Cord fabric production defect identification and analysis system based on machine vision

Abstract

The invention belongs to the technical field of image processing, and particularly relates to a cord fabric production defect identification and analysis system based on machine vision, which comprises a multispectral imaging module, an image preprocessing module, a target detection module, a result verification module, a weight correction module and a control response module, the method comprises the following steps: acquiring a multispectral image sequence of each partition of a cord fabric forming surface, combining motion compensation, geometric calibration and noise filtering to improve preprocessing precision, and generating a composite spectral image by adopting adaptive weight fusion on the basis of considering an image signal-to-noise ratio and texture definition; and detecting defects based on frequency division domain texture and morphological gradient parameter bimodal matching, further verifying detection credibility through defect sensitive spectral image feature enhancement degree, dynamically correcting a fusion weight and performing iterative optimization, and finally triggering a production control instruction according to a detection result. And closed-loop management of high-precision defect detection and production control linkage of the cord fabric is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and specifically relates to a defect recognition and analysis system for cord fabric production based on machine vision. Background Art

[0002] As the core skeleton material of tires, the quality of cord fabric directly affects the strength, durability and safety performance of tires. With the global tire industry upgrading towards lightweight and high reliability, defect detection in cord fabric production has become a key technology for improving quality.

[0003] Traditional manual inspection has limitations such as low efficiency, strong subjectivity and high omission rate, and it is difficult to meet the requirements of modern industrial automation. The defect recognition and analysis system based on machine vision provides an efficient and accurate intelligent quality control solution for cord fabric production by integrating high-precision image acquisition, deep learning algorithms and real-time data processing technologies.

[0004] In the prior art, there are already a large number of defect recognition solutions based on machine vision that can be directly applied to the cord fabric production process. For example, a machine vision-based textile detection method and system with the Chinese patent publication number CN117115147B, which is based on textile images, uses a convolutional neural network to extract features to generate a preliminary recognition library, and combines technologies such as recurrent neural networks, multi-spectral image acquisition, computer vision, and blockchain to form a detection system, which can accurately detect defects and hidden defects, realize real-time quality monitoring and adjustment, and build a reliable traceability system with the help of blockchain and the Internet of Things.

[0005] Another machine vision-based textile surface defect detection method with the Chinese patent publication number CN115294097A, which obtains the gray information of textile images, determines the window shape and size according to the template image, performs denoising by determining the adaptive weighted denoising mean based on gray and pixel distance information, and calculates the denoised data to accurately detect textile surface defects.

[0006] As can be seen from the above solutions, the prior art can generate a fused image through multi-spectral image acquisition and denoising processing to support defect detection, but there are still significant technical bottlenecks in the dynamic production scenario of cord fabric. Firstly, during the high-speed continuous production process of cord fabric, the operation of the loom and the movement of the conveyor belt cause multi-source interference coupling in image acquisition, including spatial deformation caused by motion blur, registration deviation between multi-spectral imaging modules due to mechanical vibration or timing asynchrony, and the superimposed pollution of loom vibration noise. The image preprocessing in the prior art lacks the ability of dynamic perception and optimization, and it is difficult to achieve the spatio-temporal synchronization repair of multi-modal data under complex working conditions, resulting in structural distortion and signal-to-noise ratio deterioration of the fused image, seriously affecting the baseline data quality of defect features.

[0007] Second, the existing technology lacks a reliability verification mechanism for multi-spectral fusion results. Its fusion weight allocation depends on static parameters calibrated based on historical experience, which easily leads to the significant expression of key defects in specific spectral dimensions being suppressed by unbalanced fusion weights, resulting in the attenuation of feature information of micro-defects. In addition, the lack of multi-spectral reverse verification and confidence evaluation of the detection results after fusion increases the misjudgment rate of pseudo-defects under the complex texture background of cord fabric, forming a systematic contradiction between detection accuracy and generalization. Summary of the Invention

[0008] To overcome the deficiencies in the background art, embodiments of the present invention provide a cord fabric production defect recognition and analysis system based on machine vision, which can effectively solve the problems involved in the above background art.

[0009] The object of the present invention can be achieved by the following technical solutions: A cord fabric production defect recognition and analysis system based on machine vision includes: a multi-spectral imaging module, an image preprocessing module, a target detection module, a result verification module, a weight correction module, and a control response module.

