A fabric periodic texture detection method based on deep learning

Abstract

The present invention relates to a fabric periodic texture detection method based on deep learning, which is used for automatically detecting the periodic texture primitives of fabrics. Specifically, the input fabric image is first calibrated using the tilt transform and the non-local variation algorithm. During the detection process, the pre-processed image is input into the network. After the convolution operation and the non-maximum suppression algorithm, an activation peak that follows the feature rules is generated in each feature layer. The position of the periodic texture primitive is further found through the Hough voting strategy and the centroid coordinates. The present invention is based on deep learning technology, and replaces the traditional manually designed key point detection, feature extraction and clustering for fabric periodic texture detection. By fusing multiple activations of different levels and scales, it can capture higher-level pixel and regional information, and has stronger versatility and robustness.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial visual inspection, and in particular to a method for detecting periodic texture of fabrics based on deep learning. Background Art

[0002] Fabrics are fundamental materials for modern industries such as clothing, medical devices, and the decorative industry. During textile processing and testing, a large number of issues involve fabric surface texture and morphology. Periodicity is one of the most important visual characteristics of texture images, and printed fabrics are a typical example. Printed fabrics are generally composed of periodic texture patterns, while the original patterns are usually composed of identical elements. Because the unit pixels that make up digital images are very small, the human eye cannot accurately identify unit pixels. To meet the needs of digital manufacturing in the textile industry, it is necessary to find a method to automatically and accurately extract periodic texture patterns from printed fabrics to accelerate the pattern design process or improve the accuracy of defect detection. Generally speaking, fabric images can be considered to be composed of similar, periodic texture primitives that can be removed by carefully designed filters. Therefore, traditional algorithms often use handcrafted features, such as the invention patent application CN201010205357.5 titled "A Method for Automatic Measurement of Texture Periodicity in Fabric Images." This method applies a fast Fourier transform to the row and column vectors of a grayscale fabric image to obtain the corresponding continuous spectrum. The primary and secondary periods corresponding to the maximum and second-largest peak points of the continuous spectrum are extracted, and the corresponding distance matching function values ​​are calculated to determine the final period number. Application number: CN201710216317.2. Name: Invention patent for a method and device for detecting defects in periodic texture images. The method involves subtracting a source image from a first translation image and a second translation image obtained by translating the source image to both sides by N periods in the periodic direction, binarizing the subtraction results, and obtaining a first subtraction image and a second subtraction image. The first subtraction image and the second subtraction image are then ANDed to obtain a defect image. Terzopoulus et al. (Finding structure in co-occurrence matrices for texture analysis [J] Computer graphics and image processing 12 (3) 286-308) attempted to use gray-level co-occurrence matrices to analyze periodic texture structures, but the robustness was poor. Lin et al. (Extracting periodicity of a regular texture based on autocorrelation functions [J] North-Holland 18 (5) 433-443) used the autocorrelation function method to directly extract texture periodicity by matching the self-similarity of image functions in the spatial domain. This method has a certain tolerance to local distortion, but requires smoothing the autocorrelation function and takes a long time to calculate. The method of using sum and difference histograms (SDH) for texture analysis was proposed by Unser et al. (Sum and Difference Histograms for Texture Classification [J] IEEE Transactions on Pattern Analysis & Machine Intelligence 8 (1) 118.), which greatly shortens the calculation time. Matsuyama et al. (Structural analysis of natural textures by Fourier transformation [J] Computer Vision Graphics & Image Processing 24(3)347-362.) Using Fourier transform to analyze texture, the period of a repeating pattern element can be determined by the pulse distribution of the Fourier spectrum. However, when the pattern contains fewer periodic elements, the Fourier spectrum peak is not significant; Kuo et al. (Repeat Pattern Segmentation of Printed Fabrics by Hough Transform Method[J]Textile Research Journal 75(11)779-783.) converted the printed fabric into a full-color image and used the fuzzy mean clustering algorithm to segment the pattern to obtain the printed fabric pattern. The Hough transform was used to complete the segmentation of the periodic pattern primitives. Dekel et al. (Revealing and modifying non-local variations in a single image [J] ACM Transactions on Graphics (TOG) 34 (6) 1-11.) proposed the HCDH method to convert the color image into a grayscale image. The uniform features were then calculated using the sum difference histogram (SDH) algorithm. The size of the periodic primitive was determined based on the distance between adjacent local maximum uniform values.

