Fabric pleat retentivity intelligent visual rating method

Abstract

The invention discloses an intelligent visual rating method for fabric pleat retentivity. The intelligent visual rating method comprises the following steps: S1, acquiring a fabric sample image; s2, preprocessing the fabric image; s3, feature point extraction and matching; s4, camera calibration and image correction; s5, performing three-dimensional reconstruction on the fabric; s6, fabric appearance pleat rating; by adopting machine vision and image analysis technologies, the parallax and depth of all points in the effective area of the fabric sample can be quickly calculated, the spatial point cloud information of the fabric is obtained, the three-dimensional reconstruction of the apparent form of the fabric is realized, the problems of interference of patterns, textures and the like of objective rating of the pleat appearance of the fabric and the like are solved; therefore, the grade of the fabric pleats can be evaluated more efficiently and accurately, and the blank of fabric appearance detection rating standards in the textile field in China is filled up.

Description

Technical Field

[0001] The invention belongs to the field of fabric detection, and in particular relates to an intelligent visual rating method for fabric pleat retention. Background Art

[0002] The pleats of fabrics are both functional and aesthetic, showing the layering and three-dimensional sense of clothing, which is particularly important in the evaluation of shirts, pleated skirts, trousers and other products. As clothing is affected by clothing care such as wearing, washing and drying, the pleat effect of the fabric will weaken, which will have a negative impact on the functionality and aesthetics of the clothing. Accurate rating of the pleat appearance retention of fabrics can provide an evaluation basis for clothing companies in production, which is an important guarantee for product quality control and improvement. At present, the rating method of fabric pleat retention is mostly subjective evaluation method, which is time-consuming and labor-intensive, with low precision and poor repeatability. It is easily affected by the subjective factors of the rater. Therefore, there is an urgent need for a fabric pleat retention detection technology with high efficiency, small rating result differences, good repeatability and precision.

[0003] The invention patent with patent number CN2023117073102 discloses an automatic rating method for the appearance retention level of fabric pleats. It uses two-dimensional images to detect fabric pleats, mainly by extracting surface features such as grayscale area, shadow area, and contour of the image, and comparing and converting the pleat height, thereby realizing the detection and evaluation of pleats.

[0004] The invention patent with patent number CN2021104904857 discloses a three-dimensional reconstruction system and method for complex fabric surfaces, which includes: a box, a base, a pressing unit, and an image processing module. In the published literature,

[0005] Fabric pleat grade assessment based on photometric stereo vision [Liu Ruixin, Li Liqing, Wang Jun, Chen Xia. Journal of Donghua University, 2013, 39(1): 48-53.] A method for assessing fabric pleat grade using photometric stereo vision algorithm is proposed. Multiple fabric images are taken under illumination from 8 different directions to obtain a two-dimensional grayscale image. The height of each point on the fabric surface is reconstructed through a differential operation, and the fabric surface shape information is obtained by three-dimensional reconstruction. The height value and sharpness are extracted through experiments to characterize the pleat features, so as to objectively evaluate the grade of fabric pleats.

[0006] Objective evaluation method of fabric pleat grade [Han Yanna. Wool Textile Technology, 2019, 47(4): 21-25.], attempts to use laser measurement technology to objectively evaluate fabric pleats. The specific steps are as follows: AATCC fabric pleat template and fabric pleats are scanned by a 3D laser scanner. After preprocessing the scanned data with GEOMAGIC software, the pleat feature values ​​of the template and fabric are extracted respectively. Then, the relationship between the pleat grade of the template and the characteristic parameters is established through multivariate linear regression, and the regression equation is applied to the objective evaluation of the fabric pleat grade.

[0007] Fabric structure and appearance analysis based on multi-eye stereo vision [Wang Yiliang.] applies artificial intelligence technologies such as multi-eye stereo vision, digital image processing, and machine learning algorithms to the evaluation of textile appearance performance and structural performance analysis. The proposed method uses a stereo vision algorithm to reconstruct the three-dimensional morphology of the fabric surface, and then extracts spatial features on the three-dimensional model to analyze and identify the fabric appearance and structure, effectively avoiding the interference of the texture and color of the fabric surface on the objective evaluation, which not only improves the detection accuracy, but also increases the applicability of the algorithm. It provides an objective and accurate solution for the appearance performance evaluation and organizational structure identification of complex fabrics.