[0010] The multi-spectral imaging module is connected to the image preprocessing module, the image preprocessing module is connected to the target detection module, the target detection module is connected to the result verification module, the result verification module is respectively connected to the weight correction module and the control response module, and the weight correction module is connected to the target detection module.

[0011] The multi-spectral imaging module is composed of an industrial camera linear array integrating a dichroic mirror and an adaptive supplementary light source. Each camera in the array covers a rectangular monitoring area on the forming surface of the cord fabric at the loom exit end according to a preset field of view mapping relationship and synchronously acquires a multi-spectral image sequence of its covered area.

[0012] The image preprocessing module performs preprocessing operations including motion compensation registration and noise filtering on the multi-spectral image sequences of each area.

[0013] The target detection module performs weighted fusion on the preprocessed multi-spectral image sequences of each area to generate a composite spectral image, and based on the multi-scale texture features of the composite spectral image, outputs a target detection result including the number of defects, defect types, and position coordinates of each area.

[0014] The result verification module associates the mapping relationship between the defect type and the corresponding sensitive spectral channel, and verifies the credibility of the target detection result of the composite spectral image according to the feature enhancement degree of the target defect type in each area in the image of the sensitive spectral channel. If the verification is credible, it executes the control response module, otherwise it executes the weight correction module.

[0015] The weight correction module corrects the fusion weights of the multi-spectral image sequence according to the feature enhancement intensity deviation and feeds them back to the target detection module to regenerate the composite spectral image.

[0016] The control response module sends the target detection result to the production control terminal, and the terminal responds to the corresponding control instruction.

[0017] Compared with the prior art, the embodiments of the present invention have at least the following advantages or beneficial effects: (1) By collecting the multi-spectral image sequences of each partition on the forming surface of the cord fabric, the present invention combines motion compensation, geometric calibration and noise filtering to improve the preprocessing accuracy, generates a composite spectral image and performs defect target detection, verifies the credibility of the detection based on the feature enhancement intensity of the sensitive spectral image where the defect type is located in the target detection result, dynamically executes fusion weight correction, and finally triggers a production control instruction according to the detection result, realizing the closed-loop management of high-precision defect detection and production control linkage.

[0018] (2) The preprocessing module of the present invention uses motion compensation, sub-pixel geometric calibration and cross-channel noise filtering to repair the multi-source interference caused by loom vibration and motion blur in the multi-spectral image, ensuring the spatio-temporal alignment inside the multi-spectral image sequence, thereby effectively avoiding the structural distortion and signal-to-noise ratio deterioration of the subsequent composite spectral image.

[0019] (3) The present invention detects defects based on the dual-modal feature matching of the frequency-domain texture and morphological gradient parameters of the composite spectral image, covering the frequency-domain characteristics and spatial structure characteristics of the defects, thereby effectively avoiding the limitations of a single feature mode, significantly enhancing the recognition ability of micro-defects under complex texture backgrounds, and reducing the missed detection rate.

[0020] (4) The present invention establishes a defect type-sensitive spectral channel mapping relationship, realizes the reverse verification of the credibility of the composite spectral image detection result based on the feature enhancement intensity of the sensitive spectral channel image where the defect type is located. If the verification is not credible, a positive or negative adjustment factor is further constructed based on the deviation ratio to correct the weight for closed-loop iterative optimization, suppressing the interference of non-sensitive channels, reducing the misjudgment of pseudo-defects, and improving the detection generalization. Description of the Drawings

[0021] The present invention is further described with reference to the accompanying drawings. However, the embodiments in the drawings do not constitute any limitation to the present invention. For those of ordinary skill in the art, other drawings can also be obtained according to the following drawings without creative efforts.

[0022] Figure 1 It is a schematic diagram of the module connection of the present invention.

[0023] Figure 2 It is a schematic diagram of the structure of the cord fabric multi-spectral dynamic imaging array of the present invention.

[0024] Figure 3 It is a schematic diagram of the logic for obtaining the object detection result in the object detection module of the present invention.

[0025] Reference numerals: 1. Industrial camera; 2. Adaptive supplementary light source; 3. Loom; 4. Cord fabric. Specific embodiments

[0026] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0027] Refer to Figure 1 As shown, the present invention provides a cord fabric production defect identification and analysis system based on machine vision, including: a multispectral imaging module, an image preprocessing module, an object detection module, a result verification module, a weight correction module, and a control response module.