[0003] Overall, most periodic texture extraction and detection methods are essentially based on classical statistics or hand-crafted features. They fail to consider the spatial dependencies between pixels and their neighborhoods, are limited to simple periodic texture patterns, and fail to account for the rich texture features and colors of fabrics. Their adaptability is poor and cannot meet the needs of the textile industry. Therefore, it is necessary to propose a new fabric periodic texture detection method based on deep learning technology to replace the traditional hand-crafted feature extraction, description, and clustering stages, achieving greater robustness. Summary of the Invention

[0004] In view of the above-mentioned deficiencies in the prior art, the purpose of the present invention is to provide a method for detecting periodic texture of fabrics based on deep learning. It uses deep learning technology as the basis to detect periodic texture of fabrics, and is used to replace the traditional manually designed key point detection, feature extraction and clustering to achieve stronger robustness.

[0005] The present invention discloses a method for detecting periodic texture of fabric based on deep learning, which specifically includes the following steps:

[0006] Step 1: To reduce the angle deviation of the industrial camera and the interference of dust and wrinkles on the fabric surface, the input fabric image is first calibrated using the tilt transformation and non-local variation algorithm;

[0007] Step 2: Determine the primitives of the fabric periodic texture and obtain the displacement vector, specifically:

[0008] Step 2.1: Input the calibrated image into the trained network. After convolution and non-maximum suppression, an activation peak that follows the feature rules is generated at each feature layer.

[0009] Step 2.2: Let convolutional layer l∈L, where L is the set of convolutional layers, let f l ∈F l Represents a filter, where F lis a set of filters, let is the filter f l The activation peak position vector of is f l The set of activation peak positions, for each pair p i , Both compute a set of displacement vectors d representing the potential size of the periodic texture primitives i,j , the formula is shown in formula (1):

[0010]

[0011] Among them, |p i -p j | represents the absolute value of the elements of the vector, is the filter f l The set of displacement vectors;

[0012] Step 2.3: Use Hough space H to fuse the vectors of the connected activation peak pairs; each vote is subject to d i,j The two-dimensional normal distribution centered on , the voting process formula is shown in formula (2) and formula (3):

[0013]

[0014]

[0015] Where (x,y) corresponds to the pixel coordinates of the input image, σ l is the covariance matrix of convolutional layer l;

[0016] Step 2.4: Extract the most consistent displacement vector d * As the maximum value of the voting space on the x-axis and y-axis, it is specifically as follows:

[0017] d * =(argmax x H (x,0) ,argmax y H (0,y) ) (4)

[0018] Step 3: Select the filter, specifically:

[0019] Step 3.1: Select the displacement vector d * The filter with the most consistent activation peak is set to * Consistent displacement vector d i,j The set of all votes for The formula is shown in formula (5):

[0020]

[0021] Among them, γ is based on The average prior estimate of the expected number of repetitions calculated from the distribution of

[0022] Step 3.2: Unanimous votes are weighted The weight of a particular filter is given as the sum of the weights of its consistent votes, as shown in Equations (6) and (7):

[0023]

[0024]

[0025] where β represents the radius of the immediate neighborhood around the selected displacement vector at level l;

[0026] Step 3.3: Use weights Sort the filters in each layer to select a consistent set of filters that will participate in the deep learning model for discriminating periodic textures

[0027] Step 4: Segment the periodic texture primitives, specifically: simplify the two-dimensional rectangular coordinates (x, y) of the filter to obtain the relative position of the displacement vector vote; use the IPM algorithm to obtain the centroid coordinates of the periodic pattern primitives; and segment the final periodic texture primitives using the centroid coordinates and the previously determined primitive size.

