Program method for inspecting appearance of steel wire mesh

Abstract

The invention relates to the technical field of industry, and discloses a program method for appearance inspection of steel wire mesh cloth, which effectively solves the industrial problems of confusion of oil stains and shadows and misjudgment of gaps between holes and textures in the traditional method by synchronously integrating static images, dynamic video streams and three-dimensional surface data and constructing a multi-dimensional feature analysis model. A spectrum verification technology is innovatively introduced, misjudgment results under complex scenes such as metal reflection are automatically corrected based on a characteristic absorption peak intelligent matching mechanism, oil stain defect recognition reliability is improved, detection parameters are dynamically adjusted by adopting an online learning algorithm, and autonomous balance of precision and speed is realized. The transmission requirement of a high-speed production line is met, a quality evaluation report automatically triggers sorting of a high-precision mechanical arm, the detection efficiency is greatly improved, manual intervention is reduced, quality loss is effectively reduced, and verification is stable and reliable on various kinds of industrial-grade screen cloth for interference scenes such as steel wire reflection and mesh deformation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industry, in particular to a program method for appearance inspection of steel wire mesh cloth. BACKGROUND

[0002] Steel wire mesh is a general term for mesh materials woven or welded from low-carbon steel wire, medium-carbon steel wire, high-carbon steel wire, stainless steel wire and the like, and the production process includes ordinary weaving type, embossed weaving type and spot welding type. It is mainly made of steel wire as raw material and processed into a mesh shape by professional equipment, hence the name steel wire mesh.

[0003] However, the prior art has the following problems:

[0004] 1. Lack of defect dynamic tracking;

[0005] 2. Fragmentation of multi-modal data;

[0006] 3. Insufficient adaptive capability.

[0007] Therefore, the present application provides a program method for appearance inspection of steel wire mesh cloth, which develops a method that integrates multi-modal perception, dynamic defect evolution analysis and fusion of defect classification data to break through the bottleneck of the prior art. SUMMARY

[0008] (I) Technical problems solved

[0009] In view of the deficiencies of the prior art, the present application provides a program method for appearance inspection of steel wire mesh cloth, which solves the problems raised in the background art.

[0010] (II) Technical solutions

[0011] To achieve the above object, the present application provides the following technical solutions: a program method for appearance inspection of steel wire mesh cloth, comprising the following steps:

[0012] S1, collecting static image data and dynamic video stream data of the steel wire mesh cloth;

[0013] S2, performing multi-modal feature fusion processing on the static image data and dynamic video stream data to generate a multi-scale feature matrix of the surface of the steel wire mesh cloth;

[0014] S3, performing initial screening of defect regions on the multi-scale feature matrix based on an adaptive threshold segmentation algorithm to generate a set of potential defect coordinates;

[0015] S4, performing defect type identification on the set of potential defect coordinates through a convolutional neural network model to generate defect classification data, wherein the defect types include broken wire, hole, oil stain and warp and weft deviation;

[0016] S5, constructing a dynamic defect evolution map, tracking the trajectory of defects in the dynamic video stream data based on time series analysis to generate defect dynamic behavior data;

[0017] S6, fusing the defect classification data and defect dynamic behavior data to generate a steel mesh cloth comprehensive quality evaluation report;

[0018] S7, triggering an automatic sorting instruction according to the comprehensive quality evaluation report to control the mechanical arm to perform a physical separation operation on unqualified steel mesh cloth.

[0019] Preferably, the S1 comprises:

[0020] S11, acquiring a static image of the surface of the steel mesh cloth under constant illumination conditions by a high-resolution linear array camera to generate a static image data matrix I static ;

[0021] S12, acquiring a continuous frame video stream during the transmission of the steel mesh cloth by an industrial high-speed camera to generate a dynamic video stream data V dynamic ;

[0022] S13, reconstructing the three-dimensional surface topology of the steel mesh cloth by using photometric stereo vision algorithm to generate a surface depth map D map .