[0028] The multispectral imaging module is connected to the image preprocessing module, the image preprocessing module is connected to the object detection module, the object detection module is connected to the result verification module, the result verification module is respectively connected to the weight correction module and the control response module, and the weight correction module is connected to the object detection module.

[0029] Refer to Figure 2 As shown, the multispectral imaging module is composed of a linear array of industrial cameras integrated with dichroic mirrors and adaptive supplementary light sources. Each camera in the array covers a rectangular monitoring area of the cord fabric forming surface at the loom outlet end according to a preset field of view mapping relationship and synchronously acquires a multispectral image sequence of its covered area.

[0030] The image preprocessing module performs preprocessing operations including motion compensation registration and noise filtering on the multispectral image sequences of each area.

[0031] In a preferred embodiment of the present invention, the preprocessing operation refers to the following execution process: (a) Based on the fixed conveying speed of the cord fabric and the camera exposure response duration, determine the displacement amount of the cord fabric at the camera sampling moment, and perform reverse translation compensation on the multispectral image sequences of each area to eliminate motion blur.

[0032] It should be added that the displacement amount of the cord fabric at the camera sampling moment is the product of the fixed conveying speed of the cord fabric and the camera exposure response duration.

[0033] (b) Perform intra-group geometric calibration on the multi-spectral image sequence of the same partition, including: establishing an affine transformation matrix between spectral channels based on the optical characteristics of the dichroic mirror, performing global geometric correction on the images of each spectral channel in the same partition, and converting the corrected images of each spectral channel to the frequency domain. Obtain the sub-pixel translation amount between each spectral channel through the phase correlation method, and perform sub-pixel translation compensation on the images of each spectral channel using an interpolation algorithm.

[0034] It should be added that due to the spectral splitting characteristics of the dichroic mirror, there will be optical path differences in different spectral channels, which are specifically manifested as translation, rotation, or scaling between images. Based on the pre-set optical design parameters of the dichroic mirror, establish an affine transformation matrix between spectral channels and apply it to each spectral channel image one by one to correct the geometric distortion caused by the optical path difference. The affine transformation matrix includes a scaling factor to eliminate the difference in optical path magnification, a shear factor to eliminate the tilt of the optical axis, and a translation amount to eliminate the optical path offset.

[0035] The above-mentioned obtaining the sub-pixel translation amount between each spectral channel through the phase correlation method specifically refers to: after the images of each spectral channel are converted to the frequency domain, obtain the spectral information through Fourier transform, calculate the cross-power spectrum between the images of two spectral channels based on the spectral information, perform inverse Fourier transform on the cross-power spectrum to obtain the impulse response function, retrieve the peak position of the impulse response function and perform Gaussian fitting synchronously to quantify the sub-pixel translation amount between the images of two spectral channels, and thus the sub-pixel translation amount between each spectral channel can be obtained.

[0036] The above-mentioned interpolation algorithm specifically refers to the bicubic interpolation algorithm.

[0037] (c) Extract the gray-scale distribution of the background region of the images of each spectral channel in the same partition, analyze the noise distribution characteristics of the multi-spectral channels and fit the noise probability density function, set a dynamic threshold to distinguish the real features of the images from abnormal noise points, and then filter out the isolated noise points caused by the mechanical vibration of the loom based on the cross-channel signal correlation.

[0038] The preprocessing module of the present invention uses motion compensation, sub-pixel geometric calibration, and cross-channel noise filtering to repair the multi-source interference caused by loom vibration and motion blur in the multi-spectral images, ensure the spatio-temporal alignment inside the multi-spectral image sequence, and thus effectively avoid the structural distortion and signal-to-noise ratio degradation of the subsequent composite spectral images.

[0039] The target detection module performs weighted fusion on the preprocessed multi-spectral image sequences of each partition to generate a composite spectral image, and based on the multi-scale texture features of the composite spectral image, outputs the target detection results including the number of defects, defect types, and position coordinates of each partition.

[0040] In a preferred embodiment of the present invention, the weighted fusion processing of the multi-spectral image sequence adopts an adaptive weight allocation method. The corresponding fusion weight coefficients of each spectral channel image take its predefined reference weight as the initial parameter, and are dynamically adjusted in combination with the image signal-to-noise ratio and texture clarity obtained in real time, wherein the predefined reference weight is determined through an offline calibration experiment based on the intrinsic spectral reflection characteristics of the cord fabric material.

