbor point sets are searched, and the principal component analysis algorithm PCA is used to fit the nearest neighbor point sets to calculate the covariance matrix; the eigenvector corresponding to the minimum eigenvalue of the covariance matrix is ​​the normal vector of the data point; based on the three eigenvalues ​​of the covariance matrix, the surface curvature of the data point is calculated; the normal vector XYZ components and curvature values ​​of all data points are used to generate a normal vector map and a curvature map aligned with the hyperspectral image through orthographic projection, which together constitute a three-dimensional geometric feature map.

4. The method for recycling waste fabrics based on machine vision according to claim 1, characterized in that: The step of generating a three-dimensional geometric feature map representing cloth wrinkles and textures includes: For each data point in the point cloud, all its neighboring points within a radius of 5 mm are selected, and the normal vector of the point is calculated by the principal component analysis method; based on the positional relationship between the normal vector and the neighboring points, the average curvature of the point is calculated; the z component and average curvature value of the normal vector of all data points are orthographically projected onto a two-dimensional grid of a preset resolution to form the dual-channel three-dimensional geometric feature map.

5. The method for recycling waste fabrics based on machine vision according to claim 1, characterized in that: The step of constructing a polarization optical characteristic graph characterizing the weave and glossiness of the fabric yarn comprises: Get four polarized light images I0, I2 at the polarizer angles of 0°, 45°, 90°, and 135° 45 , I 90 , I 135 According to the formula , , , calculate the Stokes parameters S0, S1, S2 of each pixel; use the formula Calculate the degree of linear polarization DoLP using the formula The polarization angle AoLP is calculated; the calculated linear polarization degree and polarization angle are used as two channels to form a polarization optical characteristic map.

6. The method for recycling waste fabrics based on machine vision according to claim 1, characterized in that: The step of constructing a polarization optical characteristic graph characterizing the weave and glossiness of the fabric yarn comprises: Obtain three polarized light images I0, I2 at the polarizer angles of 0°, 45°, and 90°. 45 , I 90 According to the formula , , , calculate the Stokes parameters S0, S1, S2 of each pixel; according to the formula Calculate the degree of linear polarization DoLP; according to the formula Calculate the polarization angle AoLP; combine the DoLP value and the AoLP value of each pixel into a dual-channel image as the polarization optical characteristic map.

7. The method for recycling waste fabrics based on machine vision according to claim 1, characterized in that: The classification network includes: Three parallel, weight-unshared convolutional neural networks are set as backbones to receive the hyperspectral feature map, the three-dimensional geometric feature map, and the polarization optical feature map respectively; each backbone independently performs deep feature extraction on the input feature map and outputs a deep feature map from the end of each backbone.

8. The method for recycling waste fabrics based on machine vision according to claim 7, characterized in that: The convolutional neural network adopts the ResNet-18 architecture.

9. The method for recycling waste fabrics based on machine vision according to claim 1, characterized in that: The steps of the cross-modal attention fusion module include: The deep feature maps of the three modalities of hyperspectral, three-dimensional geometry, and polarization optics are combined in pairs, and the cross-attention weights are calculated respectively to obtain the enhanced feature map of each modality after being enhanced by the information of the other two modalities; the original three deep feature maps and the corresponding three enhanced feature maps are spliced ​​or summed to generate the multimodal discriminant features after the weighted fusion.

10. The method for recycling waste fabrics based on machine vision according to claim 1, characterized in that: The step of outputting the classification results of the material composition, fabric structure and recycling grade of the waste fabrics includes: The multimodal discriminant features are input into a fully connected classification layer with three parallel output heads, each of which corresponds to the classification tasks of material composition, fabric structure and recyclability level; the material composition classification head outputs the probability that the fabric belongs to cotton, polyester or wool; the fabric structure classification head outputs the probability that the fabric belongs to plain weave, twill or satin; and the recyclability level classification head outputs the probability that the fabric belongs to grade one, grade two or grade three.

Images