A knitted fabric defect detection method and system based on projection method

Abstract

The invention relates to the technical field of image processing, and in particular to a knitted fabric defect detection method and system based on a projection method, comprising: firstly collecting a surface image of the knitted fabric, and preprocessing the surface image into a first image; if the first image is a non-jacquard fabric image, directly using it as a second image, and if it is a jacquard fabric image, using Gabor filtering to extract texture features to separate high-frequency and low-frequency areas, using a PM model to perform diffusion processing on them respectively, and then combining them into a second image; multi-directionally projecting the second image, generating grayscale integral curves at different angles and calculating their first-order derivatives, identifying potential defect areas through mutation points and locally amplifying them into a third image; using a YOLOV7 model to detect defects, calculating surface defect values, collecting the content ratio, physical strength and chemical content of the fabric, respectively calculating scoring values, weighted summing the scoring values ​​and the surface defect values, generating quality assessment results, and improving the detection accuracy of the knitted fabric through multi-dimensional characteristic analysis.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a knitted fabric defect detection method and system based on projection method. Background Art

[0002] Knitted fabrics are widely used in clothing and household goods, and are popular due to their unique elasticity and comfort. However, during the production process, knitted fabrics are prone to different types of defects, such as thread ends, color differences, stains, and uneven fabrics. These defects not only affect the appearance quality of knitted fabrics, but also their performance. Therefore, efficient and accurate detection of knitted fabric defects has become an urgent problem to be solved in the textile industry.

[0003] In recent years, with the vigorous development of artificial intelligence and image recognition technology, intelligent defect detection technology has been widely used in the textile industry, which not only improves the detection speed, but also improves the stability of product quality. However, existing detection schemes are mostly focused on smooth knitted fabrics, and the detection of jacquard fabric defects is relatively weak, which may lead to limited detection accuracy on jacquard fabrics. In addition, most existing schemes use a single-direction projection method, which is usually limited to longitudinal and transverse projections, lacking comprehensive analysis of multi-directional projections, and it is difficult to accurately capture the complex and delicate texture characteristics of knitted fabrics. At the same time, existing schemes often ignore the multi-dimensional characteristics of fabrics, such as physical properties, chemical composition, and fabric composition, which limits the accuracy of detection to a certain extent.

[0004] Therefore, a knitted fabric defect detection method and system based on projection method are proposed. Summary of the invention

[0005] The object of the present invention is to provide a knitted fabric defect detection method and system based on projection method, which collects the surface image of the knitted fabric and pre-processes it into a first image; if the first image is a non-jacquard fabric image, it is directly used as the second image; if it is a jacquard fabric image, Gabor filtering is used to extract texture features, and after separating the high-frequency and low-frequency areas, the PM model is used to diffuse them respectively, and then they are combined into a second image; multi-directional projection is performed on the second image to generate grayscale integral curves at different angles and calculate their first-order derivatives, and potential defect areas are identified through mutation points and partially enlarged into a third image; the YOLOV7 model is used to detect defects, calculate surface defect values, collect the content ratio, physical strength and chemical content of the fabric, calculate score values ​​respectively, and weighted sum the score values ​​and the surface defect values ​​to generate quality assessment results, thereby improving the detection accuracy of knitted fabrics.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] Part 1: A knitted fabric defect detection method and system based on projection method, comprising:

[0008] Collecting a knitted fabric surface image, and preprocessing the knitted fabric surface image to obtain a first image, wherein the first image includes a jacquard fabric image and a non-jacquard fabric image;

[0009] If the first image is the non-jacquard fabric image, the first image is directly recorded as the second image; if the first image is the jacquard fabric image, the texture features of the jacquard fabric image are extracted using Gabor filtering to obtain a high-frequency area and a low-frequency area; the PM model is used to perform diffusion processing on the high-frequency area and the low-frequency area respectively, and the processed high-frequency area and the low-frequency area are combined to obtain the second image; wherein the PM model is used to smooth the jacquard texture in the high-frequency area and eliminate the noise in the low-frequency area;

[0010] Performing grayscale integral projection on the second image using a multi-directional projection method to obtain integral curves at different projection angles; calculating the first-order derivative of the integral curve to obtain derivative values ​​at different positions, and obtaining mutation points according to the derivative values; performing intersection processing on all the mutation points to obtain a potential defect area;

[0011] The potential defect area is intercepted and partially enlarged to obtain a third image; the third image is detected using the trained YOLOV7 model to obtain actual defects; and the surface defect value is obtained according to the number and size of the actual defects;

[0012] The content ratio, physical strength and chemical content of the knitted fabric are obtained, the scoring values ​​of the content ratio, the physical strength and the chemical content are calculated respectively, and the scoring values ​​and the surface defect values ​​are weightedly summed to obtain a quality assessment result.

[0013] Furthermore, the process of acquiring the first image includes:

[0014] Graying the surface image of the knitted fabric to obtain a grayscale image;

[0015] Performing denoising on the grayscale image using Gaussian filtering to obtain a denoised image;

[0016] The denoised image is classified using a trained CNN model to obtain the jacquard fabric image and the non-jacquard fabric image.

[0017] Furthermore, the processing of the high frequency region includes:

[0018] assigning a first diffusion coefficient to the high frequency region;

[0019] Calculating the gradient of each pixel in the high-frequency area to obtain an edge intensity value;

[0020] The diffusion intensity is calculated by using a selective diffusion function according to the first diffusion coefficient and the edge intensity value;

[0021] The high-frequency region is subjected to diffusion processing using a diffusion equation of the PM model according to the diffusion intensity to obtain the processed high-frequency region.

[0022] Furthermore, the processing of the low-frequency region includes:

[0023] Setting a second diffusion coefficient for the low-frequency region, wherein the second diffusion coefficient is greater than the first diffusion coefficient;

[0024] The low-frequency region is subjected to diffusion processing using the diffusion equation of the PM model according to the second diffusion coefficient to obtain the processed low-frequency region.