[0041] It should be added that the specific acquisition process of the above image signal-to-noise ratio and texture clarity is as follows: Select a preset pixel size window in the texture area of the spectral channel image and calculate the gray mean value of this window. Select the textureless background area of the spectral channel image and calculate the gray standard deviation of this area. Substitute the gray mean value and the gray standard deviation into the standard SNR formula to obtain the image signal-to-noise ratio of the spectral channel image.

[0042] Apply the Sobel operator to each pixel of the spectral channel image to obtain the horizontal gradient and vertical gradient of each pixel and calculate the gradient amplitude therefrom. Define the texture clarity as the sum of the squares of the gradient amplitudes of all pixels in the spectral channel image.

[0043] The specific acquisition process of the corresponding fusion weight coefficients of the above each spectral channel image is as follows: Normalize the image signal-to-noise ratio and texture clarity of each spectral channel image so that the values are in the interval. Set the corresponding adjustment factors for the image signal-to-noise ratio and texture clarity to control the contribution ratio of the image signal-to-noise ratio and texture clarity. Multiply the image signal-to-noise ratio by its adjustment factor, add it to the product of the texture clarity and its adjustment factor, and use the product of the calculation result and the predefined reference weight of the spectral channel image as the corresponding fusion weight coefficient of the spectral channel image. Finally, it is still necessary to normalize the corresponding fusion weight coefficients of each spectral channel image. The setting of the corresponding adjustment factors for the image signal-to-noise ratio and texture clarity is based on the weighted fusion focus target of the multi-spectral image sequence. If noise suppression is emphasized, the adjustment factor given to the image signal-to-noise ratio is greater than the adjustment factor of the texture clarity, which can be exemplified as 0.7 and 0.3. If detail retention is emphasized, the adjustment factor given to the image signal-to-noise ratio is less than the adjustment factor of the texture clarity, which can be exemplified as 0.3 and 0.7. It is constantly required that the sum of the adjustment factors of the image signal-to-noise ratio and texture clarity is 1.

[0044] Refer to Figure 3 As shown, in a preferred embodiment of the present invention, the target detection result is obtained through the following process: S1. Divide the composite spectral image into several image units, extract the frequency-domain texture parameter set and morphological gradient parameter set of each image unit, and respectively organize them into a texture parameter matrix and a morphological parameter matrix.

[0045] It should be added that the above frequency-domain texture parameter set at least includes high-frequency subband energy, low-frequency subband energy, subband entropy value, mean direction response, and variance of direction response. The morphological gradient features at least include the multi-scale gradient mean of each unit of the image, gradient extreme density, direction gradient dispersion, and meridional / zonal gradient concentration.

[0046] S2. Perform similarity matching between the texture parameter matrix of each image unit and the standard texture parameter interval matrix of each defect type of the pre-stored cord fabric to generate a texture similarity score, and simultaneously perform similarity matching between the morphological parameter matrix and the standard morphological parameter interval matrix to generate a morphological similarity score.

[0047] S3. Calculate the product value of the bimodal similarity scores of each image unit and each defect type. If the product value of a certain image unit and a specific defect type exceeds a preset value, mark the image unit with the corresponding defect type label.

[0048] S4. Perform connected component retrieval on adjacent image units with the same defect type label, merge them to form the same defect region, and use the combination of the vertex coordinates of the minimum bounding rectangle and the centroid coordinates of the defect region as the defect position coordinates, and output the number of defects, defect types, and position coordinates of the corresponding partition of the composite spectral image accordingly.

[0049] In a preferred embodiment of the present invention, the similarity matching in step S3 is as follows: assign a preset reference matching score to each element in the parameter matrix.

[0050] Compare the value of each element in the parameter matrix with the numerical interval of the element at the same position in the corresponding standard parameter interval matrix. If the element value is within the corresponding numerical interval, retain the reference matching score of the element. If it exceeds the numerical interval, calculate a score correction coefficient based on the deviation degree of the element value from the interval boundary, and reduce the reference matching score according to the correction coefficient.

[0051] It should be noted that the process of calculating the score correction coefficient based on the deviation degree of the element value from the interval boundary is as follows: if the element value is less than the lower limit value of the interval, use the ratio of the absolute difference between the element value and the lower limit value of the interval to the lower limit value of the interval as the deviation degree. If the element value is greater than the lower limit value of the interval, use the ratio of the difference between the element value and the upper limit value of the interval to the upper limit value of the interval as the deviation degree, and substitute the deviation degree into the standard exponential decay function with the control decay factor set to 1 to obtain the score correction coefficient.

