A curtain fabric defect recognition method based on feature cube and fuzzy detection dual recognition

Abstract

The present invention discloses a curtain fabric defect recognition method based on feature cube and fuzzy detection dual recognition, comprising the following steps: capturing curtain fabric images at different angles, and performing multi-angle synthesis to obtain a curtain fabric panoramic image; obtaining material features and texture features, and correlating them with exposure levels to obtain a curtain fabric feature cube; graying the panoramic image, enhancing it based on a deep learning-based adaptive image enhancement algorithm, and adjusting the enhancement parameters to obtain an enhanced curtain fabric panoramic image; filtering and segmenting the enhanced image to obtain a fabric dynamic segmentation map, and inputting it into a detection model to identify fabric and texture defects, and obtaining a first recognition result; optimizing the edge of the enhanced image, and inputting it into an edge detection model to identify edge defects, and obtaining a second recognition result. The present invention achieves the purpose of improving the accuracy, robustness, adaptability and dynamic adjustment capability of curtain fabric defect detection.

Description

Technical Field

[0001] The present invention relates to the field of intelligent recognition technology, and in particular to a curtain fabric defect recognition method based on dual recognition of feature cube and fuzzy detection. Background Art

[0002] During the manufacturing process of curtain fabrics, due to the influence of factors such as material properties, processing technology or external environment, different types of defects often occur, such as yarn defects, texture deviations, irregular edges, etc. These defects not only affect the appearance quality of curtain fabrics, but also have an adverse effect on subsequent processing and application. Therefore, high-precision recognition and detection technology for curtain fabric defects has become an important research direction in the current field of intelligent manufacturing and quality control.

[0003] Existing curtain fabric defect detection methods mainly include: manual detection and image processing detection. Image processing detection methods usually detect fabric defects through algorithms such as grayscale detection, edge extraction or texture analysis. However, these methods have low recognition capabilities for fabric defects with complex textures, especially under multiple lighting conditions, different exposure levels or complex materials. Their detection accuracy and robustness are difficult to meet actual needs. In addition, image processing detection methods are highly dependent on specific types of defects and cannot effectively adapt to the comprehensive detection of multiple types of defects.

[0004] With the development of artificial intelligence technology, deep learning-based defect detection methods have gradually emerged. These methods use convolutional neural networks (CNN) or deep neural networks (DNN) to extract features and classify defects from fabric images, greatly improving detection accuracy and generalization capabilities. However, such methods still have the following problems:

[0005] (1) Insufficient processing capabilities for multi-exposure level and multi-angle images: Defects in curtain fabrics are often affected by lighting conditions and shooting angles. Different exposure levels and shooting angles may cause defect features to be lost or blurred. Existing deep learning methods lack an effective feature fusion mechanism when processing multi-exposure level and multi-angle images, resulting in unstable detection results.

[0006] (2) Insufficient effective extraction of material and texture features: There are many types of curtain fabrics, and the material and texture characteristics vary greatly. Existing methods are unable to fully mine and utilize the high-level semantic features of curtain fabrics, making it difficult to effectively identify defects in fabrics of different materials and textures.

[0007] (3) Low detection accuracy of edge defects: The detection of edge defects usually requires high-precision edge processing of fabric images. However, existing methods are easily affected by texture interference and noise during edge optimization and detection, resulting in poor edge defect recognition results.

[0008] (4) Limited real-time feedback and parameter adaptation capabilities: In the actual production process, different batches of fabrics may have differences in lighting, materials, and textures. Existing methods usually rely on fixed detection models and parameters, lack the ability of real-time feedback and dynamic adjustment, and are difficult to adapt to complex and changing actual scenarios. Summary of the invention

[0009] In order to overcome the shortcomings of the prior art, the present invention provides a curtain fabric defect recognition method based on dual recognition of feature cube and fuzzy detection, which is used to solve the above-mentioned technical problems existing in the existing deep learning-based defect detection method for curtain fabric defect detection, thereby achieving the purpose of improving the accuracy, robustness, adaptability and dynamic adjustment capability of curtain fabric defect detection.

[0010] To solve the above problems, the technical solution adopted by the present invention is as follows:

[0011] A curtain fabric defect recognition method based on feature cube and fuzzy detection dual recognition includes the following steps:

[0012] Capturing curtain fabric images at different angles at multiple exposure levels, and performing multi-angle synthesis for each exposure level to obtain a panoramic image of the curtain fabric at each exposure level;

[0013] Obtain material features and texture features according to the model of the curtain fabric to be identified, and correlate them with the exposure level to obtain a feature cube of the curtain fabric at each exposure level;

[0014] After the panoramic image is grayed out, it is enhanced using an adaptive image enhancement algorithm based on deep learning, and the enhancement parameters are dynamically adjusted through a real-time feedback mechanism to obtain an enhanced panoramic image of the curtain fabric at each exposure level.