[0008] Although the currently published patents and related research involve methods for detecting fabric pleats, there is still a lack of in-depth discussion on methods for detecting pleats of fabrics with complex patterns, textures and multiple colors, and the core problems faced in practical applications have not been effectively solved. Summary of the invention

[0009] The purpose of the present invention is to provide an intelligent visual rating method for fabric pleat retention, which adopts machine vision and image analysis technology to realize three-dimensional reconstruction of fabric surface morphology, solve the problem of interference of pattern, texture, etc. in objective rating of fabric pleat appearance, so as to more efficiently and accurately grade fabric pleats and fill the gap in fabric appearance inspection and rating standards in my country's textile field.

[0010] In order to solve the above technical problems, the following technical solutions are adopted

[0011] An intelligent visual rating method for fabric pleat retention, characterized by comprising the following steps:

[0012] S1: fabric sample image acquisition: including fabric sample processing and using an image acquisition device to acquire an image of the processed fabric sample to obtain a fabric image;

[0013] S2: fabric image preprocessing;

[0014] S3: Feature point extraction and matching: After the fabric image is preprocessed, feature points are extracted from the fabric image. The detected feature points include corner points, edges and significant texture points. These feature points are used for subsequent matching and 3D reconstruction processes.

[0015] S4: Camera calibration and image correction: Zhang Zhengyou camera calibration method is used to collect two-dimensional images of multiple fabric samples to calculate the internal and external parameters of the camera; through camera calibration, the relative position and internal parameter matrix between the two cameras are obtained, and image correction is performed on this basis;

[0016] S5: Fabric 3D reconstruction: Convolutional neural network is used to reconstruct fabric 3D, including fusion extraction based on multi-scale features, rough matching based on Euclidean distance, purification to achieve precise matching, and acquisition of fabric 3D point cloud, so as to describe the microscopic morphology and undulation of the fabric surface and realize 3D reconstruction of the fabric;

[0017] S6: Fabric appearance pleat rating: With the bottom surface of the stage as the reference in space, the point cloud height information is converted to the brightness of the image to form a two-dimensional depth map, in which the higher part is the pleat, and the height of each point is converted to the brightness of the image. The higher the height, the brighter the brightness, and the fabric sample is rated for pleats.

[0018] After optimization, the step S1, fabric sample processing comprises:

[0019] (1) Sample cutting: Cut three fabric samples of the same size from the fabric to be tested;

[0020] (2) Pleat setting: Fold the cut fabric sample in half and use an iron to press a stable pleat in the middle of each fabric sample;

[0021] (3) Fabric sample processing: Wash, dry and adjust the humidity of the fabric samples, and finally perform shaping treatment on the processed fabric samples.

[0022] After optimization, in step S1, the image acquisition device includes a box, the box is used to place the fabric sample, and its bottom plate serves as a stage, and when the box is closed, a space is formed inside the box that is not disturbed by external light sources; a camera and a light source are installed in the box, two cameras are provided for acquiring images of the fabric samples, and the light source is used for providing illumination; the camera is connected to a computer via a data cable.

[0023] After optimization, the step S2, fabric image preprocessing includes:

[0024] (1) Valid area marking: Mark the valid area of ​​the collected fabric image, retain the content of the marked area, and assign the rest of the area to 0;

[0025] (2) Grayscale processing;

[0026] (3) Image denoising;

[0027] (4) Image enhancement: Use the image histogram to optimize the denoised fabric image, enhance the image contrast, and make the fabric surface features more obvious;

[0028] (5) Sharpening;

[0029] (6) Light balance: Use white balance technology to adjust the light balance.

[0030] After optimization, the step S4 comprises:

[0031] (1) Camera calibration: Using a chessboard as a calibration object, capture template images at different angles and orientations. Then, perform corner detection on the template image to detect feature points in the calibration object. Then, calculate the camera’s intrinsic and extrinsic parameters based on the projection information of these feature points to achieve camera calibration.

[0032] (2) Image correction: First, a distortion model and a data fitting objective function are established. Then, the corner point detection method is used to obtain the coordinates of the corner points on the template image. Then, the coordinates of the corresponding corner points are calculated using the established distortion model. Finally, these two sets of coordinates are brought into the objective function formula for data fitting and the corresponding distortion parameter values ​​are calculated to eliminate camera distortion and complete image correction. The distortion model is as follows:

[0033]

[0034] In the formula, (x i ,y i ) is the actual coordinate of the target corner point, (X i ,Y i ) are the normalized coordinates of the target corner points.