[0052] According to the predefined defect type weight rule, obtain the weight coefficient of each defect type of the cord fabric for each element in the parameter matrix, and perform weighted summation on the matching scores of each element in the parameter matrix according to the weight to generate the similarity score.

[0053] Among them, the parameter matrix includes a texture parameter matrix or a morphological parameter matrix, and the standard parameter interval matrix includes a corresponding standard texture parameter interval matrix or a standard morphological parameter interval matrix.

[0054] The embodiment of the present invention detects defects based on the dual-modal feature matching of the frequency-domain texture and morphological gradient parameters of the composite spectral image, covering the frequency-domain characteristics and spatial structure characteristics of the defects, thereby effectively avoiding the limitations of a single feature mode and significantly enhancing the recognition ability of micro-defects under a complex texture background to reduce the missed detection rate.

[0055] The result verification module associates the mapping relationship between the defect type and the corresponding sensitive spectral channel, and verifies the credibility of the target detection result of the composite spectral image according to the feature enhancement degree of each partition target defect type in the sensitive spectral channel image. If the verification is credible, the control response module is executed; otherwise, the weight correction module is executed.

[0056] In a preferred embodiment of the present invention, the feature enhancement degree of each partition target defect type in the sensitive spectral channel image is obtained as follows: Based on the weight coefficients of each element in the parameter matrix for each defect type of the cord fabric, the weight coefficients of the target defect type for each element in the parameter matrix are extracted, and the elements whose weight coefficients exceed the average element weight coefficient of the parameter matrix are selected as the enhanced elements of the target defect type in the parameter matrix, and thus each enhanced element of each partition target defect type in the texture parameter matrix and the morphological parameter matrix is retrieved.

[0057] It should be added that the enhanced elements of the above target defect type in the texture parameter matrix and the morphological parameter matrix can be exemplarily referred to Table 1.

[0058] Table 1 Mapping relationship table between enhanced elements and defect types

[0059]

[0060] Obtain the texture parameter matrix and the morphological parameter matrix of the target defect position area in each spectral channel image of each partition, retrieve the numerical values of each enhanced element in the matrix and calculate the defect feature manifestation degree in the enhanced direction, and take the average value to obtain the feature manifestation degree of the target defect position area in each spectral channel image of each partition.

[0061] The above retrieval of the numerical values of each enhanced element in the matrix and the calculation of the defect feature manifestation degree in the enhanced direction can be explained by taking the high-frequency sub-band energy of the above strong twist defect in the texture parameter matrix as an example: When a strong twist defect appears in the cord fabric, its spiral texture will cause an increase in high-frequency energy, so the numerical enhancement direction should be a positive increase. The relative increase ratio of the high-frequency sub-band energy value in the matrix to the preset high-frequency sub-band energy standard value in the normal state of the cord fabric is used as the defect feature manifestation degree in the corresponding enhanced direction.

[0062] Taking the low-frequency sub-band energy of the hole defect in the above table in the texture parameter matrix as an example for explanation: when there is a hole defect in the cord fabric, the hole causes a decrease in the low-frequency energy of the light-transmitting area, so the numerical enhancement direction should be a reverse decrease. The relative reduction ratio of the low-frequency sub-band energy value in the matrix to the preset low-frequency sub-band energy standard value under the normal state of the cord fabric is used as the defect feature manifestation degree in the corresponding enhancement direction.

[0063] Among them, the preset high- and low-frequency sub-band energy standard values under the normal state of the cord fabric can be based on the detection requirements of the physical properties of the cord fabric in the industry standard specifications, or can be calibrated based on the experimental tests of defect-free cord fabric samples at the initial stage of system development.

[0064] Calculate the minimum improvement rate of the feature manifestation degree of the target defect type in each partition in the sensitive spectral channel image relative to the feature manifestation degree in other spectral channel images, and multiply the improvement rate by the feature manifestation degree in the sensitive spectral channel image to obtain the feature enhancement degree of the target defect type in each partition in the sensitive spectral channel image.

[0065] In a preferred embodiment of the present invention, the condition for verifying the credibility of the target detection result of the composite spectral image is that the feature enhancement degree of the target defect type in each partition in the sensitive spectral channel image is greater than or equal to the preset feature enhancement degree standard threshold.

[0066] The weight correction module corrects the fusion weights of the multi-spectral image sequence according to the feature enhancement degree deviation and feeds them back to the target detection module to regenerate the composite spectral image.