[0015] Based on the feature cube of curtain fabric, the enhanced image is filtered and the local dynamic area threshold segmentation is performed to obtain the dynamic segmentation map of the fabric and input it into Type fuzzy - Identify fabric defects and texture defects in the detection model to obtain a first identification result;

[0016] The enhanced image is optimized at the edge by using an edge optimization algorithm, and is input into an edge detection model to identify edge defects, thereby obtaining a second identification result;

[0017] The first and second recognition results are attribute-mapped to form an interactive defect map for displaying the location and type of each defect.

[0018] As a preferred embodiment of the present invention, when performing multi-angle synthesis, it includes:

[0019] The phase correlation method is used to obtain the displacement between two adjacent angle images at the same exposure level and determine the overlapping area;

[0020] Extract in the overlapped area Features, and use similarity measurement method to perform rough matching of feature points;

[0021] According to the plane projection model, The method estimates the perspective transformation matrix between the images to be stitched, and optimizes the matching point pairs to complete the transformation alignment;

[0022] The transformed image is decomposed into multi-scale space through the curtain fabric image pyramid, and weighted fusion is performed using the fade-in and fade-out weighted method according to the horizontal displacement of the image.

[0023] As a preferred embodiment of the present invention, when optimizing the matching point pairs, it includes:

[0024] Using the first curtain fabric image collected as a reference image, obtaining a set of errors of matching point pairs after transformation relative to the reference image;

[0025] The errors after transformation of all groups of matching point pairs are summed based on the error function to obtain the total error value;

[0026] By minimizing the total error value, the homography matrix of an angle image to the reference image is adjusted to optimize the matching point pairs.

[0027] As a preferred embodiment of the present invention, when obtaining a curtain fabric feature cube, it includes:

[0028] The texture features are defined by a texture node, wherein the URL field of the texture node is associated with a standard texture map;

[0029] The material characteristics are defined through the Material class node in the appearance object, and the URL field of the Material class node is associated with a standard material map;

[0030] The material characteristics, texture characteristics and exposure levels defined are respectively , , The axes form an interrelated feature cube as high-level semantic features.

[0031] As a preferred embodiment of the present invention, when performing enhancement, it includes:

[0032] Segmenting the curtain fabric panoramic image by a threshold method to obtain curtain fabric sub-image blocks in bright and dark areas, and obtaining average brightness responses in the bright and dark areas;

[0033] The average brightness response of the bright and dark areas is combined with the adaptive gamma factor to obtain the curtain fabric enhancement sequence, as shown in Formula 5:

[0034] (5);

[0035] In the formula, is the intensity value of each pixel in the curtain fabric panoramic image, is the maximum pixel intensity value, is the adaptive gamma factor, Enhanced sequence for curtain fabrics;

[0036] Wherein, the first adaptive gamma factor is determined according to the pixel probability density and the weighted probability density, as shown in Formula 6 and Formula 7;

[0037] (6);

[0038] (7);

[0039] In the formula, , is the average brightness response of bright and dark areas, is the weighted probability density, is the pixel probability density, , are the first and second adaptive indexes, To take the greater of the two, , is the maximum and minimum pixel probability density, is the first adaptive gamma factor;

[0040] According to the first adaptive gamma factor, the second and third adaptive gamma factors are obtained, as shown in Formula 8:

[0041] (8);

[0042] In the formula, , are the second and third adaptive gamma factors.

[0043] As a preferred implementation mode of the present invention, when dynamically adjusting the enhancement parameters, it includes:

[0044] Input the curtain fabric enhancement sequence into a fully convolutional network for weight map prediction, upsample the weight map through a guided filter, and dynamically adjust the upsampling parameters through a real-time feedback mechanism;

[0045] Wherein, the real-time feedback mechanism includes:

[0046] Initialize the first parameter , the second parameter ;

[0047] Get the fitness of each individual and sort them to determine the best position ;

[0048] Update the current individual position, and update the current individual position again;

[0049] Update the first parameter , the second parameter , update the best position and return;

[0050] Among them, the first parameter To avoid collision variables, the second parameter A random variable that balances exploration and exploitation.

[0051] As a preferred embodiment of the present invention, when filtering and local dynamic area threshold segmentation are performed, it includes:

[0052] Obtaining an average grayscale value of the enhanced curtain fabric panoramic image at each exposure level, and performing detection based on a grayscale detection condition;

[0053] Based on the exposure level information contained in the curtain fabric feature cube, the enhanced curtain fabric panoramic image at the corresponding exposure level that does not meet the grayscale detection condition is eliminated;

[0054] The enhanced curtain fabric panoramic image of the remaining exposure levels is subjected to local dynamic region threshold segmentation to obtain a fabric dynamic segmentation map.