[0023] Preferably, the S2 comprises:

[0024] S21, wavelet multi-resolution decomposition is performed on I static to extract low-frequency texture features F texture and high-frequency edge features F edge ;

[0025] S22, optical flow field calculation is performed on V dynamic to generate a motion vector field M flow ;

[0026] S23, aligning F texture , F edge , M flow and D map according to the spatial coordinates to construct a four-dimensional feature tensor T feature = F texture , F edge , M flow , D map .

[0027] Preferably, the S3 comprises

[0028] S31, calculating the dynamic segmentation threshold of each channel in T feature by using Otsu algorithm:

[0029] θ k= μ k + α·σ k , k ∈ {1, 2, 3, 4}

[0030] wherein μ k is the mean value of channel k, σ k is the standard deviation, and α is an adaptive adjustment coefficient;

[0031] S32, mark the pixel points satisfying T feature (x, y, z, k) > θ k as potential defect points, and generate a binary mask B mask ;

[0032] S33, extract the minimum circumscribed rectangle coordinates of the defect region in B mask by connected component analysis, and generate a potential defect coordinate set

[0033] Preferably, after the S3, it further comprises:

[0034] S3A, performing multispectral verification analysis:

[0035] acquire the reflectance spectrum curve R(λ) of the defect region by the near-infrared spectrometer, the wavelength range is 1300nm;

[0036] calculate the feature absorption peak matching degree:

[0037]

[0038] wherein n is the number of absorption peaks, is the measured absorption peak position, is the corresponding peak position in the standard defect spectrum library, and σ is the standard deviation parameter of the Gaussian function, used to control the sensitivity of the matching degree calculation;

[0039] When M match <0.85, correct the defect type identified by S4, and generate verified defect classification data

[0040] Preferably, the S4 comprises:

[0041] S41, construct a double-branch convolutional neural network, the first branch adopts ResNet-34 architecture to process the static image block, and the second branch adopts 3D-CNN architecture to process the dynamic video cube;

[0042] S42, for each coordinate region in C defect , extract the image block P static from I static , and extract the spatiotemporal cube C cube from V dynamic ;

[0043] S43, merging the double-branch outputs through the feature fusion layer to generate a defect probability vector:

[0044] [p break ,p hole ,p stain ,p offset ]=Softmax(W f ·[F static ,F dynamic ])

[0045] where p break is the probability of broken filament defects, p hole is the probability of hole defects, p stain is the probability of oil stain defects, p offset is the probability of warp and weft offset defects, W f is the weight matrix of the feature fusion layer, F static is the feature vector of the static image branch, and F dynamic is the feature vector of the dynamic video branch;

[0046] S44, labeling defects with a maximum probability value exceeding 0.95 as confirmed defects to generate defect classification data D class .

[0047] Preferably, the S5 comprises:

[0048] S51, constructing a Kalman filter tracking model for each type of defect in D class .

[0049] X t =A·X t-1 +B·u t +w t

[0050] where X t is a defect state vector at time t, containing position, velocity, and size information, A is a state transition matrix, X t-1 is a defect state vector at time t-1, B is a control input matrix, u t is a control input vector, and w t is process noise.

[0051] S52, matching defect instances in consecutive frames through the Hungarian algorithm to generate a defect trajectory chain L track ={T1,T2,…,T m}.

[0052] S53, calculating the spatiotemporal feature quantities of each trajectory, including the average moving speed v avg , the area change rate ΔS, and the life cycle t life, generate defect dynamic behavior data D dynamic .

[0053] Preferably, the S6 comprises:

[0054] S61, construct defect severity evaluation function:

[0055]

[0056] where s size is the defect area ratio, β i is the weight coefficient, v avg is the average moving speed, ΔS is the area change rate, t is time, is the derivative of the area change rate with respect to time;

[0057] S62, generate quality score matrix Q class according to D severity and S score ;

[0058] S63, when Q score <η, mark as unqualified products, generate comprehensive quality evaluation report R report .

[0059] Preferably, after the S6, it further comprises:

[0060] S6A, implement detection parameter adaptive optimization:

[0061] Construct detection effect evaluation function:

[0062] E val = ω1·Recall+ ω2·Precision- ω3·T process

[0063] where Recall is the defect recall rate, Precision is the detection accuracy, T process is the single frame processing time, ω1, ω2, ω3 are weight coefficients.