[0067] In a preferred embodiment of the present invention, the correction of the fusion weights of the multi-spectral image sequence according to the feature enhancement degree deviation includes: extracting the feature enhancement degree of the target defect type in the untrusted partition in the sensitive spectral channel image, constructing a positive adjustment factor with the deviation ratio of the feature enhancement degree to the preset feature enhancement degree standard threshold, and upwardly correcting the corresponding fusion weight coefficient of the sensitive spectral channel image.

[0068] It should be noted that the above positive adjustment factor can be exemplarily constructed as a linear adjustment factor, that is, the product combination of the normalized deviation ratio and the preset gain coefficient. Upwardly correcting the corresponding fusion weight coefficient of the sensitive spectral channel image means multiplying the original fusion weight coefficient by the positive adjustment factor as the upward correction amount, which is increased on the basis of the original fusion weight coefficient.

[0069] Construct a reverse adjustment factor with the deviation ratio of the feature manifestation degree of the target defect type in the untrusted partition in the sensitive spectral channel image relative to the maximum feature manifestation degree in other spectral channel images, and downwardly correct the corresponding fusion weight coefficients of other spectral channel images.

[0070] It should be noted that the above reverse adjustment factor can be exemplarily constructed as an exponential decay factor, that is, substituting the minimum deviation ratio into the exponential decay function under the preset suppression intensity coefficient, and downwardly correcting the corresponding fusion weight coefficients of each other spectral channel image. That is, multiplying the original fusion weight coefficient by the reverse adjustment factor as the downward correction amount, and reducing it on the basis of the original fusion weight coefficient.

[0071] Normalize the corresponding fusion weight coefficients of each spectral channel image after correction to regenerate the fusion weights of the multi-spectral image sequence.

[0072] In a preferred embodiment of the present invention, the weight correction module further includes an iterative verification and rollback mechanism: set the maximum number of iterations for correcting the fusion weights of the multi-spectral image sequence. If it is still verified as untrustworthy after reaching the maximum number of iterations, automatically roll back the predefined reference weight configuration of each spectral channel image, trigger an alarm signal and generate an error log. The error log includes the weight adjustment trajectory during the iteration process and the change curve of the credibility verification data.

[0073] In the embodiment of the present invention, by establishing the defect type-sensitive spectral channel mapping relationship, the reverse verification of the credibility of the composite spectral image detection result is realized based on the feature enhancement degree of the sensitive spectral channel image where the defect type is located. If the verification is untrustworthy, further construct a positive or reverse adjustment factor based on the deviation ratio to correct the weight for closed-loop iterative optimization, suppress the interference of non-sensitive channels, reduce the misjudgment of pseudo-defects, and improve the detection generalization.

[0074] The control response module sends the target detection result to the production control terminal, and the terminal responds to the corresponding control instruction.

[0075] In a preferred embodiment of the present invention, the control instruction includes the following operations: when no defect is detected in the target detection result, send a maintenance operation instruction to the production control terminal to maintain the current production state.

[0076] When there are the number of partition defects, defect types and position coordinates in the target detection result, send a shutdown instruction to the production control terminal to abort the cord fabric production process, and convert the defect position coordinates in the image into the actual defect position coordinates on the cord fabric forming surface according to the mapping relationship between the composite spectral image and the actual physical coordinates of the cord fabric, generate a notification window including the actual defect position coordinates, defect types and quantities, and feedback it to the operator terminal to trigger manual intervention.

[0077] It should be added that the preset parameters or rules of the system of the present invention are all stored in the cloud database and support direct calling during use. The data sources of all preset parameters are preset during the system development stage to ensure the standardization and reliability of parameter calling.

[0078] In the embodiment of the present invention, by collecting multi-spectral image sequences of each partition on the forming surface of the cord fabric, combining motion compensation, geometric calibration, and noise filtering to improve the preprocessing accuracy, generating a composite spectral image and performing defect target detection, verifying the credibility of the detection based on the feature intensity of the sensitive spectral image where the defect type is located in the target detection result, dynamically executing fusion weight correction, and finally triggering a production control instruction according to the detection result, a closed-loop management of high-precision defect detection and production control linkage is realized.

[0079] The above content is only an example and illustration of the structure of the present invention. Those skilled in the art of the present technology can make various modifications or supplements to the described specific embodiments or use similar methods for substitution, as long as they do not deviate from the structure of the invention or exceed the scope defined by the present invention, they should fall within the protection scope of the present invention.