[0055] As a preferred embodiment of the present invention, when identifying cloth defects and texture defects, it includes:

[0056] Input the cloth dynamic segmentation map into Type fuzzy - The input layer of the detection model is used, and the material, texture, and exposure level features in the associated curtain fabric feature cube are used as inputs to the three neuron nodes of the input layer respectively;

[0057] The output of the input layer is received by the nodes of the membership layer, and the Type fuzzy membership function, obtain the corresponding membership value, and output it to the output layer, as shown in Formula 15:

[0058] (15);

[0059] In the formula, is the output of the membership layer, The first Output node, symbol superscript represents the membership layer, , is the input variable of the upper and lower membership functions The membership degree of For the Input variables, The average value of the membership function of the term, , For the Input variables, The standard deviation of the upper and lower membership functions of the term;

[0060] The output layer provides a mapping from input nodes to output nodes, and uses elastic back-propagation training to adjust parameters to optimize the recognition results of cloth defects and texture defects.

[0061] As a preferred embodiment of the present invention, when performing edge optimization, it includes:

[0062] The edge optimization value is obtained based on the high-frequency amplitude of the wavelet decomposition of the edge area of ​​the enhanced curtain fabric panoramic image. ;

[0063] Based on the edge optimization value, the maximum value of the sub-bands to be enhanced in different directions, and the nonlinearity, and generate an edge enhancement function to enhance the edge area;

[0064] Based on the entropy values ​​and edge signal modulus values ​​of the smooth area and the texture area of ​​the enhanced curtain fabric panoramic image, soft threshold denoising is performed on the smooth area and the texture area.

[0065] As a preferred embodiment of the present invention, when edge defects are identified, the following steps are included:

[0066] Will The network is used as the main framework for target classification and bounding box regression tasks, and adopts As the backbone network, build the initial lightweight model;

[0067] Add after the feature extraction network of the last layer of the initial lightweight model An attention mechanism module is used to obtain the edge detection model;

[0068] The edge-optimized curtain fabric panoramic image is input into the edge detection model. The channel attention submodule in the attention mechanism module generates a channel attention map to enhance important channel features, and the spatial attention submodule generates a spatial attention map to enhance important spatial features.

[0069] Compared with the prior art, the present invention has the following beneficial effects:

[0070] (1) Efficient fusion of multi-exposure level and multi-angle images

[0071] The present invention captures curtain fabric images at different angles at multiple exposure levels, and adopts phase correlation method, multi-scale pyramid decomposition and fade-in and fade-out weighted fusion technology to generate high-quality panoramic images of curtain fabrics. It can effectively retain the defect features of the fabric under different lighting conditions and angles, overcome the problem of defect feature loss caused by changes in lighting and angles in traditional methods, and improve the stability and robustness of the detection results.

[0072] (2) Feature depth correlation between material, texture and exposure level

[0073] By constructing a feature cube that correlates material features, texture features, and exposure levels, the high-level semantic features of curtain fabrics are extracted, which can more accurately characterize the diverse material and texture information of fabrics and meet the needs of defect detection under various fabric types. Compared with existing methods, the present invention can more comprehensively explore the characteristics of fabrics and effectively improve the accuracy of defect identification.

[0074] (3) High-precision defect classification and detection capabilities

[0075] By combining The present invention realizes the accurate classification and detection of fabric defects, texture defects and edge defects by combining a fuzzy detection model and an elastic back propagation optimization mechanism of deep learning. The membership degree of the dynamic segmentation map of the cloth is calculated by using a fuzzy membership function, which can better adapt to the fuzziness and diversity of defect characteristics and significantly enhance the detection ability of complex textures and tiny defects.

[0076] (4) Optimized detection of edge defects

[0077] The present invention adopts an edge optimization algorithm based on high-frequency amplitude calculation of wavelet decomposition combined with edge enhancement function, and introduces an attention mechanism module to enhance channel and spatial features, making defect detection in edge areas more accurate. Compared with the problem that traditional methods are susceptible to texture interference, the present invention has stronger noise resistance and recognition accuracy in edge processing, and is particularly suitable for defect recognition in edge areas.

[0078] (5) Adaptive image enhancement and real-time feedback mechanism

[0079] Through the deep learning adaptive image enhancement algorithm and the real-time feedback mechanism of dynamically adjusting parameters, the present invention can dynamically adjust the enhancement parameters and detection models according to the characteristics and defect features of different fabrics to adapt to various production scenarios. The real-time feedback mechanism can optimize the convolutional network weights and enhancement parameters, improving the efficiency and robustness of defect detection.