[0064] Adjust the key parameter set Θ = {α, β, γ, τ} dynamically through the Bayesian optimization algorithm:

[0065]

[0066] where κ is the regularization coefficient, KL is the KL divergence, E val (Θ) is the detection effect evaluation function value when the parameter is Θ, p(Θ) is the probability distribution of the parameter Θ, p prior (Θ) is the prior probability distribution.

[0067] When E valWhen the boost rate is less than 1%, a convolutional neural network fine-tuning is triggered:

[0068]

[0069] Wherein is the validation set loss function containing difficult samples, mu is the momentum factor, eta is the learning rate, D val is the validation data set, W is the network weight, W prev is the previous weight update amount;

[0070] The optimized parameter set Θ new and the network weight W new are updated to the online detection system.

[0071] Preferably, the S7 comprises:

[0072] S71, convert the spatial coordinates of unqualified products in R report into pose data in the mechanical arm coordinate system;

[0073] S72, generate a collision-free grabbing trajectory through a path planning algorithm;

[0074] S73, control the six-axis mechanical arm to perform the sorting operation, and update the product batch database synchronously.

[0075] (Three) beneficial effects

[0076] Compared with the prior art, the present application provides a program method for appearance inspection of steel wire mesh cloth, which has the following beneficial effects:

[0077] 1. Multi-modal fusion improves detection accuracy

[0078] By synchronously integrating static images, dynamic video streams and three-dimensional surface data, a multi-dimensional feature analysis model is constructed, effectively solving the industry problems of oil stain and shadow confusion, hole and texture gap misjudgment in traditional methods, and improving the defect recognition accuracy.

[0079] 2. Multi-spectrum-vision collaborative anti-interference mechanism

[0080] By innovatively introducing spectral verification technology, based on the intelligent matching mechanism of characteristic absorption peak, the misjudgment result under complex scenes such as metal reflection is automatically corrected, and the recognition reliability of oil stain defects is improved.

[0081] 3. Self-adaptive intelligent optimization system

[0082] By adopting an online learning algorithm to dynamically adjust the detection parameters, the precision and speed are balanced automatically, and the high-speed production line transmission demand is met.

[0083] 4. Full-process automatic closed-loop control

[0084] Through the quality evaluation report, high-precision mechanical arm sorting is automatically triggered, the detection efficiency is greatly improved, manual intervention is reduced, and quality loss is effectively reduced.

[0085] 5. Strong adaptability to complex working conditions

[0086] Through the interference scene of steel wire reflection and mesh deformation, the false positive rate is reduced through feature fusion and parameter self-optimization mechanism, and stability and reliability are verified on various industrial meshes. BRIEF DESCRIPTION OF DRAWINGS

[0087] Figure 1 It is a schematic diagram of the overall system architecture of the present application. DETAILED DESCRIPTION

[0088] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0089] Please refer to Figure 1 The program method for appearance inspection of the steel wire mesh cloth includes the following steps:

[0090] S1, collecting static image data and dynamic video stream data of the steel wire mesh cloth;

[0091] S2, performing multi-modal feature fusion processing on the static image data and dynamic video stream data to generate a steel wire mesh cloth surface multi-scale feature matrix;

[0092] S3, performing initial screening of the multi-scale feature matrix based on an adaptive threshold segmentation algorithm to generate a potential defect coordinate set;

[0093] S4, performing defect type identification on the potential defect coordinate set through a convolutional neural network model to generate defect classification data, wherein the defect types include broken wires, holes, oil stains, and warp and weft offsets;

[0094] S5, constructing a dynamic defect evolution map, performing trajectory tracking on defects in the dynamic video stream data based on time series analysis to generate defect dynamic behavior data;

[0095] S6, fusing the defect classification data and the defect dynamic behavior data to generate a steel wire mesh cloth comprehensive quality evaluation report;

[0096] S7, triggering an automatic sorting instruction according to the comprehensive quality evaluation report to control a mechanical arm to perform physical separation operation on unqualified steel wire mesh cloth;