[0035] After optimization, the step S5 is based on the fusion extraction of multi-scale features: by using multiple convolutional layers to extract features of the left and right fabric images, constructing related feature pyramids, and forming multi-level feature maps; fusing high-level feature maps with low-level feature maps to obtain image features, specifically: adding the feature map of the lowest layer in the feature pyramid network to the adjacent upper layer to form a new feature map, and performing the same operation on the upper layer, so as to fuse all the information in the pyramid network together and realize the fusion of high and low-level feature maps.

[0036] After optimization, the step S5, rough matching by Euclidean distance: rough matching is performed on the feature map extracted in the above step, specifically: a key point of any fabric image is selected in the left and right fabric images, and the first two key points with the closest Euclidean distance to the corresponding fabric image are found; among these two key points, if the closest distance divided by the second closest distance is less than a certain ratio threshold, the pair of matching points is accepted; wherein if the Euclidean distance is d, the threshold is set to T=αd, 0<α<1.

[0037] After optimization, the step S5 is to purify and realize accurate matching: after rough matching, the obtained matching point set is purified by repeated testing and continuous iteration to realize accurate matching.

[0038] After optimization, the step S5, obtaining the three-dimensional point cloud of the fabric: after obtaining the matching points corresponding to the left and right fabric images, the parallax of all points in the image is obtained in combination with the camera calibration, the depth is obtained by parallax calculation, and then the spatial point cloud information of the image is obtained.

[0039] After optimization, the step S6 comprises:

[0040] (1) Data input and preprocessing: First, adjust the input data format to make it consistent with the training data format, and then input the fabric image spatial point cloud information into the trained classification model;

[0041] (2) Grade prediction: The classification model calculates the pleat grade of the fabric based on the input spatial point cloud information, and the output result is a rating from 1 to 5;

[0042] (3) Result comparison and verification: Compare the pleat level predicted by the classification model with the pleat level of the actual sample; use the real pleat data as a reference to test whether the prediction of the classification model is consistent with the actual situation;

[0043] (4) Accuracy optimization: By comparing different classification models, the nearest neighbor algorithm with the highest recognition rate is used for pleat grade evaluation; when pleat rating, the classification model is verified multiple times;

[0044] (5) Result output: Through the above steps, the classification model will output the pleat level of each fabric sample and classify it into levels 1 to 5.

[0045] The above technical solution has the following beneficial effects:

[0046] (1) To solve the problem of interference of pattern and color in the objective evaluation of the appearance of fabric pleats, the present invention uses feature pyramid to realize the fusion extraction of multi-scale features. The extracted data is screened by threshold method to obtain some erroneous matching points. Then, the screened erroneous matching points are purified by repeated testing and continuous iteration to achieve accurate matching. Finally, accurate spatial point cloud information is obtained. These methods can effectively perform three-dimensional digitization on fabrics with complex patterns, colors and textures, thereby solving the problem of interference of pattern, color and texture in the objective evaluation of the appearance of fabric pleats.

[0047] (2) Automatically grading the pleat level. The present invention realizes the automatic grading of fabric pleat levels. The nearest neighbor algorithm (KNN) with the highest recognition rate is used to automatically grade the reconstructed fabric pleats to achieve accuracy optimization. The pleat level calculated by the model is compared with the pleat level of the actual sample to ensure the validity of the test results. Finally, the accurate and reliable pleat level of each sample is output and classified from 1 to 5. The automated process ensures the consistency and reliability of the rating results, while improving the detection efficiency, providing great convenience for clothing companies to grade fabric pleats.

[0048] (3) The present invention proposes a method for intelligently rating the pleat retention of fabrics by using three-dimensional reconstruction of the fabric surface morphology, accurately calculating the pleat grade of the fabric, and providing an evaluation basis for clothing companies in production. Objective, fast, and accurate machine vision and image analysis technologies show good application prospects. Using machine vision technology, the parallax and depth of all points in the effective area of ​​the fabric sample can be quickly calculated to obtain the spatial point cloud information of the fabric, thereby realizing the three-dimensional reconstruction of the appearance of the fabric pleats. Using image analysis technology, combined with the reconstruction of the three-dimensional morphology of the appearance of the fabric pleats, an intelligent rating of the fabric pleat retention can be achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The present invention will be further described below in conjunction with the accompanying drawings:

[0050] Figure 1 A flowchart of the rating method of the present invention;

[0051] Figure 2 It is a structural schematic diagram of the image acquisition device.