Claims

1. A machine vision-based defect recognition and analysis system for cord fabric production, characterized in that: include: The multispectral imaging module is composed of a linear array of industrial cameras with integrated dichroic mirrors and adaptive fill light sources. Each camera in the array covers the rectangular monitoring partition of the cord fabric forming surface at the outlet of the loom according to the preset field of view mapping relationship and synchronously collects the multispectral image sequence of the covered partition; An image preprocessing module performs motion compensation registration and noise filtering preprocessing operations on multispectral image sequences; The target detection module generates a composite spectral image by weighted fusion of the preprocessed multispectral image sequence, and outputs the target detection results of the number of defects, defect types and position coordinates of each partition based on the texture features of the composite spectral image; The result verification module associates the mapping relationship between the defect type and the corresponding sensitive spectral channel, verifies the credibility of the composite spectral image target detection result according to the feature enhancement degree of the target defect type in the sensitive spectral channel image, and executes the control response module if it is credible, otherwise executes the weight correction module; The weight correction module corrects the fusion weight of the multi-spectral image sequence according to the feature enhancement deviation and regenerates the composite spectral image; The control response module sends the target detection results to the control terminal to execute the control instructions.

2. The machine vision-based cord fabric production defect recognition and analysis system according to claim 1, characterized in that: The preprocessing operation is as follows: (a) based on the fixed transmission speed of the cord fabric and the exposure response time of the camera, the displacement of the cord fabric at the camera sampling moment is determined, and the multispectral image sequence of each partition is reversely translated to eliminate motion blur; (b) performing intra-group geometric calibration on the multispectral image sequence of the same partition, including: establishing an affine transformation matrix between spectral channels based on the optical characteristics of the dichroic mirror, performing global geometric correction on the images of each spectral channel in the same partition, and converting the corrected images of each spectral channel to the frequency domain, obtaining the sub-pixel translation between each spectral channel by a phase correlation method, and performing sub-pixel translation compensation on each spectral channel image by an interpolation algorithm; (c) The grayscale distribution of the background area of ​​each spectral channel image in the same partition is extracted, the noise distribution characteristics of the multi-spectral channel are analyzed and the noise probability density function is fitted. The dynamic threshold is set to distinguish the real features of the image from abnormal noise points, and then the isolated noise points caused by the mechanical vibration of the loom are filtered out based on the cross-channel signal correlation.

3. The machine vision-based cord fabric production defect recognition and analysis system according to claim 1, characterized in that: The weighted fusion processing of the multispectral image sequence adopts an adaptive weight allocation method. The fusion weight coefficient corresponding to each spectral channel image takes its predefined benchmark weight as an initial parameter, and is dynamically adjusted in combination with the image signal-to-noise ratio and texture clarity acquired in real time. The predefined benchmark weight is determined by an offline calibration experiment based on the intrinsic spectral reflectance characteristics of the cord fabric material.

4. The machine vision-based cord fabric production defect recognition and analysis system according to claim 1, characterized in that: The target detection result is obtained by referring to the following acquisition process: S1. Dividing the composite spectral image into a plurality of image units, extracting a frequency domain texture parameter set and a morphological gradient parameter set of each image unit, and organizing them into a texture parameter matrix and a morphological parameter matrix respectively; S2. The texture parameter matrix of each image unit is matched with the pre-stored standard texture parameter interval matrix of each defect type of the cord fabric to generate a texture similarity score, and the morphological parameter matrix is ​​matched with the standard morphological parameter interval matrix to generate a morphological similarity score; S3. Calculate the product value of the bimodal similarity score of each image unit and each defect type. If the product value of an image unit and a specific defect type exceeds a preset value, mark the image unit as a corresponding defect type label; S4. Connected domain retrieval is performed on adjacent image units with the same defect type label, and they are merged to form the same defect area. The combination of the vertex coordinates and the centroid coordinates of the minimum circumscribed rectangle of the defect area is used as the defect position coordinates, thereby outputting the defect quantity, defect type and position coordinates of the corresponding partition of the composite spectral image.