[0025] Furthermore, the integral curve acquisition process includes:

[0026] Select different projection angles, and calculate the projection value according to the different projection angles and the gray value of the second image, which is expressed as:

[0027] ;

[0028] in, for Angle at distance The projection value at ; For the second image at point The gray value of and are respectively the horizontal and vertical coordinates of the second image; It is along Angle at distance The set of pixels at ;

[0029] Will Set as the horizontal coordinate of the curve, The vertical coordinate of the curve is set, and the integral curve is obtained by plotting the horizontal coordinate of the curve and the vertical coordinate of the curve.

[0030] Furthermore, the mutation point acquisition process includes:

[0031] Obtaining the integral curves at different projection angles, calculating the first-order derivative for each integral curve, and obtaining the derivative values ​​at different positions;

[0032] If the derivative value exceeds the first threshold, the coordinate position corresponding to the derivative value is marked according to the integral curve, and the coordinate position is recorded as a mutation point.

[0033] Furthermore, the potential defect area marking process includes:

[0034] Collecting all the mutation points and combining them into a mutation point set; performing the intersection processing on the mutation point set to obtain overlapping mutation points and non-overlapping mutation points;

[0035] Taking the overlapping mutation point and the non-overlapping mutation point as centers, respectively draw a first rectangle and a second rectangle, wherein the side length of the first rectangle is greater than the side length of the second rectangle;

[0036] If there is an overlapping part between the first rectangles, the overlapping part between the second rectangles, or the overlapping part between the first rectangle and the second rectangle, the boundary of the minimum circumscribed rectangle is calculated based on the overlapping part, and the potential defect area is obtained based on the boundary of the minimum circumscribed rectangle.

[0037] Furthermore, the training steps of the YOLOV7 model include:

[0038] Step S1: collecting historical knitted fabric images containing defects, and marking the defect areas and defect types in the historical knitted fabric images to obtain a first data set;

[0039] Step S2: inputting the first data set into a basic model, and the basic model outputting a predicted value for the first data set;

[0040] Step S3: using a loss function to calculate a loss value according to the predicted value, the defect area and the defect type;

[0041] Step S4: using the back propagation method to calculate the gradient information of the loss value; using the Adam optimizer to update the network parameters of the basic model according to the gradient information;

[0042] Step S5: If the number of training rounds of the basic model is equal to the preset number of rounds, the training of the basic model is terminated, otherwise, steps S1 to S5 are repeated.

[0043] Furthermore, the calculation process of the quality assessment result includes:

[0044] Obtaining the scoring values ​​of the content ratio, the physical strength, and the chemical content, and normalizing the scoring values ​​and the surface defect values ​​to obtain a first value, a second value, a third value, and a fourth value;

[0045] collecting user satisfaction with the knitted fabric at fixed time intervals, and calculating, according to the satisfaction, the first value, the second value, the third value, and the fourth value, the correlation coefficient between the satisfaction and the first value, the second value, the third value, and the fourth value using the Pearson correlation coefficient method;

[0046] The correlation coefficient is recorded as a weight coefficient, and the score value and the surface defect value are weighted and summed according to the weight coefficient to obtain the quality assessment result.

[0047] The second part is a knitted fabric defect detection system based on projection method, including:

[0048] An image preprocessing module, used for collecting a knitted fabric surface image, and obtaining a first image after preprocessing the knitted fabric surface image;

[0049] An image reprocessing module, if the first image is a non-jacquard fabric image, it is directly recorded as the second image; if the first image is a jacquard fabric image, Gabor filtering is used to extract the texture features of the jacquard fabric image to obtain a high-frequency area and a low-frequency area; the PM model is used to diffuse the high-frequency area and the low-frequency area respectively, and the processed high-frequency area and the low-frequency area are combined to obtain the second image; wherein the PM model is used to smooth the jacquard texture in the high-frequency area and eliminate the noise in the low-frequency area;

[0050] A surface defect calculation module is used to perform grayscale integral projection on the second image using a multi-directional projection method to obtain integral curves at different projection angles; calculate the first-order derivative of the integral curve to obtain a mutation point; perform intersection processing on all the mutation points to obtain a potential defect area; intercept the potential defect area to obtain a third image; use the trained YOLOV7 model to detect the third image to obtain actual defects; and obtain a surface defect value according to the number and size of the actual defects;

[0051] The comprehensive defect calculation module is used to obtain the content ratio, physical strength and chemical content of the knitted fabric, calculate the score values ​​of the content ratio, the physical strength and the chemical content respectively, and perform weighted summation of the score value and the surface defect value to obtain a quality assessment result.

[0052] Compared with the prior art, the present invention has the following beneficial effects:

[0053] 1. The images are classified by convolutional neural network, and the fabrics are divided into jacquard fabrics and non-jacquard fabrics. For jacquard fabrics, Gabor filtering is further used to extract its texture features, distinguishing the fine texture and large uniform areas of knitted fabrics. Then, these areas are diffused separately. The improved PM model can flexibly adjust the diffusion coefficient according to the gradient of the image, which can smooth out the noise in the complex texture while retaining the defect details, so that the detection process focuses on the real defects and improves the accuracy of knitted fabric detection.

[0054] 2. The multi-directional projection method is used for grayscale integral projection. By generating integral curves at multiple projection angles, the adaptability of the projection is improved, and subtle defects of complex texture fabrics can be captured. Then, the first-order derivatives of all integral curves are calculated to identify texture mutation points and locate potential defect areas. In addition, the potential defect area is locally enlarged, and the YOLOv7 model is used for defect detection to identify the specific location, number and size of surface defects, reducing the possibility of misjudgment, thereby improving the accuracy and comprehensiveness of fabric defect detection.