[0097] S1 includes:

[0098] S11, collect static images of the steel wire mesh cloth surface under constant illumination conditions by a high-resolution linear array camera, and generate a static image data matrix I static ;

[0099] S12, collect continuous frame video streams during the transmission of the steel wire mesh cloth by an industrial high-speed camera, and generate dynamic video stream data V dynamic ;

[0100] S13, reconstruct the three-dimensional surface topology of the steel wire mesh cloth by using photometric stereo vision algorithm, and generate a surface depth map D map ;

[0101] S2 includes:

[0102] S21, perform wavelet multi-resolution decomposition on I static , extract low-frequency texture features F texture and high-frequency edge features F edge ;

[0103] S22, perform optical flow field calculation on V dynamic , and generate a motion vector field M flow ;

[0104] S23, align F texture , F edge , M flow and D map according to spatial coordinates, and construct a four-dimensional feature tensor T feature =F texture ,F edge ,M flow ,D map ;

[0105] S3 includes

[0106] S31, calculate the dynamic segmentation threshold of each channel in T featuref by using Otsu algorithm:

[0107] θ k =μ k +α·σ k ,k∈{1,2,3,4}

[0108] wherein μ k is the mean value of channel k, σ k is the standard deviation, and α is an adaptive adjustment coefficient;

[0109] S32, mark the pixel points satisfying T feature (x,y,z,k)>θ k as potential defect points, and generate a binary mask Bmask ;

[0110] S33, extracting the minimum circumscribed rectangle coordinates of the defect area in the connected domain through connected domain analysis to generate a potential defect coordinate set mask

[0111] S3 also includes:

[0112] S3A, performing multispectral verification analysis:

[0113] Acquiring the reflection spectrum curve R(λ) of the defect area by a near-infrared spectrometer, the wavelength range being 1300nm;

[0114] Calculating the feature absorption peak matching degree:

[0115]

[0116] Wherein n is the number of absorption peaks, is the measured absorption peak position, is the corresponding peak position in the standard defect spectrum library, and σ is the standard deviation parameter of the Gaussian function, used to control the sensitivity of the matching degree calculation;

[0117] When M match <0.85, the defect type identified by S4 is corrected to generate verified defect classification data

[0118] S4 includes:

[0119] S41, constructing a double-branch convolutional neural network, the first branch adopting ResNet-34 architecture to process static image blocks, and the second branch adopting 3D-CNN architecture to process dynamic video cubes;

[0120] S42, extracting image blocks P defect from I static for each coordinate area in C static , and extracting spatiotemporal cubes C dynamic from V cube ;

[0121] S43, merging the double-branch outputs through a feature fusion layer to generate a defect probability vector:

[0122] [p break ,p hole ,p stain ,p offset ]=Softmax(W f ·[F static ,F dynamic ])

[0123] Wherein, p break ​p is the probability of broken filament defect hole p is the probability of hole defect stain p is the probability of oil stain defect offset W is the probability of warp and weft offset defect f F is the weight matrix of feature fusion layer static F is the feature vector of static image branch dynamic F is the feature vector of dynamic video branch

[0124] S44, defects with maximum probability value exceeding 0.95 are marked as confirmed defects, and defect classification data D is generated class ;

[0125] S5 includes:

[0126] S51, Kalman filter tracking model is constructed for each type of defect in D class :

[0127] X t = A·X t-1 + B·u t + w t

[0128] where X t is the defect state vector at time t, including position, speed and size information, A is the state transition matrix, X t-1 is the defect state vector at time t-1, B is the control input matrix, u t is the control input vector, w t is the process noise

[0129] S52, defect instances in consecutive frames are matched by Hungarian algorithm to generate defect trajectory chain L track = {T1, T2, …, T m};

[0130] S53, the spatiotemporal feature quantities of each trajectory are calculated, including average moving speed v avg , area change rate ΔS and life cycle t life , and defect dynamic behavior data D dynamic is generated

[0131] S6 includes:

[0132] S61, defect severity evaluation function is constructed

[0133]