[0080] (6) Improve the comprehensiveness and intuitiveness of defect detection. The present invention generates an interactive defect map by mapping the recognition results in the form of attribute mapping, intuitively displays the type and location of defects, and supports real-time query, editing and management by users. Combined with defect distribution and statistical analysis tools, data reports can be further generated to provide a scientific basis for production process optimization and quality control.

[0081] (7) Enhance detection efficiency and system self-learning ability

[0082] Through real-time feedback and The present invention optimizes the fuzzy detection model, and significantly improves the detection efficiency while improving the detection accuracy. In addition, the model realizes self-learning ability by dynamically adjusting parameters, and can be continuously optimized in the actual production environment, reducing manual intervention, adapting to the differences of different batches of fabrics, and reducing the overall detection cost.

[0083] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] Figure 1 It is a step diagram of the curtain fabric defect recognition method based on feature cube and fuzzy detection dual recognition provided by the present invention;

[0085] Figure 2 It is an overall structural diagram of the edge detection model provided by the present invention.

[0086] Description of the accompanying figures: 1. Initial lightweight model; 2. Attention mechanism module; 3. Channel attention sub-module; 4. Spatial attention sub-module. DETAILED DESCRIPTION

[0087] The curtain fabric defect recognition method based on feature cube and fuzzy detection dual recognition provided by the present invention is as follows: Figure 1 As shown, the following steps are included:

[0088] Step S1: capturing curtain fabric images at different angles at multiple exposure levels, and performing multi-angle synthesis for each exposure level to obtain a panoramic image of the curtain fabric at each exposure level;

[0089] Step S2: obtaining material features and texture features according to the model of the curtain fabric to be identified, and correlating the material features, texture features, and exposure levels to obtain a curtain fabric feature cube at each exposure level;

[0090] Step S3: After the curtain fabric panoramic image is grayed, it is enhanced by an adaptive image enhancement algorithm based on deep learning, and the enhancement parameters are dynamically adjusted through a real-time feedback mechanism to obtain an enhanced curtain fabric panoramic image at each exposure level;

[0091] Step S4: Filter the enhanced curtain fabric panoramic image based on the curtain fabric feature cube, perform local dynamic region threshold segmentation, obtain the fabric dynamic segmentation map, and input it into Type fuzzy - Identify fabric defects and texture defects in the detection model to obtain a first identification result;

[0092] Step S5: optimizing the edge of the enhanced curtain fabric panoramic image by using an edge optimization algorithm, and inputting the image into an edge detection model to identify edge defects, thereby obtaining a second identification result;

[0093] Step S6: Mapping the attributes of the first recognition result and the second recognition result to form an interactive defect map for displaying the location and type of each defect.

[0094] In the above step S1, when performing multi-angle synthesis for each exposure level, it includes:

[0095] The phase correlation method is used to obtain the displacement between two adjacent angle images at the same exposure level and determine the overlapping area;

[0096] Extract in overlapping area Features, and use similarity measurement method to perform rough matching of feature points;

[0097] According to the plane projection model, The method estimates the perspective transformation matrix between the images to be stitched, and optimizes the matching point pairs to complete the transformation alignment;

[0098] The transformed image is decomposed into multi-scale space through the curtain fabric image pyramid, and weighted fusion is performed using the fade-in and fade-out weighted method according to the horizontal displacement of the image;

[0099] Among them, the curtain fabric image pyramid includes Group, each group has Layer, the next group of images is obtained by downsampling the previous group of images, and each previous layer in each group is obtained by Laplace transforming the next layer;

[0100] When obtaining the displacement, the cross-power spectrum between two adjacent angle images is obtained in the frequency domain, and then transformed to the spatial domain through inverse Fourier transform to obtain the position of the maximum value of the impulse response function.

[0101] Furthermore, when optimizing the matching point pairs, it includes:

[0102] The first curtain fabric image collected is used as a reference image to obtain a set of errors after the matching point pairs are transformed relative to the reference image, as shown in Formula 1 and Formula 2:

[0103] (1);

[0104] (2);

[0105] In the formula, Angle image To the adjacent angle image The homography matrix of Adjacent angle images to the homography matrix of the reference image, Angle image to the homography matrix of the reference image, , Angle image and adjacent angle images The , feature points, for Images from adjacent angles To an angle image The projection of is the error after transformation;

[0106] The errors after transformation of all groups of matching point pairs are summed based on the error function to obtain the total error value, as shown in Formula 3 and Formula 4:

[0107] (3);

[0108] (4);

[0109] In the formula, Angle image and adjacent angle images The set of matching point pairs between is the error function, is the total error value, is the preset number of pixels;

[0110] By minimizing the total error value, the homography matrix from an angle image to the reference image is adjusted to optimize the matching point pairs.