[0055] 3. Collect various factors such as the physical strength, chemical content and content ratio of the fabric and calculate the corresponding score values, so as to comprehensively evaluate the quality of the fabric. At the same time, collect user satisfaction scores for knitted fabrics at fixed time intervals, use the Pearson correlation coefficient method to dynamically calculate the weight coefficient of each factor, and calculate the quality assessment results based on these weight coefficients. This calculation method can flexibly adapt to changes in actual user feedback, making quality assessment more in line with market demand, thereby improving the accuracy of knitted fabric testing. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 A schematic flow chart of a knitted fabric defect detection method based on projection method provided by an embodiment of the present invention;

[0057] Figure 2 A schematic diagram of a process for reprocessing a jacquard fabric provided by an embodiment of the present invention;

[0058] Figure 3 A schematic diagram of a knitted fabric surface detection process provided by an embodiment of the present invention;

[0059] Figure 4 A schematic structural diagram of a knitted fabric defect detection system based on projection method provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0060] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0061] See also Figures 1 to 4 The present invention provides a knitted fabric defect detection method and system based on projection method, and the technical scheme is as follows:

[0062] Embodiment 1

[0063] At present, the quality of knitted fabrics directly affects the market competitiveness and user satisfaction of products. The knitted fabric defect detection method adopted by Company A in actual production mainly relies on simple image processing methods. However, when dealing with complex textures and small-scale defects, there is a problem of inaccurate detection accuracy. Especially on high-frequency production lines, for knitted fabrics with complex patterns and texture structures, the problem of inaccurate detection leads to increased rework rate and production costs, which makes it difficult to meet the actual needs of knitted fabric defect detection. Therefore, there is an urgent need for a new solution that can accurately identify knitted fabric defects.

[0064] Figure 1 A schematic flow chart of a knitted fabric defect detection method based on projection method provided in this embodiment.

[0065] like Figure 1 As shown, a knitted fabric defect detection method based on projection method comprises:

[0066] The surface image of the knitted fabric is collected by a high-resolution camera to ensure image clarity and detail integrity; the surface image of the knitted fabric is preprocessed to obtain a first image, which includes a jacquard fabric image and a non-jacquard fabric image.

[0067] Furthermore, the process of acquiring the first image includes:

[0068] Graying the surface image of the knitted fabric to obtain a grayscale image, thereby eliminating the interference of color on defect detection and focusing on texture extraction during later detection;

[0069] Using Gaussian filtering to perform denoising on the grayscale image to obtain a denoised image, which can effectively remove high-frequency noise in the grayscale image;

[0070] Among them, Gaussian filtering uses a Gaussian filter with a kernel size of 5×5 and a standard deviation of 1.5.

[0071] The denoised image is classified using a trained CNN model to obtain the jacquard fabric image and the non-jacquard fabric image.

[0072] Among them, in this embodiment, the input layer of the CNN model receives a 224×224 denoised image; three convolutional layers are used, which contain 64, 128 and 256 filters respectively; a maximum pooling layer is added after each convolutional layer, and the pooling window size is 2×2; the fully connected layer contains 128 neurons, and then the jacquard fabric image and the non-jacquard fabric image are classified by connecting the Softmax activation function.

[0073] Specifically, 1000 images of jacquard fabrics and non-jacquard fabrics were used as training sets, 800 images were used as test sets, the learning rate was set to 0.001, the batch size was 32, and the accuracy was 94% after 200 rounds of training. As shown in Table 1, by comparing the test results with other classification models, it can be seen that the CNN model has the highest accuracy in classifying knitted fabrics and can better identify jacquard features.

[0074] Table 1 Comparison of test results of classification models

[0075]

[0076] By preprocessing the surface images of knitted fabrics and inputting the preprocessed images into the trained CNN model, jacquard and non-jacquard fabrics can be accurately classified, effectively avoiding the interference of texture type on subsequent detection, thereby improving the accuracy and pertinence of knitted fabric detection.

[0077] If the first image is the non-jacquard fabric image, it is directly recorded as the second image;

[0078] Figure 2 This is a schematic diagram of the process of reprocessing the jacquard fabric provided in this embodiment.

[0079] like Figure 2 As shown, if the first image is the jacquard fabric image, Gabor filtering is used to extract the texture features of the jacquard fabric image to obtain a high-frequency area and a low-frequency area; the PM model is used to perform diffusion processing on the high-frequency area and the low-frequency area respectively, and the processed high-frequency area and the low-frequency area are combined to obtain the second image; wherein the PM model is used to smooth the jacquard texture in the high-frequency area and eliminate the noise in the low-frequency area;

[0080] Among them, Gabor filtering is a classic method for extracting texture features in image processing, which can effectively capture the texture information of knitted fabrics. Set a lower frequency, such as 0.3, and apply a low-frequency Gabor filter to convolve the first image to extract the low-frequency area, that is, the large-scale texture and background structure in the knitted fabric, including the basic fabric structure and color changes; then design a higher frequency, such as 0.8, and apply a high-frequency Gabor filter to convolve the first image to extract the high-frequency area, that is, the details in the knitted fabric, including jacquard and small defects. Save the low-frequency area and the high-frequency area as two independent images for subsequent diffusion processing.

[0081] Furthermore, the processing of the high frequency region includes:

[0082] Specify a first diffusion coefficient for the high frequency region , used to control the diffusion rate;

[0083] Among them, in this embodiment, Set to 0.2;

[0084] Calculate the gradient of each pixel in the high-frequency area to obtain the edge strength value ;

[0085] Among them, the edge strength value Indicates the intensity of pixel value change, When it is larger, it means that the pixel changes more dramatically, corresponding to image defects and edge areas. When it is smaller, it corresponds to the area where the image is smoother and more regular.

[0086] The diffusion intensity is calculated using the selective diffusion function according to the first diffusion coefficient and the edge strength value, and is expressed as:

[0087] ;

[0088] in, In order to increase the diffusion intensity, the diffusion will be reduced for high gradient areas (such as defective parts) to ensure that defects and edge details are retained. For areas with continuous gradients (such as jacquard texture parts), the jacquard pattern will be gradually smoothed to weaken fine textures.

[0089] The high frequency region is diffused using the diffusion equation of the PM model according to the diffusion intensity to obtain the processed high frequency region, which is expressed as:

[0090] ;

[0091] in, It is The first image after iterations, is the number of iterations; is the step size parameter, which controls the change amplitude of each iteration and is set to 0.2; div is the divergence operation.

[0092] The number of iterations is preset to 10 times. After the iteration is completed, the image is observed. If the details of the edge and defect are clear and the noise is less, the iteration is not continued. If the image still shows a lot of noise or the defect is difficult to distinguish, the number of iterations is appropriately increased until the effect is achieved. In this embodiment, the number of iterations is 20 times.