[0028] Furthermore, the present invention also defines the network structure in step 2.1 to select a spatial pyramid aggregation module to learn multi-scale features, and the convolution module is composed of a convolution layer C i|i∈{0,1,...,4} The composition is, specifically, 96, 256, 384, 384, and 256 filters, and the filter sizes are (11×11), (5×5), (3×3), (3×3), and (3×3), respectively.

[0029] By adopting the above technology, compared with the existing technology, the beneficial effects of the present invention are as follows: the present invention adopts deep learning technology to detect the periodic texture of fabrics, replacing the traditional manual design of key point detection, feature extraction and clustering, and by fusing multiple activations of different levels and scales to achieve the purpose of capturing higher-level pixel and regional information, it has stronger versatility and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 Schematic diagram of a fabric image acquisition device;

[0031] Figure 2 This is the test result diagram of periodic texture fabric.

[0032] In the figure: 1. Workbench; 2. Fabric; 3. Light source; 4. Industrial camera. DETAILED DESCRIPTION

[0033] The present invention will be further described below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to illustrate the present invention and are not intended to limit the present invention.

[0034] The fabric image acquisition device used in the fabric periodic texture detection based on deep learning of the present invention is as follows: Figure 1 As shown, the fabric image acquisition device includes a workbench 1, a light source 3 and an industrial camera 4. The industrial camera 4 is arranged above the workbench 1, and the light source 3 is arranged on both sides of the industrial camera 4. When in use, the fabric 2 is placed on the workbench 1, the light source 3 is turned on, and the fabric is photographed by the industrial camera 4 directly above.

[0035] The fabric periodic texture detection method based on deep learning of the present invention specifically includes the following steps:

[0036] Step 1: Place fabric 2 on workbench 1 and turn on light source 3. To reduce the angle deviation of the industrial camera 4 and the interference of dust and wrinkles on the fabric surface, the input fabric image is first calibrated using tilt transformation and non-local variation algorithm.

[0037] Step 2: To determine the primitives of the fabric periodic texture, first obtain the displacement vector, specifically:

[0038] Step 2.1: Input the image calibrated in step 1 into the trained network. The network structure uses the spatial pyramid pooling module to learn multi-scale features. The convolution module consists of the convolution layer C i|i∈{0,1,...,4} The network is composed of 96, 256, 384, 384, and 256 filters, with filter sizes of (11×11), (5×5), (3×3), (3×3), and (3×3), respectively. After convolution and non-maximum suppression, activation peaks that follow feature rules are generated at each feature layer.

[0039] Step 2.2: Let convolutional layer l∈L, where L is the set of convolutional layers, let f l ∈F l Represents a filter, where F l is a set of filters, let is the filter f l The activation peak position vector of is f l The set of activation peak positions; for each pair of p i , Both compute a set of displacement vectors d representing the potential size of the periodic texture primitives i,j , the formula is as follows:

[0040]

[0041] Among them, |p i -p j | represents the absolute value of the elements of the vector, is the filter f l The set of displacement vectors;

[0042] Step 2.3: Use a Hough space H to fuse the vectors of the connected activation peak pairs; each vote is subject to d i,j The voting process is a two-dimensional normal distribution centered on . The formula for the voting process is as follows:

[0043]

[0044]

[0045] Where (x,y) corresponds to the pixel coordinates of the input image, σ l is the covariance matrix of convolutional layer l;

[0046] Step 2.4: Extract the most consistent displacement vector d * As the maximum value of the voting space on the x-axis and y-axis, it is as follows:

[0047] d * =(argmax x H (x,0) ,argmax y H (0,y) ) (4)

[0048] Step 3: Select the appropriate filter, specifically:

[0049] Step 3.1: Select the displacement vector d * The filter with the most consistent activation peak is set to * Consistent displacement vector d i，j The set of all votes for The formula is as follows:

[0050]

[0051] where γ is based on The average prior estimate of the expected number of repetitions calculated from the distribution of

[0052] Step 3.2: Unanimous votes are weighted The weight of a particular filter is given as the sum of the weights of its consistent votes, as shown in Equations (6) and (7):

[0053]