[0111] Preferably, the preset number of pixels is 1-3.

[0112] In the above step S2, when obtaining the curtain fabric feature cube at each exposure level, it includes:

[0113] Texture features are defined through texture nodes, and the URL field of the texture node is associated with a standard texture map;

[0114] The material characteristics are defined through the Material class node in the appearance object. The URL field of the Material class node is associated with a standard material map.

[0115] The material characteristics, texture characteristics and exposure levels defined are respectively , , The axes form an interrelated feature cube as high-level semantic features.

[0116] In the above step S3, when the adaptive image enhancement algorithm based on deep learning is used for enhancement, it includes:

[0117] The curtain fabric panoramic image is segmented by a threshold method to obtain curtain fabric sub-image blocks in bright and dark areas, and the average brightness response of the bright and dark areas is obtained;

[0118] The average brightness response of the bright and dark areas is combined with the adaptive gamma factor to obtain the curtain fabric enhancement sequence, as shown in Formula 5:

[0119] (5);

[0120] In the formula, is the intensity value of each pixel in the curtain fabric panoramic image, is the maximum pixel intensity value, is the adaptive gamma factor, Enhanced sequence for curtain fabrics;

[0121] Wherein, the first adaptive gamma factor is determined according to the pixel probability density and the weighted probability density, as shown in Formula 6 and Formula 7;

[0122] (6);

[0123] (7);

[0124] In the formula, is the average brightness response of the bright area, is the average brightness response of the dark area, is the weighted probability density, is the pixel probability density, is the first adaptive index, is the second adaptive index, To take the greater of the two, is the maximum pixel probability density, is the minimum pixel probability density, is the first adaptive gamma factor;

[0125] According to the first adaptive gamma factor, the second and third adaptive gamma factors are obtained, as shown in Formula 8:

[0126] (8);

[0127] In the formula, is the second adaptive gamma factor, is the third adaptive gamma factor;

[0128] Among them, multiple exposure levels are divided into three groups according to the exposure intensity from weak to strong, corresponding to the first, second and third adaptive gamma factors respectively.

[0129] In the above step S3, when the enhancement parameters are dynamically adjusted through the real-time feedback mechanism, it includes:

[0130] The curtain fabric enhancement sequence is input into a fully convolutional network for weight map prediction, and the weight map is upsampled through a guided filter, and the upsampling parameters are dynamically adjusted through a real-time feedback mechanism;

[0131] The real-time feedback mechanism includes:

[0132] Initialize the first parameter , the second parameter ;

[0133] Get the fitness of each individual and sort them to determine the best position ;

[0134] Update the current individual position, and update the current individual position again;

[0135] Update the first parameter , the second parameter , update the best position and return;

[0136] Among them, the first parameter , the second parameter , as shown in Formula 9 and Formula 10:

[0137] (9);

[0138] (10);

[0139] In the formula, is the control variable, is the current iteration number, is the maximum number of iterations, is the regularization parameter;

[0140] When updating the current individual position, as shown in formula 11:

[0141] (11);

[0142] In the formula, is the updated current individual position, is the distance between the current position and the optimal position, are the three components in the spatial coordinate system, Current iteration number The best position under

[0143] When updating the current individual position again, as shown in formula 12:

[0144] (12);

[0145] In the formula, Current iteration number The position with lower fitness value, is the current individual position after updating again;

[0146] Among them, the first parameter To avoid collision variables, the second parameter A random variable that balances exploration and exploitation.

[0147] In the above step S4, when filtering and local dynamic area threshold segmentation are performed, it includes:

[0148] The average grayscale value of the enhanced curtain fabric panoramic image at each exposure level is obtained, and detection is performed based on the grayscale detection condition, as shown in Formula 13:

[0149] (13);

[0150] In the formula, is the average gray value, is the low threshold in the ideal grayscale range, is the high threshold in the ideal grayscale range;

[0151] Based on the exposure level information contained in the feature cube of the curtain fabric, the enhanced curtain fabric panoramic image at the corresponding exposure level that does not meet the grayscale detection condition is eliminated;

[0152] The enhanced curtain fabric panoramic image of the remaining exposure levels is subjected to local dynamic region threshold segmentation to obtain a fabric dynamic segmentation map.