[0093] By using the selective diffusion function, the high-frequency area of ​​the knitted fabric is controlled by the edge strength during the diffusion process, retaining the defect details so that the defect features will not be lost due to over-smoothing. For the jacquard part of the high-frequency area, the diffusion intensity will be slightly increased, so that after each iteration, the details of the jacquard will be gradually smoothed, thereby improving the accuracy of fabric defect detection.

[0094] Furthermore, the processing of the low-frequency region includes:

[0095] Specify a second diffusion coefficient for the low frequency region , the second diffusion coefficient Greater than the first diffusion coefficient ;

[0096] Among them, in this embodiment, Set to 1;

[0097] The low-frequency region is diffused using the diffusion equation of the PM model according to the second diffusion coefficient to obtain the processed low-frequency region, which is expressed as:

[0098] ;

[0099] in, It is The first image after iterations, is the number of iterations; is the step size parameter, set to 0.2; div is the divergence operation.

[0100] The number of iterations in the low-frequency region is generally the same as that in the high-frequency region.

[0101] Specifically, the diffusion of low-frequency areas is to make the background more uniform, so there is no need to use selective diffusion for low-frequency areas. Through a higher diffusion coefficient, the subtle noise in the low-frequency area is effectively smoothed, and the texture details in the large-scale area are reduced, so that in the subsequent steps, more attention is paid to defects rather than texture noise, thereby improving the reliability and accuracy of knitted fabric detection.

[0102] Furthermore, after the high-frequency area and the low-frequency area are diffused separately, the two parts are combined into a new image using a weighted fusion method. In this embodiment, the high-frequency area and the low-frequency area are respectively set to 0.5, so that different frequency bands can transition naturally and retain the defect information of the high-frequency area and the uniform background of the low-frequency area.

[0103] Figure 3 This is a schematic diagram of the knitted fabric surface detection process provided in this embodiment.

[0104] like Figure 3 As shown, the second image is projected with grayscale integral using a multi-directional projection method to obtain integral curves at different projection angles; the first-order derivative of the integral curve is calculated to obtain derivative values ​​at different positions, and mutation points are obtained based on the derivative values; all the mutation points are subjected to intersection processing to obtain a potential defect area; the potential defect area is intercepted and locally enlarged to obtain a third image; the third image is detected using a trained YOLOV7 model to obtain actual defects; and the surface defect value is obtained based on the number and size of the actual defects.

[0105] Furthermore, the integral curve acquisition process includes:

[0106] Select different projection angles, and calculate the projection value according to the different projection angles and the gray value of the second image, which is expressed as:

[0107] ;

[0108] in, for Angle at distance The projection value at ; For the second image at point The gray value of and are respectively the horizontal and vertical coordinates of the second image; It is along Angle at distance The set of pixels at ;

[0109] Will Set as the horizontal coordinate of the curve, The vertical coordinate of the curve is set, and the integral curve is obtained by plotting the horizontal coordinate of the curve and the vertical coordinate of the curve.

[0110] In this embodiment, the projection angles are selected as 0°, 45°, 90° and 135°, and pixel projection is performed for each angle; in addition, the distance It does not refer to the actual distance between pixels, but is expressed as a serial number.

[0111] The distance As the horizontal axis, the projection value The vertical coordinate is set as the vertical coordinate, and the integral curve is obtained by plotting the horizontal coordinate and the vertical coordinate.

[0112] Specifically, taking a 5×5 grayscale matrix as an example, it is expressed as:

[0113] ;

[0114] in, is the grayscale matrix, i.e. the second image;

[0115] If the projection angle is 0°, the grayscale values ​​are summed along each column of the grayscale matrix, and the projection values ​​are distance , , , , ; Here, the column number is set to the distance ; According to the projection value, the integral curve can be drawn, which is expressed as:

[0116] ;

[0117] Specifically, the multi-directional projection method was compared with the two-directional projection detection method in the warp and weft directions, and the same 10 pieces of defective knitted fabrics with different pattern directions were used for verification. The proportion of successfully identifying defective areas in the former was generally increased by 20% to 50% compared with the latter. Multi-directional projection helps to detect details and defects in different directions and extract the characteristic distribution of images at different angles. For example, the integral curve with a projection angle of 45° can capture the inclined texture and oblique defects in knitted fabrics, and achieve more comprehensive image analysis, thereby improving the reliability and accuracy of knitted fabric detection.

[0118] Furthermore, the mutation point acquisition process includes:

[0119] Obtaining the integral curves at different projection angles, calculating the first-order derivative for each integral curve, and obtaining the derivative values ​​at different positions;

[0120] If the derivative value exceeds the first threshold, the coordinate position corresponding to the derivative value is marked according to the integral curve, and the coordinate position is recorded as a mutation point.

[0121] Specifically, taking a 5×5 grayscale matrix as an example, assuming that the integral curve with a projection angle of 45° is expressed as:

[0122] ;

[0123] The derivative value of each point is calculated using the difference method, expressed as:

[0124] ;

[0125] in, For point The derivative value of is the projection distance, is the projection value;

[0126] After calculating the derivative, the result is:

[0127] ;

[0128] Assuming that the first threshold is set to an absolute value of 300, there is an abnormality in the projection value at a distance of 8. If the lower left corner of the second image is taken as the origin, the coordinate position of the mutation point is located at (5, 0).

[0129] Under projection at multiple angles, the derivative of the integral curve is calculated and all possible mutation points are marked, which can effectively detect key feature changes in the surface image of knitted fabrics, such as defective areas of edges and textures, thereby improving the accuracy of knitted fabric anomaly detection.

[0130] Furthermore, the potential defect area marking process includes:

[0131] Collecting all the mutation points and combining them into a mutation point set; performing the intersection processing on the mutation point set to obtain overlapping mutation points and non-overlapping mutation points;

[0132] Taking the overlapping mutation point and the non-overlapping mutation point as centers, respectively draw a first rectangle and a second rectangle, wherein the side length of the first rectangle is greater than the side length of the second rectangle;

[0133] If there is an overlapping part between the first rectangles, the overlapping part between the second rectangles, or the overlapping part between the first rectangle and the second rectangle, the boundary of the minimum circumscribed rectangle is calculated based on the overlapping part, and the potential defect area is obtained based on the boundary of the minimum circumscribed rectangle.