5. The machine vision-based cord fabric production defect recognition and analysis system according to claim 4 is characterized by: The similarity matching in step S3 refers to the following process: Assigning a preset benchmark matching score to each element in the parameter matrix; Compare the value of each element in the parameter matrix with the value interval of the element at the same position in the corresponding standard parameter interval matrix. If the element value is within the corresponding value interval, the benchmark matching score of the element is retained. If it exceeds the value interval, the score correction coefficient is calculated based on the deviation between the element value and the interval boundary, and the benchmark matching score is reduced according to the correction coefficient. According to the predefined defect type weight rule, the weight coefficient of each defect type of the cord fabric for each element in the parameter matrix is ​​obtained, and the matching score of each element in the parameter matrix is ​​weighted and summed according to the weight to generate the similarity score; The parameter matrix includes a texture parameter matrix or a morphology parameter matrix, and the standard parameter interval matrix includes a corresponding standard texture parameter interval matrix or a standard morphology parameter interval matrix.

6. The machine vision-based cord fabric production defect recognition and analysis system according to claim 5, characterized in that: The characteristic enhancement degree of each subarea target defect type in the sensitive spectral channel image refers to the following acquisition process: Based on the weight coefficients of each defect type of the cord fabric to each element in the parameter matrix, the weight coefficients of the target defect type to each element in the parameter matrix are extracted, and the elements whose weight coefficients exceed the average element weight coefficient of the parameter matrix are selected as the strengthening elements of the target defect type in the parameter matrix, so as to retrieve the strengthening elements of the target defect type in each partition in the texture parameter matrix and the morphology parameter matrix; Obtain the texture parameter matrix and morphological parameter matrix of the target defect position area in each spectral channel image of each partition, retrieve the value of each strengthening element in the matrix and calculate the degree of defect feature manifestation in the strengthening direction, and take the average value to obtain the characteristic manifestation degree of the target defect position area in each spectral channel image of each partition; The characteristic appearance degree of each partition target defect type in the sensitive spectral channel image is calculated, and the minimum improvement rate relative to the characteristic appearance degree in other spectral channel images is multiplied by the improvement rate and the characteristic appearance degree in the sensitive spectral channel image to obtain the characteristic enhancement degree of each partition target defect type in the sensitive spectral channel image.

7. The machine vision-based cord fabric production defect recognition and analysis system according to claim 1, characterized in that: The condition for verifying the reliability of the composite spectral image target detection result is that the feature enhancement degree of each partition target defect type in the sensitive spectral channel image is greater than or equal to a preset feature enhancement degree reaching threshold.

8. The machine vision-based cord fabric production defect recognition and analysis system according to claim 1, characterized in that: The method of correcting the multispectral image sequence fusion weight according to the feature enhancement degree deviation includes: extracting the feature enhancement degree of the untrustworthy partition target defect type in the sensitive spectral channel image, constructing a positive adjustment factor based on the deviation ratio between the feature enhancement degree and a preset feature enhancement degree reaching threshold, and upwardly correcting the corresponding fusion weight coefficient of the sensitive spectral channel image; A reverse adjustment factor is constructed based on the deviation ratio of the characteristic appearance degree of the untrustworthy partition target defect type in the sensitive spectral channel image relative to the maximum characteristic appearance degree in other spectral channel images, and the corresponding fusion weight coefficients of other spectral channel images are corrected downward; The corrected fusion weight coefficients corresponding to each spectral channel image are normalized to regenerate the fusion weights of the multispectral image sequence.

9. The machine vision-based cord fabric production defect recognition and analysis system according to claim 1, characterized in that: The weight correction module also includes an iterative verification and rollback mechanism: setting the maximum number of iterations for multispectral image sequence fusion weight correction. If the verification is still unreliable after reaching the maximum number of iterations, the predefined benchmark weight configuration of each spectral channel image is automatically rolled back, triggering an alarm signal and generating an error log. The error log contains the weight adjustment trajectory during the iteration process and the credibility verification data change curve.

10. The machine vision-based cord fabric production defect recognition and analysis system according to claim 1, characterized in that: The control instructions include the following operations: When no defects are detected in the target detection results, a maintenance operation instruction is sent to the production control terminal to maintain the current production status; When the target detection results contain the number of partition defects, defect types and position coordinates, a stop command is sent to the production control terminal to terminate the cord fabric production process. Based on the mapping relationship between the composite spectral image and the actual physical coordinates of the cord fabric, the defect position coordinates in the image are converted into the actual defect position coordinates of the cord fabric molding surface. A notification window containing the actual defect position coordinates, defect type and quantity is generated and fed back to the operator terminal to trigger manual intervention.

Images