[0153] In the above step S4, when identifying fabric defects and texture defects, it includes:

[0154] Input the cloth dynamic segmentation map into Type fuzzy - The input layer of the detection model is used, and the material, texture, and exposure level features in the associated curtain fabric feature cube are used as inputs to the three neuron nodes of the input layer, as shown in Formula 14:

[0155] (14);

[0156] In the formula, the superscript symbol represents the input layer, is a neuron node, is the output of the neuron, Indicates that the input of the neuron node is the material feature, Indicates that the input of the neuron node is texture feature, Indicates that the input of the neuron node is the exposure level, is the output of the input layer;

[0157] The nodes of the membership layer receive the output of the input layer and execute Type fuzzy membership function, obtain the corresponding membership value, and output it to the output layer, as shown in Formula 15:

[0158] (15);

[0159] In the formula, is the output of the membership layer, The first Output node, symbol superscript represents the membership layer, is the input variable of the upper membership function The membership degree of is the input variable of the lower membership function The membership degree of For the Input variables, The average value of the membership function of the term, For the Input variables, The standard deviation of the upper membership function of the term, For the Input variables, The standard deviation of the lower membership function of the term;

[0160] The output layer provides a mapping from input nodes to output nodes based on the activation functions of the hidden layer and the output layer, the bias of the hidden unit, the bias of the output unit, and the weights between neurons, as shown in Formula 16:

[0161] (16);

[0162] In the formula, is the output of the output layer, and the symbol is superscript represents the output layer, is the number of neurons in the hidden layer, is the activation function of the output layer, To hide the neurons Connect to output neuron The weight of is the activation function of the hidden layer, For hidden neurons The bias on The output neuron The bias on To input the neuron Connecting to hidden neurons The weight of is the number of output nodes in the membership layer;

[0163] By using elastic back-propagation training to adjust parameters, the recognition results of cloth defects and texture defects are optimized.

[0164] In the above step S5, when performing edge optimization, it includes:

[0165] The edge optimization value is obtained based on the high-frequency amplitude of the wavelet decomposition of the edge area of ​​the panoramic image of the curtain fabric. , as shown in Formula 17:

[0166] (17);

[0167] In the formula, is the high frequency amplitude of wavelet decomposition, is a constant, is a threshold point in the ideal high-frequency threshold range, is another threshold point in the ideal high-frequency threshold range;

[0168] Edge-based optimization The edge enhancement function is generated by taking the value, the maximum value of the sub-bands to be enhanced in different directions, and the nonlinearity to enhance the edge area, as shown in Formula 18: (18);

[0169] In the formula, is the edge area after edge enhancement, for The coefficient sign is used to keep the phase unchanged. is the maximum value of the sub-band to be enhanced in different directions, , , are the first, second and third nonlinearities, used for jointly controlling the gain function;

[0170] Based on the entropy value and edge signal modulus of the smooth area and texture area of ​​the panoramic image of the enhanced curtain fabric, the smooth area and the texture area are subjected to soft threshold denoising, as shown in Formula 19 and Formula 20:

[0171] (19);

[0172] (20);

[0173] In the formula, is the entropy value of the smooth area and the texture area; is the soft threshold, which is determined based on the number of wavelet transform layers; is the edge signal modulus, These are the smooth area and texture area after denoising.

[0174] In the above step S5, when edge defects are identified, the following steps are included:

[0175] Will The network is used as the main framework for target classification and bounding box regression tasks, and adopts As the backbone network, construct the initial lightweight model 1;

[0176] Add after the feature extraction network of the last layer of the initial lightweight model 1 Attention mechanism module 2, obtains the edge detection model, such as Figure 2 As shown;

[0177] The edge-optimized panoramic image of curtain fabric is input into the edge detection model. The channel attention submodule 3 in the attention mechanism module 2 enhances important channel features, and the spatial attention submodule 4 enhances important spatial features;

[0178] Among them, when enhancing important channel features, it includes:

[0179] Perform global average pooling on the original feature map of each channel to obtain the feature description in the channel dimension, and send the feature description to the shared connection layer to obtain the weight value of each channel to generate a channel attention map;

[0180] Apply the channel attention map to the original feature map to weight the features of different channels to enhance the important channel features and obtain the intermediate feature map;

[0181] When enhancing important spatial features, include:

[0182] For each channel, perform maximum pooling and average pooling operations on the original feature map to obtain the maximum value information and average value information at each position;

[0183] After the maximum value information and the average value information are concatenated, they are sent to the convolution layer to generate a spatial attention map, and the spatial attention map is multiplied by the intermediate feature map to enhance important spatial features and suppress unimportant spatial features;

[0184] Among them, the channel attention map contains the importance information of each channel in the feature map, and the weight value contains the importance of each channel to the entire feature map;

[0185] The mechanism of the channel attention module is shown in Formula 21:

[0186] (twenty one);

[0187] In the formula, is the output after the channel attention module, is the channel attention, is element-wise multiplication, is the original feature map;

[0188] The mechanism of the spatial attention module is shown in Formula 22:

[0189] (twenty two);

[0190] In the formula, is the output after the spatial attention module, For spatial attention.