[0134] Specifically, taking the projection angles of 0° and 90° as examples, assuming that the mutation point for 90° is {(1, 1), (2, 1), (2, 3)}, and the mutation point for 0° is {(1, 1), (1, 2), (1, 3)}, then the overlapping mutation point is {(1, 1)}, and the non-overlapping mutation point is {(1, 2), (1, 3), (2, 1), (2, 3)}. For the overlapping mutation point, a rectangle with a side length of 10 pixels is drawn with (1, 1) as the center, and for the non-overlapping mutation point, a rectangle with a side length of 5 pixels is drawn with {(1, 2), (1, 3), (2, 1), (2, 3)} as the center. Check whether there is any overlap between the rectangles, and merge them if there is overlap between the rectangles. For example, the center point of rectangle 1 is (1, 1), and the range is from (-4, -4) to (6, 6), and the center point of rectangle 2 is (1, 2), and the range is from (-1.5, -0.5) to (3.5, 3.5). Calculate their minimum circumscribed rectangle boundaries, the minimum x value of the horizontal range is -4, the maximum x value is 6, the minimum y value of the vertical range is -4, and the maximum y value is 6. Then the merged rectangle range is (-4, -4) to (6, 6), and this rectangle is recorded as a potential defect area. By processing the intersection of mutation points, abnormal changes in the surface image of knitted fabrics can be found in multiple directions, and multiple rectangular areas can be merged by calculating the minimum circumscribed rectangle of overlapping rectangles, making the detection area more accurate and concentrated, thereby improving the accuracy of image defect detection and reducing the complexity of later re-inspection.

[0135] Furthermore, the training steps of the YOLOV7 model include:

[0136] Step S1: collecting historical knitted fabric images containing defects, and marking the defect areas and defect types in the historical knitted fabric images to obtain a first data set;

[0137] The defect area is the bounding box coordinates of the defect, and the defect types include cracks and stains.

[0138] Step S2: input the first data set into the basic model for forward propagation, and the basic model outputs the predicted value of the first data set; wherein the basic model is the state of initialization of the YOLOV7 model;

[0139] The basic model extracts the features of the image through its convolutional network, and then outputs the location, confidence, and defect type of each defect area;

[0140] Step S3: using a loss function to calculate a loss value according to the predicted value, the defect area and the defect type;

[0141] Among them, the loss function includes:

[0142] Position loss, which is the difference between the predicted bounding box and the annotated bounding box;

[0143] Confidence loss, that is, the confidence error of the target;

[0144] Class loss, which is the difference between the predicted class and the actual defect class;

[0145] Step S4: using the back propagation method to calculate the gradient information of the loss value; using the Adam optimizer to update the network parameters of the basic model according to the gradient information, so that the model can gradually predict the defect location and category more accurately;

[0146] Step S5: If the number of training rounds of the basic model is equal to the preset number of rounds, the training of the basic model is terminated, otherwise, steps S1 to S5 are repeated.

[0147] Specifically, in this embodiment, 1,000 knitted fabric images containing different types of defects, including stains and cracks, were collected. The defective areas in the images were labeled using the LabelImg annotation tool. The number of training rounds was set to 100. The mAP (Mean Average Precision) of the YOLOV7 model reached 0.93, indicating that the model has a high accuracy in detecting and classifying defects. After the model training was completed, 200 knitted fabric images containing different types of defects were used for testing. In order to prove the superiority of the YOLOV7 model, two comparison models were selected. As shown in Table 2, the YOLOv7 model has the highest mAP, which shows that the YOLOv7 model can detect defects more accurately and the YOLOV7 model can be deployed. Re-detection by the YOLOv7 model achieves accurate knitted fabric defect detection and improves detection efficiency and accuracy.

[0148] Table 2 Comparison of test results of target detection models

[0149]

[0150] Furthermore, in this embodiment, a custom four-point detection method is used to score each square meter of fabric, including:

[0151] If the fabric has no flaws and no surface defects, it will be scored as 4 points, which means it is a high-quality product;

[0152] If the fabric defect is less than 1 cm and the number of defects is less than 3, it will be scored as 3 points, which has little impact on the fabric quality;

[0153] If the fabric defect is less than 3 cm and greater than or equal to 1 cm, and the number of defects is less than 5 and greater than or equal to 3, it will be scored as 2 points, which has a certain impact on the fabric;

[0154] If the fabric defect is less than 5 cm and greater than or equal to 3 cm, and the number of defects is less than 7 and greater than or equal to 5, it will be scored as 1 point, which has a significant impact on the appearance and use of the fabric;

[0155] If the fabric defect is greater than or equal to 5 cm and the number of defects is greater than or equal to 7, it will be scored as 0 points and the fabric may not be able to be used normally.

[0156] Furthermore, the method further comprises:

[0157] The content ratio, physical strength and chemical content of the knitted fabric are obtained, the score values ​​of the content ratio, the physical strength and the chemical content are calculated respectively, and the score values ​​and the surface defect values ​​are normalized to obtain a first value, a second value, a third value and a fourth value.

[0158] Among them, the content ratio of knitted fabrics can be obtained by using infrared spectroscopy to analyze the spectral characteristics of different components to obtain the proportion of each component; the physical strength of knitted fabrics can be obtained by using a tensile strength testing machine to obtain mechanical performance indicators such as tensile strength and tear strength; the chemical content of knitted fabrics can be obtained by using dissolution analysis and other methods to obtain the formaldehyde content, pH value and metal composition of knitted fabrics.

[0159] Specifically, assuming that the ideal composition ratio of knitted fabric is , the actual content ratio was detected to be , calculate the difference between the actual content ratio and the standard content ratio of each component, and set weights for different components , can be set to 0.3 respectively, expressed as:

[0160] ;

[0161] in, Score the content ratio;

[0162] Assume that the standard tensile strength and tear strength are The actual detection value is , the physical strength score is:

[0163] ;

[0164] in, is the physical strength score, and The weight factors can be set to 0.7 and 0.3 respectively;

[0165] Assuming the upper limit of standard formaldehyde content is The actual detection value is , the formaldehyde score can be expressed as:

[0166] ;

[0167] in, The formaldehyde score and other chemical contents can also be calculated in this way. The chemical score value is obtained by weighted summation, where the weight factors of formaldehyde content, pH value and metal component can be set to 0.4, 0.3 and 0.3 respectively.