[0054]

[0055] where β represents the radius of the immediate surrounding neighborhood of the selected displacement vector at level l;

[0056] Step 3.3: Use weights Sort the filters in each layer to select a consistent set of filters that will participate in the deep learning model for discriminating periodic textures

[0057] Step 4: Segment the periodic texture primitives. The segmentation results are as follows: Figure 2 As shown in the figure, the specific operation steps are: simplify the two-dimensional rectangular coordinates (x, y) of the filter to obtain the relative position of the displacement vector vote; use the IPM algorithm to obtain the centroid coordinates of the periodic pattern primitive; and segment the final periodic texture primitive by the centroid coordinates and the previously determined primitive size.

[0058] The contents described in the embodiments of this specification are merely an enumeration of the implementation forms of the inventive concept. The scope of protection of the present invention should not be regarded as limited to the specific forms described in the embodiments. The scope of protection of the present invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A method for detecting periodic texture of fabric based on deep learning, characterized in that The steps include: Step 1: Place the fabric (2) on the workbench (1), turn on the light source (3), and use the industrial camera (4) directly above to shoot the fabric (2). In order to reduce the angle deviation of the industrial camera (1) when shooting and the interference of dust and wrinkles on the fabric surface, the tilt transformation and non-local variation algorithm are used to calibrate the input fabric image; Step 2: Determine the primitives of the fabric periodic texture. First, obtain the displacement vector, specifically: Step 2.1: Input the image calibrated in step 1 into the trained network. After convolution and non-maximum suppression, an activation peak that follows the feature rules is generated at each feature layer. Step 2.2: Let convolutional layer l∈L, where L is the set of convolutional layers, let f l ∈F l Represents a filter, where F l is a set of filters, let is the filter f l The activation peak position of is f l The set of activated peak positions; for each pair Both compute a set of displacement vectors d representing the potential size of the periodic texture primitives i，j , the formula is as follows: Among them, |p i -p j | represents the absolute value of the elements of the vector, is the filter f l The set of displacement vectors; Step 2.3: Use Hough space H to fuse the vectors of the connected activation peak pairs; each vote is subject to d i，j The voting process is a two-dimensional normal distribution centered on . The formula for the voting process is as follows: Where (x, y) corresponds to the pixel coordinates of the input image, σ l is the covariance matrix of convolutional layer l; Step 2.4: Extract the most consistent displacement vector d * As the maximum value of the voting space on the x-axis and y-axis, it is as follows: d * =(argmax x H (x，0) ,argmax y H (0，y) ) (4) Step 3: Select the filter, specifically: Step 3.1: Select the displacement vector d * The filter with the most consistent activation peak is set to * Consistent displacement vector d i，j The set of all votes for The formula is as follows: Among them, γ is based on The average prior estimate of the expected number of repetitions calculated from the distribution of Step 3.2: Unanimous votes are weighted The weight of a particular filter is given as the sum of the weights of its consensus votes, as follows: where β represents the radius of the immediate neighborhood around the selected displacement vector at level l; Step 3.3: Use weights Sort the filters in each layer to select a consistent set of filters that will participate in the deep learning model for discriminating periodic textures Step 4: Segment the periodic texture primitives.

2. A fabric periodic texture detection method based on deep learning according to claim 1, characterized in that The network structure in step 2.1 uses a spatial pyramid pooling module to learn multi-scale features, and the convolution module consists of a convolution layer C i|i∈{0，1，...，4} The composition is, specifically, 96, 256, 384, 384, and 256 filters, and the filter sizes are (11×11), (5×5), (3×3), (3×3), and (3×3), respectively.

3. The method for detecting periodic texture of fabric based on deep learning according to claim 1, characterized in that The specific steps of segmenting the periodic texture primitives in step 4 are as follows: simplifying the two-dimensional rectangular coordinates (x, y) of the filter to obtain the relative position of the displacement vector vote; using the IPM algorithm to obtain the centroid coordinates of the periodic pattern primitives; and segmenting the final periodic texture primitives using the centroid coordinates and the previously determined primitive size.

Images