[0191] In the above step S6, when forming the interactive defect map, it includes:

[0192] Constructing ontology metadata, through The parser realizes the conversion of ontology metadata into resource description framework, and parses and stores ontology metadata in the form of triples of <subject>, <defect type>, and <defect location>, and supports query language to query the stored triples;

[0193] Associate the <subject> in the triplet with the model and batch number of the curtain fabric, and associate the <defect type> and <defect location> in the triplet with Type fuzzy - The detection model and the edge detection model are associated;

[0194] After defect identification is completed, the model and batch number of the curtain fabric, the first identification result, and the second identification result are mapped to triples through attributes, and the mapped triples are stored in An interactive defect map is formed in the graph database;

[0195] Among them, the model and batch number of the curtain fabric are mapped to <subject>, the fabric defects, texture defects and their defect positions in the first recognition result are mapped to <defect type>, <defect position> respectively, and the edge defects and their defect positions in the second recognition result are mapped to <defect type>, <defect position> respectively.

[0196] The above-mentioned embodiments are only preferred embodiments of the present invention and cannot be used to limit the scope of protection of the present invention. Any non-substantial changes and substitutions made by technicians in this field on the basis of the present invention shall fall within the scope of protection required by the present invention.

Claims

1. A curtain fabric defect recognition method based on feature cube and fuzzy detection dual recognition, characterized in that: The following steps are involved: Capturing curtain fabric images at different angles at multiple exposure levels, and performing multi-angle synthesis for each exposure level to obtain a panoramic image of the curtain fabric at each exposure level; Obtain material features and texture features according to the model of the curtain fabric to be identified, and correlate them with the exposure level to obtain a feature cube of the curtain fabric at each exposure level; After the panoramic image is grayed out, it is enhanced using an adaptive image enhancement algorithm based on deep learning, and the enhancement parameters are dynamically adjusted through a real-time feedback mechanism to obtain an enhanced panoramic image of the curtain fabric at each exposure level. Based on the feature cube of curtain fabric, the enhanced image is filtered and the local dynamic area threshold segmentation is performed to obtain the dynamic segmentation map of the fabric, which is input into Type fuzzy - The input layer of the detection model is used, and the material, texture, and exposure level features in the associated curtain fabric feature cube are used as inputs of three neuron nodes of the input layer to identify fabric defects and texture defects, thereby obtaining a first recognition result. The enhanced image is optimized at the edge by using an edge optimization algorithm, and is input into an edge detection model to identify edge defects, thereby obtaining a second identification result; The first and second recognition results are attribute-mapped to form an interactive defect map for displaying the location and type of each defect.

2. The curtain fabric defect recognition method based on feature cube and fuzzy detection dual recognition according to claim 1 is characterized in that: When performing multi-angle synthesis, it includes: The phase correlation method is used to obtain the displacement between two adjacent angle images at the same exposure level and determine the overlapping area; Extract in the overlapped area Features, and use similarity measurement method to perform rough matching of feature points; According to the plane projection model, The method estimates the perspective transformation matrix between the images to be stitched, and optimizes the matching point pairs to complete the transformation alignment; The transformed image is decomposed into multi-scale space through the curtain fabric image pyramid, and weighted fusion is performed using the fade-in and fade-out weighted method according to the horizontal displacement of the image.

3. The curtain fabric defect recognition method based on feature cube and fuzzy detection dual recognition according to claim 2 is characterized in that: When optimizing matching point pairs, include: Using the first curtain fabric image collected as a reference image, obtaining a set of errors of matching point pairs after transformation relative to the reference image; The errors after transformation of all groups of matching point pairs are summed based on the error function to obtain the total error value; By minimizing the total error value, the homography matrix of an angle image to the reference image is adjusted to optimize the matching point pairs.

4. The curtain fabric defect recognition method based on feature cube and fuzzy detection dual recognition according to claim 3 is characterized in that: When obtaining the feature cube of curtain fabric, it includes: The texture features are defined by a texture node, wherein the URL field of the texture node is associated with a standard texture map; The material characteristics are defined through the Material class node in the appearance object, and the URL field of the Material class node is associated with a standard material map; The material characteristics, texture characteristics and exposure levels defined are respectively used as , , The axes form an interrelated feature cube as high-level semantic features.