[0168] The user's satisfaction with the knitted fabric is collected at fixed time intervals, and the correlation coefficient between the satisfaction and the first value, the second value, the third value and the fourth value is calculated using the Pearson correlation coefficient method according to the satisfaction, the first value, the second value, the third value and the fourth value, which is expressed as:

[0169] ;

[0170] in is the weight coefficient of the ith value, For user satisfaction, is the jth user’s satisfaction, is the mean satisfaction value, represents the i-th numerical rating value of the j-th user, is the mean score of the ith value.

[0171] The correlation coefficient is recorded as the weight coefficient, and the score value and the surface defect value are weighted and summed according to the weight coefficient to obtain the quality assessment result , expressed as:

[0172] ;

[0173] in, , , and are weight coefficients of the first value, the second value, the third value and the fourth value respectively; , , and They are the first value, the second value, the third value and the fourth value respectively.

[0174] Under this quality assessment result, if the set qualification standard is a total score greater than 0.7 points, the fabric meets the standard and is assessed as a qualified product.

[0175] The correlation between user satisfaction and various indicators is calculated through the Pearson correlation coefficient, so that the weight of each indicator corresponds to the actual feedback of the user, ensuring that the quality assessment is more in line with user needs. At the same time, it also makes the assessment results more flexible to adapt to quality changes, thereby improving the accuracy of the test results.

[0176] In order to prove the effect of the method of the present invention, the method of the present invention is compared with the knitted fabric defect detection method used in the actual production of a certain company A. The detection rate of the surface defects of the fabric by the method of the present invention reaches 96.5%, and the detection rate of the surface defects of the fabric by the comparison method is 90.5%, which shows that the method has obvious advantages in defect detection. In addition, the method of the present invention also calculates the overall performance of the fabric, which can meet the needs of users and more directly judge whether the quality of the fabric meets the standards. In short, by using Gabor filtering and PM diffusion model, it has higher adaptability in the processing of jacquard fabrics, separates high and low frequency features, and makes the defects more obvious when smoothing the pattern; then, through multi-angle grayscale integral curve analysis and mutation point recognition, the potential defect area can be accurately located, and then the defects are accurately identified by introducing the YOLOV7 deep learning model, which effectively improves the accuracy of defect detection; finally, the content ratio, physical strength, chemical content and surface defects are weighted and summed to generate a more comprehensive quality assessment result, thereby improving the overall detection accuracy of the fabric and providing a clear basis for subsequent product improvements.

[0177] Embodiment 2

[0178] Factory B mainly produces jacquard fabrics. With the diversification of jacquard fabric patterns and weaves, the existing detection methods face the challenge of insufficient detection efficiency and accuracy. In order to solve this problem, it is urgent to introduce a new knitted fabric detection system to improve the accuracy of detection.

[0179] Figure 4 A schematic diagram of a knitted fabric defect detection system based on projection method provided in this embodiment.

[0180] like Figure 4 As shown, a knitted fabric defect detection system based on projection method comprises:

[0181] An image preprocessing module, used for collecting a knitted fabric surface image, and obtaining a first image after preprocessing the knitted fabric surface image;

[0182] An image reprocessing module, if the first image is a non-jacquard fabric image, it is directly recorded as the second image; if the first image is a jacquard fabric image, Gabor filtering is used to extract the texture features of the jacquard fabric image to obtain a high-frequency area and a low-frequency area; the PM model is used to diffuse the high-frequency area and the low-frequency area respectively, and the processed high-frequency area and the low-frequency area are combined to obtain the second image; wherein the PM model is used to smooth the jacquard texture in the high-frequency area and eliminate the noise in the low-frequency area;

[0183] A surface defect calculation module is used to perform grayscale integral projection on the second image using a multi-directional projection method to obtain integral curves at different projection angles; calculate the first-order derivative of the integral curve to obtain a mutation point; perform intersection processing on all the mutation points to obtain a potential defect area; intercept the potential defect area to obtain a third image; use the trained YOLOV7 model to detect the third image to obtain actual defects; and obtain a surface defect value according to the number and size of the actual defects;

[0184] The comprehensive defect calculation module is used to obtain the content ratio, physical strength and chemical content of the knitted fabric, calculate the score values ​​of the content ratio, the physical strength and the chemical content respectively, and perform weighted summation of the score value and the surface defect value to obtain a quality assessment result.

[0185] Furthermore, the process of acquiring the first image includes:

[0186] Graying the surface image of the knitted fabric to obtain a grayscale image;

[0187] Performing denoising on the grayscale image using Gaussian filtering to obtain a denoised image;

[0188] The denoised image is classified using a trained CNN model to obtain the jacquard fabric image and the non-jacquard fabric image.

[0189] Furthermore, the processing of the high frequency region includes:

[0190] assigning a first diffusion coefficient to the high frequency region;

[0191] Calculating the gradient of each pixel in the high-frequency area to obtain an edge intensity value;

[0192] The diffusion intensity is calculated by using a selective diffusion function according to the first diffusion coefficient and the edge intensity value;

[0193] The high-frequency region is subjected to diffusion processing using a diffusion equation of the PM model according to the diffusion intensity to obtain the processed high-frequency region.

[0194] Furthermore, the processing of the low-frequency region includes:

[0195] Setting a second diffusion coefficient for the low-frequency region, wherein the second diffusion coefficient is greater than the first diffusion coefficient;

[0196] The low-frequency region is subjected to diffusion processing using the diffusion equation of the PM model according to the second diffusion coefficient to obtain the processed low-frequency region.