[0052] in, Figure 2 The reference numerals are: box 1, fabric sample 2, folding door 3, camera 4, light source 5. DETAILED DESCRIPTION

[0053] The present invention aims to provide an intelligent visual rating method for fabric pleat retention, which uses machine vision and image analysis technology to achieve three-dimensional reconstruction of fabric surface morphology, solve the problem of interference such as pattern and texture in objective rating of fabric pleat appearance, and thus more efficiently and accurately rate fabric pleats, filling the gap in fabric appearance detection and rating standards in my country's textile field. The technical solution of the present invention is described in detail below in conjunction with specific embodiments:

[0054] S1: Fabric sample image acquisition

[0055] (1) Fabric sample processing

[0056] Sample cutting: Cut three samples of the same size from the fabric to be tested according to relevant standards.

[0057] Pleat setting: Fold the cut samples in half and use an iron to press a stable pleat in the middle of each sample.

[0058] Sample processing: Wash, dry and adjust the humidity of the sample according to relevant regulations. If there are some obvious wrinkles in the sample in addition to the set pleats, iron them properly to avoid affecting the pleats. Finally, the processed textile fabrics are shaped to ensure the stability of the sample.

[0059] (2) Fabric image acquisition:

[0060] The image acquisition device is used to acquire images of the processed fabric sample 2. The image acquisition device can filter out interference from the external light source 5, receive uniform illumination, and present the fabric surface information more clearly. It also supports operations such as shooting, storing, and transmitting images, and provides image data for subsequent processing, which is an important guarantee for achieving fabric pleat appearance rating. The specific structure of the device is as follows ( Figure 2 shown):

[0061] a. Box 1: Box 1 is made of metal sheet, closed on all sides and back, and folding door 3 on the front to facilitate placement and removal of fabric sample 2. The bottom plate also serves as a fabric stage, so a stronger and smoother metal material should be selected to ensure that the fabric sample 2 remains flat and does not deform during the collection process, avoiding measurement errors caused by deformation of the stage. The inner wall of box 1 is neutral gray, forming diffuse reflection, reducing the impact of reflected light from the inner wall on image collection.

[0062] b. Camera 4: Two stereo cameras 4 with a high resolution of more than 2 million pixels are used as core imaging devices to clearly and accurately capture the left and right views. To adapt to different lighting conditions, the stereo camera 4 also needs to be equipped with automatic focus and exposure adjustment functions. The camera 4 is connected to the computer via a data cable, and the computer performs operations such as shooting, storing and transmitting images, and provides image data for subsequent processing.

[0063] c. Light source 5: A light source 5 with high illumination efficiency is selected and installed above the box 1 to ensure that the light can evenly cover the entire fabric surface for arrangement, thereby avoiding the error influence on image acquisition caused by uneven illumination.

[0064] During the fabric image acquisition process, two stereo cameras 4 with exactly the same internal parameters are used to acquire images from the same horizontal viewing angle. When shooting, it is necessary to ensure that the camera 4 can completely capture the entire fabric sample 2 and the surface of the fabric sample 2 is evenly illuminated to obtain a clear and complete fabric image for subsequent inspection and rating.

[0065] S2: Fabric image preprocessing:

[0066] After image acquisition, the following preprocessing steps are required to ensure the accuracy and effectiveness of subsequent analysis:

[0067] (1) Effective area marking: In order to focus on analyzing the pleat information on the fabric surface, it is necessary to mark the effective area of ​​the image captured by the stereo camera, retain the content of the marked area, and assign the rest to 0 to prevent the rest from interfering with subsequent results.

[0068] (2) Grayscale processing: Grayscale the marked image to simplify the image information, reduce the amount of calculation for subsequent processing, and improve processing efficiency.