5. The curtain fabric defect recognition method based on feature cube and fuzzy detection dual recognition according to any one of claims 1 to 4, characterized in that: When making enhancements, include: Segmenting the curtain fabric panoramic image by a threshold method to obtain curtain fabric sub-image blocks in bright and dark areas, and obtaining average brightness responses in the bright and dark areas; The average brightness response of the bright and dark areas is combined with the adaptive gamma factor to obtain the curtain fabric enhancement sequence, as shown in Formula 5: （5）； In the formula, is the intensity value of each pixel in the curtain fabric panoramic image, is the maximum pixel intensity value, is the adaptive gamma factor, Enhanced sequence for curtain fabrics; Wherein, the first adaptive gamma factor is determined according to the pixel probability density and the weighted probability density, as shown in Formula 6 and Formula 7; （6）； （7）； In the formula, , is the average brightness response of bright and dark areas, is the weighted probability density, is the pixel probability density, , are the first and second adaptive indexes, To take the greater of the two, , are the maximum and minimum pixel probability densities, is the first adaptive gamma factor; According to the first adaptive gamma factor, the second and third adaptive gamma factors are obtained, as shown in Formula 8: （8）； In the formula, , are the second and third adaptive gamma factors.

6. The curtain fabric defect recognition method based on feature cube and fuzzy detection dual recognition according to claim 5 is characterized in that: When dynamically adjusting enhancement parameters, including: Input the curtain fabric enhancement sequence into a fully convolutional network for weight map prediction, upsample the weight map through a guided filter, and dynamically adjust the upsampling parameters through a real-time feedback mechanism; Wherein, the real-time feedback mechanism includes: Initialize the first parameter , the second parameter ; Get the fitness of each individual and sort them to determine the best position ; Update the current individual position, and update the current individual position again; Update the first parameter , the second parameter , update the best position and return; Among them, the first parameter To avoid collision variables, the second parameter A random variable that balances exploration and exploitation.

7. The curtain fabric defect recognition method based on feature cube and fuzzy detection dual recognition according to claim 6 is characterized in that: When filtering and local dynamic area threshold segmentation, it includes: Obtaining an average grayscale value of the enhanced curtain fabric panoramic image at each exposure level, and performing detection based on a grayscale detection condition; Based on the exposure level information contained in the curtain fabric feature cube, the enhanced curtain fabric panoramic image at the corresponding exposure level that does not meet the grayscale detection condition is eliminated; The enhanced curtain fabric panoramic image of the remaining exposure levels is subjected to local dynamic region threshold segmentation to obtain a fabric dynamic segmentation map.

8. The curtain fabric defect recognition method based on feature cube and fuzzy detection dual recognition according to claim 7 is characterized in that: When identifying fabric defects and texture defects, it includes: The output of the input layer is received by the nodes of the membership layer, and the Type fuzzy membership function, obtain the corresponding membership value, and output it to the output layer, as shown in Formula 15: （15）； In the formula, is the output of the membership layer, The first Output nodes, symbols superscript represents the membership layer, , is the input variable of the upper and lower membership functions The membership degree of For the Input variables, The average value of the membership function of the term, , For the Input variables, The standard deviation of the upper and lower membership functions of the term; The output layer provides a mapping from input nodes to output nodes, and uses elastic back-propagation training to adjust parameters to optimize the recognition results of cloth defects and texture defects.

9. The curtain fabric defect recognition method based on feature cube and fuzzy detection dual recognition according to claim 1 or 6, characterized in that: When optimizing at the edge, include: The edge optimization value is obtained based on the high-frequency amplitude of the wavelet decomposition of the edge area of ​​the enhanced curtain fabric panoramic image. ; Based on the edge optimization value , the maximum value and nonlinearity of the sub-bands to be enhanced in different directions, and generate an edge enhancement function to enhance the edge area; Based on the entropy values ​​and edge signal modulus values ​​of the smooth area and the texture area of ​​the enhanced curtain fabric panoramic image, soft threshold denoising is performed on the smooth area and the texture area.

10. The curtain fabric defect recognition method based on feature cube and fuzzy detection dual recognition according to claim 9, characterized in that: When identifying edge defects, include: Will The network is used as a framework for target classification and bounding box regression tasks, and adopts As the backbone network, build the initial lightweight model; Add after the feature extraction network of the last layer of the initial lightweight model An attention mechanism module is used to obtain the edge detection model; The edge-optimized curtain fabric panoramic image is input into the edge detection model. The channel attention submodule in the attention mechanism module generates a channel attention map to enhance important channel features, and the spatial attention submodule generates a spatial attention map to enhance important spatial features.

Images