[0197] Furthermore, the integral curve acquisition process includes:

[0198] Select different projection angles, and calculate the projection value according to the different projection angles and the gray value of the second image, which is expressed as:

[0199] ;

[0200] in, for Angle at distance The projection value at ; For the second image at point The gray value of and are respectively the horizontal and vertical coordinates of the second image; It is along Angle at distance The set of pixels at ;

[0201] Among them, in this embodiment, the projection angles are selected as 0°, 45°, 90° and 135°;

[0202] Will Set as the horizontal coordinate of the curve, The vertical coordinate of the curve is set, and the integral curve is obtained by plotting the horizontal coordinate of the curve and the vertical coordinate of the curve.

[0203] Furthermore, the mutation point acquisition process includes:

[0204] Obtaining the integral curves at different projection angles, calculating the first-order derivative for each integral curve, and obtaining the derivative values ​​at different positions;

[0205] If the derivative value exceeds the first threshold, the coordinate position corresponding to the derivative value is marked according to the integral curve, and the coordinate position is recorded as a mutation point.

[0206] Furthermore, the potential defect area marking process includes:

[0207] Collecting all the mutation points and combining them into a mutation point set; performing the intersection processing on the mutation point set to obtain overlapping mutation points and non-overlapping mutation points;

[0208] Taking the overlapping mutation point and the non-overlapping mutation point as centers, respectively draw a first rectangle and a second rectangle, wherein the side length of the first rectangle is greater than the side length of the second rectangle;

[0209] If there is an overlapping part between the first rectangles, the overlapping part between the second rectangles, or the overlapping part between the first rectangle and the second rectangle, the boundary of the minimum circumscribed rectangle is calculated based on the overlapping part, and the potential defect area is obtained based on the boundary of the minimum circumscribed rectangle.

[0210] Furthermore, the training steps of the YOLOV7 model include:

[0211] Step S1: collecting historical knitted fabric images containing defects, and marking the defect areas and defect types in the historical knitted fabric images to obtain a first data set;

[0212] Step S2: inputting the first data set into a basic model, and the basic model outputting a predicted value for the first data set;

[0213] Step S3: using a loss function to calculate a loss value according to the predicted value, the defect area and the defect type;

[0214] Step S4: using the back propagation method to calculate the gradient information of the loss value; using the Adam optimizer to update the network parameters of the basic model according to the gradient information;

[0215] Step S5: If the number of training rounds of the basic model is equal to the preset number of rounds, the training of the basic model is terminated, otherwise, steps S1 to S5 are repeated.

[0216] Furthermore, in this embodiment, the standard four-point scoring method is used to score the surface defects of knitted fabrics. The scoring criteria for defects in warp, weft and other directions are:

[0217] 4 points, the defect length is 3 cm or less;

[0218] 3 points: the defect length exceeds 3 cm but does not exceed 6 cm;

[0219] 2 points: the defect length exceeds 6 cm but does not exceed 9 cm;

[0220] 1 point: defect length is greater than 9 cm.

[0221] In addition, 1 point will be given for defects that appear continuously and those that seriously affect the appearance.

[0222] Furthermore, the calculation process of the quality assessment result includes:

[0223] Calculating the score values ​​of the content ratio, the physical strength, and the chemical content, and normalizing the score values ​​and the surface defect values ​​to obtain a first value, a second value, a third value, and a fourth value;

[0224] collecting user satisfaction with the knitted fabric at fixed time intervals, calculating correlation coefficients between the satisfaction and the first value, the second value, the third value, and the fourth value using a Pearson correlation coefficient method, and recording the correlation coefficients as weight coefficients;

[0225] The score value and the surface defect value are weighted and summed according to the weight coefficient to obtain the quality assessment result.

[0226] Specifically, if the quality assessment result of a piece of cloth is less than 0.6 points, it is unqualified.

[0227] Specifically, in order to prove the effect of the system of the present invention, it is compared with the existing detection system of a factory B. The detection and sampling standards formulated by AATCC (American Association of Textile Chemists and Colorists) are used for quality control, including: the number of samples is the square root of the total number of yards multiplied by eight, and the number of boxes is the square root of the total number of boxes. For ten boxes of knitted fabrics with a total fabric length of 5,000 yards, the number of sampling boxes is 3, the number of samples is 71, and the sampling test results are: the overall defect detection rate of the system of the present invention is 93.2%, while the overall defect detection rate of the original system of factory B is 89.7%. The system of the present invention has a higher defect detection rate, which is increased by 3.5 percentage points, and can more keenly capture small and overall complex defects in the fabric, improving the accuracy of the system in detecting knitted fabrics.

[0228] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A knitted fabric defect detection method based on projection method, characterized in that: include: Collecting a knitted fabric surface image, and preprocessing the knitted fabric surface image to obtain a first image, wherein the first image includes a jacquard fabric image and a non-jacquard fabric image; If the first image is the non-jacquard fabric image, the first image is directly recorded as the second image; if the first image is the jacquard fabric image, a Gabor filter is used to extract the texture features of the jacquard fabric image to obtain a high-frequency area and a low-frequency area, the Gabor filter includes a low-frequency Gabor filter and a high-frequency Gabor filter, and a first diffusion coefficient of the high-frequency Gabor filter is higher than a second diffusion coefficient of the low-frequency Gabor filter; Using the PM model to perform diffusion processing on the high-frequency region and the low-frequency region respectively, and combining the processed high-frequency region and the low-frequency region using a weighted fusion method to obtain the second image; wherein the PM model is used to smooth the jacquard texture of the high-frequency region using a selective diffusion function and a diffusion equation, and at the same time use the diffusion equation to eliminate noise in the low-frequency region; Use a multi-directional projection method to perform grayscale integral projection on the second image, and accumulate grayscale values ​​along the direction of the projection angle to obtain a projection value; set the projection distance as the abscissa of the curve, set the projection value as the ordinate of the curve, and obtain integral curves at different projection angles; use a difference method to calculate the first-order derivative of the integral curve to obtain derivative values ​​at different positions; if the derivative value exceeds a first threshold, the coordinate position corresponding to the derivative value is recorded as a mutation point; otherwise, the derivative value is directly saved; perform intersection processing on all the mutation points to obtain overlapping mutation points and non-overlapping mutation points, and draw a minimum circumscribed rectangle with the overlapping mutation points and the non-overlapping mutation points as the center to obtain a potential defect area; wherein the side length of the minimum circumscribed rectangle of the overlapping mutation point is greater than the side length of the minimum circumscribed rectangle of the non-overlapping mutation point; The potential defect area is intercepted and partially enlarged to obtain a third image; the third image is detected using the trained YOLOV7 model to obtain actual defects; and the surface defect value is obtained according to the number and size of the actual defects; Obtaining the content ratio, physical strength and chemical content of the knitted fabric, and calculating a score value according to the content ratio, the physical strength and the chemical content; normalizing the score value and the surface defect value to obtain a first value, a second value, a third value and a fourth value; collecting user satisfaction with the knitted fabric at fixed time intervals, and calculating, according to the satisfaction, the first value, the second value, the third value, and the fourth value, the correlation coefficient between the satisfaction and the first value, the second value, the third value, and the fourth value using the Pearson correlation coefficient method; The correlation coefficient is recorded as a weight coefficient, and the score value and the surface defect value are weighted and summed according to the weight coefficient to obtain a quality assessment result.