[0069] (3) Image denoising: Image denoising is an indispensable part of image preprocessing. Image filtering algorithms are usually used to suppress noise in sample images. Commonly used image filtering algorithms include mean, median and Gaussian filtering. This step can effectively retain image details, thereby improving the accuracy of fabric appearance pleat rating results.

[0070] (4) Image enhancement: Use the image histogram to optimize the denoised image and enhance the image contrast, making the surface features of the fabric more obvious.

[0071] (5) Sharpening: In order to enhance image details, sharpening is also required after image enhancement. Sharpening is to enhance the edges of the image. Usually, an operator with rotation invariance is used to sharpen the image to meet the requirements of image edge sharpening in different directions.

[0072] (6) Lighting balance: To ensure uniform lighting of the image, white balance technology is used to adjust the lighting balance to reduce the impact of lighting changes on fabric feature extraction and make subsequent feature point matching more stable.

[0073] S3: Feature point extraction and matching:

[0074] After preprocessing, feature points are extracted from the image. The feature points usually detected include corner points, edges and significant texture points. These feature points will be used in the subsequent matching and 3D reconstruction process to ensure that the details and texture information of the fabric surface are fully captured.

[0075] S4: Camera calibration and image correction:

[0076] The present invention adopts Zhang Zhengyou's camera calibration method to calculate the internal and external parameters of the camera by collecting two-dimensional images of multiple fabric samples. Through calibration, the relative position and internal parameter matrix between the two cameras are obtained, and image correction is performed on this basis to eliminate image distortion caused by lens distortion. This correction step ensures the geometric accuracy of the image and is a reliable data basis for subsequent three-dimensional reconstruction.

[0077] (1) Camera calibration: Using a chessboard as a calibration object, take template images at different angles. Then perform corner detection on the template image to detect feature points in the calibration object. Then, the camera’s intrinsic and extrinsic parameters are calculated based on the projection information of these feature points to achieve camera calibration.

[0078] (2) Image correction: Since the distortion of the camera lens will cause image distortion, the image needs to be corrected. The main process is as follows: first, establish a distortion model and a data fitting objective function, then use the corner point detection method to obtain the coordinates of the corner points on the template image, and then use the established distortion model to calculate the coordinates of the corresponding corner points. Finally, these two sets of coordinates are brought into the objective function for data fitting and the corresponding distortion parameter values ​​are calculated to eliminate camera distortion and complete image correction. The distortion model is as follows:

[0079]

[0080] In the formula, (x i ,y i ) is the actual coordinate of the target corner point, (X i ,Y i ) are the normalized coordinates of the target corner points.

[0081] S5: Fabric 3D Reconstruction

[0082] After image processing and correction, convolutional neural network is used to reconstruct the fabric in three dimensions. The specific method includes fusion extraction based on multi-scale features, rough matching based on Euclidean distance, purification to achieve precise matching, and acquisition of fabric three-dimensional point cloud, so as to accurately describe the microscopic morphology and undulation of the fabric surface, and finally realize the three-dimensional reconstruction of the fabric.

[0083] (1) Fusion extraction based on multi-scale features: By using multiple convolutional layers to extract features from the left and right images, a related feature pyramid is constructed to form a multi-level feature map. The low-level feature map contains rich details such as edges and textures, while the high-level feature map processed by multiple convolutional layers contains more semantic information that helps to rate the folds. In order to make the feature map contain more and richer information and achieve more accurate ratings, the high-level feature map is fused with the low-level feature map to obtain image features. The specific method is as follows: the feature map of the lowest layer in the feature pyramid network is added to the adjacent upper layer to form a new feature map, and the same operation is performed on the upper layer, so that all the information in the pyramid network is fused together to achieve the fusion of high- and low-level feature maps and increase the accuracy of subsequent steps.

[0084] (2) Coarse matching based on Euclidean distance: Coarse matching is performed on the feature maps extracted in the above steps. The steps are as follows: Select a key point in either image and find the first two key points with the closest Euclidean distance to the corresponding image. Among these two key points, if the closest distance divided by the next closest distance is less than a certain ratio threshold, then this pair of matching points is accepted. If the Euclidean distance is d, then the threshold is set to T = αd, 0 < α < 1.

[0085] (3) Purification to achieve accurate matching To improve the calculation accuracy, it is necessary to eliminate some erroneous matching points in the matching point set after rough matching to ensure the accuracy of the reconstruction effect. Therefore, the matching point set is purified by repeated testing and continuous iteration to achieve accurate matching.

[0086] (4) Acquisition of fabric three-dimensional point cloud: After obtaining the matching points corresponding to the left and right images by the above method, the parallax of all points in the image can be obtained by combining with camera calibration. The depth is obtained by parallax calculation, and then the spatial point cloud information of the image is obtained.

[0087] S6: Fabric appearance pleats rating

[0088] Since the camera is parallel to the plane of the fabric sample, the bottom surface of the stage is used as the reference in space, and the point cloud height information is converted to the brightness of the image to form a two-dimensional depth map, in which the higher part is the fold, and the height of each point is converted to the brightness of the image. The higher the height, the brighter the brightness. In this way, the folds of the fabric sample are graded. The specific steps are as follows:

[0089] (1) Data input and preprocessing: First, adjust the input data format to make it consistent with the training data format to improve efficiency and reduce errors. Then input the spatial point cloud information of the fabric sample into the trained classification model.

[0090] (2) Grade prediction: The classification model calculates the pleat grade of the fabric based on the input spatial point cloud information, and the output result is a rating from 1 to 5.

[0091] (3) Result comparison and verification: The pleat level predicted by the classification model is compared with the pleat level of the actual sample to ensure the effectiveness of the classification model. The real pleat data is used as a reference to test whether the prediction of the classification model is consistent with the actual situation.

[0092] (4) Accuracy optimization: In order to achieve a pleat rating accuracy of 0.1, in the feature extraction stage, the matching points are purified to achieve precise matching to obtain accurate point cloud information of the pleats on the fabric surface. By comparing different classification models, the nearest neighbor algorithm (KNN) with the highest recognition rate is used for pleat rating evaluation. When rating pleats, the model is verified multiple times to ensure its stability and accuracy.

[0093] (5) Result output: Through the above steps, the classification model will output the accurate and reliable pleat level of each sample and classify it into levels 1 to 5.

[0094] The above are only specific embodiments of the present invention, but the technical features of the present invention are not limited thereto. Any simple changes, equivalent replacements or modifications made based on the present invention to solve basically the same technical problems and achieve basically the same technical effects are all included in the protection scope of the present invention.

Claims

1. An intelligent visual rating method for fabric pleat retention, characterized in that The steps include: S1: Fabric sample image acquisition: including fabric sample processing and using an image acquisition device to acquire an image of the processed fabric sample to obtain a fabric image; S2: fabric image preprocessing; S3: Feature point extraction and matching: After the fabric image is preprocessed, feature points are extracted from the fabric image. The detected feature points include corner points, edges and significant texture points. These feature points are used for subsequent matching and 3D reconstruction processes. S4: Camera calibration and image correction: Zhang Zhengyou camera calibration method is used to collect two-dimensional images of multiple fabric samples to calculate the internal and external parameters of the camera; through camera calibration, the relative position and internal parameter matrix between the two cameras are obtained, and image correction is performed on this basis; S5: Fabric 3D reconstruction: Convolutional neural network is used to reconstruct fabric 3D, including fusion extraction based on multi-scale features, rough matching based on Euclidean distance, purification to achieve precise matching, and acquisition of fabric 3D point cloud, so as to describe the microscopic morphology and undulation of the fabric surface and realize 3D reconstruction of the fabric; S6: Fabric appearance pleat rating: With the bottom surface of the stage as the reference in space, the point cloud height information is converted to the brightness of the image to form a two-dimensional depth map, in which the higher part is the pleat, and the height of each point is converted to the brightness of the image. The higher the height, the brighter the brightness, and the fabric sample is rated for pleats.

2. The intelligent visual rating method for fabric pleat retention according to claim 1, characterized in that: The step S1, fabric sample processing comprises: (1) Sample cutting: Cut three fabric samples of the same size from the fabric to be tested; (2) Pleat setting: Fold the cut fabric sample in half and use an iron to press a stable pleat in the middle of each fabric sample; (3) Fabric sample processing: Wash, dry and adjust the humidity of the fabric samples, and finally perform shaping treatment on the processed fabric samples.