2. The knitted fabric defect detection method based on projection method according to claim 1 is characterized in that: The process of acquiring the first image includes: Graying the surface image of the knitted fabric to obtain a grayscale image; Performing denoising on the grayscale image using Gaussian filtering to obtain a denoised image; The denoised image is classified using a trained CNN model to obtain the jacquard fabric image and the non-jacquard fabric image.

3. The method for detecting knitted fabric defects based on projection method according to claim 1, characterized in that: The processing process of the high frequency area includes: assigning the first diffusion coefficient to the high frequency region; Calculating the gradient of each pixel in the high-frequency area to obtain an edge intensity value; The diffusion intensity is calculated by using the selective diffusion function according to the first diffusion coefficient and the edge intensity value; The high-frequency region is subjected to diffusion processing using the diffusion equation of the PM model according to the diffusion intensity to obtain the processed high-frequency region.

4. The method for detecting knitted fabric defects based on projection method according to claim 1, characterized in that: The processing process of the low frequency region includes: assigning the second diffusion coefficient to the low-frequency region, the second diffusion coefficient being greater than the first diffusion coefficient; The low-frequency region is subjected to diffusion processing using the diffusion equation of the PM model according to the second diffusion coefficient to obtain the processed low-frequency region.

5. The method for detecting knitted fabric defects based on projection method according to claim 1, characterized in that: The integral curve acquisition process includes: The different projection angles are selected, and the projection value is calculated according to the different projection angles and the gray value of the second image, which is expressed as: ； in, for Angle at distance The projection value at ; For the second image at point The gray value of and are respectively the horizontal and vertical coordinates of the second image; It is along Angle at distance The set of pixels at ; Will Set as the horizontal coordinate of the curve, The vertical coordinate of the curve is set, and the integral curve is obtained by plotting the horizontal coordinate of the curve and the vertical coordinate of the curve.

6. The method for detecting knitted fabric defects based on projection method according to claim 1, characterized in that: The mutation point acquisition process includes: Acquire the integral curves at the different projection angles, calculate the first-order derivative for each integral curve, and obtain the derivative values ​​at different positions; If the derivative value exceeds the first threshold, the coordinate position corresponding to the derivative value is marked according to the integral curve, and the coordinate position is recorded as the mutation point.

7. The method for detecting knitted fabric defects based on projection method according to claim 1, characterized in that: The potential defect area marking process includes: Collecting all the mutation points and combining them into a mutation point set; performing the intersection processing on the mutation point set to obtain the overlapping mutation points and the non-overlapping mutation points; Taking the overlapping mutation point and the non-overlapping mutation point as centers, respectively draw a first rectangle and a second rectangle, wherein the side length of the first rectangle is greater than the side length of the second rectangle; If there is an overlapping part between the first rectangles, the overlapping part between the second rectangles, or the overlapping part between the first rectangle and the second rectangle, the boundary of the minimum circumscribed rectangle is calculated based on the overlapping part, and the potential defect area is obtained based on the boundary of the minimum circumscribed rectangle.

8. The method for detecting knitted fabric defects based on projection method according to claim 1, characterized in that: The training steps of the YOLOV7 model include: Step S1: collecting historical knitted fabric images containing defects, and marking the defect areas and defect types in the historical knitted fabric images to obtain a first data set; Step S2: inputting the first data set into a basic model, and the basic model outputting a predicted value for the first data set; Step S3: using a loss function to calculate a loss value according to the predicted value, the defect area and the defect type; Step S4: using the back propagation method to calculate the gradient information of the loss value; using the Adam optimizer to update the network parameters of the basic model according to the gradient information; Step S5: If the number of training rounds of the basic model is equal to the preset number of rounds, the training of the basic model is terminated, otherwise, steps S1 to S5 are repeated.

9. A knitted fabric defect detection system based on projection method, characterized in that: A knitted fabric defect detection method based on projection method for implementing any one of claims 1 to 8, comprising: An image preprocessing module, used for collecting a knitted fabric surface image, and obtaining a first image after preprocessing the knitted fabric surface image; An image reprocessing module, if the first image is a non-jacquard fabric image, directly record the first image as the second image; if the first image is a jacquard fabric image, use Gabor filtering to extract the texture features of the jacquard fabric image to obtain a high-frequency area and a low-frequency area; use the PM model to perform diffusion processing on the high-frequency area and the low-frequency area respectively, and combine the processed high-frequency area and the low-frequency area to obtain the second image; A surface defect calculation module is used to perform grayscale integral projection on the second image using a multi-directional projection method to obtain integral curves at different projection angles; calculate the first-order derivative of the integral curve to obtain a mutation point; perform intersection processing on all the mutation points to obtain a potential defect area; intercept the potential defect area to obtain a third image; use the trained YOLOV7 model to detect the third image to obtain actual defects; and obtain a surface defect value according to the number and size of the actual defects; The comprehensive defect calculation module is used to obtain the content ratio, physical strength and chemical content of the knitted fabric, calculate the score values ​​of the content ratio, the physical strength and the chemical content respectively, and perform weighted summation of the score value and the surface defect value to obtain a quality assessment result.

Images