3. The intelligent visual rating method for fabric pleat retention according to claim 1, characterized in that: In step S1, the image acquisition device comprises a box, the box is used to place the fabric sample, and its bottom plate serves as a stage, and when the box is closed, a space is formed inside the box that is not disturbed by external light sources; A camera and a light source are installed in the box. Two cameras are provided for collecting images of fabric samples. The light source is used for providing illumination. The camera is connected to a computer via a data cable.

4. The intelligent visual rating method for fabric pleat retention according to claim 1, characterized in that: The step S2, fabric image preprocessing comprises: (1) Valid area marking: Mark the valid area of ​​the collected fabric image, retain the content of the marked area, and assign the rest of the area to 0; (2) Grayscale processing; (3) Image denoising; (4) Image enhancement: Use the image histogram to optimize the denoised fabric image, enhance the image contrast, and make the fabric surface features more obvious; (5) Sharpening processing; (6) Light balance: Use white balance technology to adjust the light balance.

5. The intelligent visual rating method for fabric pleat retention according to claim 1, characterized in that: The step S4 comprises: (1) Camera calibration: Using a chessboard as a calibration object, capture template images at different angles and orientations. Then, perform corner detection on the template image to detect feature points in the calibration object. Then, calculate the camera’s intrinsic and extrinsic parameters based on the projection information of these feature points to achieve camera calibration. (2) Image correction: First, a distortion model and a data fitting objective function are established. Then, the corner point detection method is used to obtain the coordinates of the corner points on the template image. Then, the coordinates of the corresponding corner points are calculated using the established distortion model. Finally, these two sets of coordinates are brought into the objective function formula for data fitting and the corresponding distortion parameter values ​​are calculated to eliminate camera distortion and complete image correction. The distortion model is as follows: In the formula, (x i ,y i ) is the actual coordinate of the target corner point, (X i ,Y i ) are the normalized coordinates of the target corner points.

6. The intelligent visual rating method for fabric pleat retention according to claim 1, characterized in that: The step S5 is based on the fusion extraction of multi-scale features: by using multiple convolutional layers to extract features from the left and right fabric images, constructing related feature pyramids, and forming multi-level feature maps; fusing high-level feature maps with low-level feature maps to obtain image features, specifically: adding the feature map of the lowest layer in the feature pyramid network to the adjacent upper layer to form a new feature map, and performing the same operation on the upper layer, so as to fuse all the information in the pyramid network together and realize the fusion of high-level and low-level feature maps.

7. The intelligent visual rating method for fabric pleat retention according to claim 6, characterized in that: The step S5, rough matching based on Euclidean distance: rough matching is performed on the feature map extracted in the above step, specifically: a key point of any fabric image is selected in the left and right fabric images, and the first two key points with the closest Euclidean distance to the corresponding fabric image are found; among these two key points, if the closest distance divided by the second closest distance is less than a certain ratio threshold, then this pair of matching points is accepted; wherein if the Euclidean distance is d, the threshold is set to T=αd, 0<α<1.

8. The intelligent visual rating method for fabric pleat retention according to claim 7, characterized in that: The step S5 is to purify and realize accurate matching: after the rough matching, the obtained matching point set is purified by repeated testing and continuous iteration to realize accurate matching.

9. The intelligent visual rating method for fabric pleat retention according to claim 8, characterized in that: The step S5, obtaining the three-dimensional point cloud of the fabric: after obtaining the matching points corresponding to the left and right fabric images, the parallax of all points in the image is obtained in combination with the camera calibration, the depth is obtained by parallax calculation, and then the spatial point cloud information of the image is obtained.

10. The intelligent visual rating method for fabric pleat retention according to claim 1, characterized in that: The step S6 comprises: (1) Data input and preprocessing: First, adjust the input data format to make it consistent with the training data format, and then input the fabric image spatial point cloud information into the trained classification model; (2) Grade prediction: The classification model calculates the pleat grade of the fabric based on the input spatial point cloud information, and the output result is a rating from 1 to 5; (3) Result comparison and verification: Compare the pleat level predicted by the classification model with the pleat level of the actual sample; use the real pleat data as a reference to test whether the prediction of the classification model is consistent with the actual situation; (4) Accuracy optimization: By comparing different classification models, the nearest neighbor algorithm with the highest recognition rate is used for pleat grade evaluation; when pleat rating, the classification model is verified multiple times; (5) Result output: Through the above steps, the classification model will output the pleat level of each fabric sample and classify it into levels 1 to 5.

Images