[0134] where s size is the defect area ratio, β i is the weight coefficient, v avgis the average moving speed, ΔS is the area change rate, and t is the time, is the derivative of the area change rate with respect to time;

[0135] S62, according to D class and S severity generate a quality score matrix Q score ;

[0136] S63, when Q score <η is marked as unqualified products, and a comprehensive quality assessment report R report is generated;

[0137] S6 further includes:

[0138] S6A, implement adaptive optimization of detection parameters:

[0139] Construct a detection effect evaluation function:

[0140] E val = ω1·Recall + ω2·Precision - ω3·T process

[0141] where Recall is the defect recall rate, Precision is the detection accuracy, T process is the single-frame processing time, and ω1, ω2, ω3 are weight coefficients.

[0142] Adjust the key parameter set Θ = {α, β, γ, τ} dynamically through the Bayesian optimization algorithm:

[0143]

[0144] where κ is the regularization coefficient, KL is the KL divergence, E val (Θ) is the detection effect evaluation function value when the parameter is Θ, p(Θ) is the probability distribution of the parameter Θ, and p prior (Θ) is the prior probability distribution.

[0145] When the E val improvement rate of 10 consecutive batches is less than 1%, trigger the convolutional neural network fine-tuning:

[0146]

[0147] where is the validation set loss function containing difficult samples, μ is the momentum factor, η is the learning rate, D val is the validation data set, W is the network weight, and W prev is the previous weight update amount.

[0148] The optimized parameter set Θ new and network weight Wnew updating to an online detection system;

[0149] S7 comprises:

[0150] S71, converting the spatial coordinates of the unqualified product in the space into pose data in the mechanical arm coordinate system; report

[0151] S72, generating a collision-free grabbing trajectory through a path planning algorithm;

[0152] S73, controlling the six-axis mechanical arm to perform the sorting operation, and synchronously updating the product batch database.

[0153] Embodiment: Industrial application example of flexible steel mesh cloth intelligent detection system

[0154] This embodiment is implemented in a steel mesh cloth production line special for automobile filter elements. The mesh cloth has high ductility characteristics, and the transmission speed is stable at 4 m / s. The detection system adopts a three-in-one architecture: a linear array camera with a polarizer is equipped at the front end to reduce metal reflection, a high-speed camera is installed laterally to capture dynamic deformation, and a laser displacement sensor is integrated at the bottom to build a three-dimensional model.

[0155] The implementation process starts from the synchronous acquisition of multi-source data. When the mesh cloth passes through the detection station, the system synchronously acquires three groups of key data: the linear array camera acquires the surface texture with a scanning accuracy of 20,000 lines per second, the high-speed camera records the microscopic deformation of the steel wire at 1,000 frames per second, and the laser sensor generates a three-dimensional point cloud with a precision of 0.02 mm. These three groups of data are accurately aligned through time stamps to form a four-dimensional data cube that integrates spatial coordinates, time series, and material characteristics.

[0156] In the feature fusion stage, a hierarchical processing strategy is adopted. First, the static image is subjected to multi-scale wavelet decomposition to extract texture features at different resolutions. Second, the dense optical flow field between consecutive frames is calculated to quantify the deformation trajectory of the mesh structure. Finally, the depth map is spatially registered with the optical features to establish a grid-based feature map. This fusion method effectively distinguishes between real steel wire fractures and visual artifacts caused by light, especially when the mesh cloth undergoes elastic stretching. It can accurately identify microscopic cracks above 0.3 mm.

[0157] In the defect recognition stage, a dual-channel intelligent analysis model is deployed. For the acquired 256x256 pixel image block, the static analysis channel uses a lightweight convolutional network to complete the positioning and classification of broken wires and holes within 3 ms. The dynamic analysis channel processes 32 frames of video clips and identifies oil diffusion patterns through spatiotemporal feature modeling. When a suspected oil stain area is detected, the system automatically triggers a near-infrared spectral probe to verify the composition of the pollutant by comparing the C-H bond characteristic absorption peak. This reduces the false positive rate of traditional methods from 25% to below 1.2%.

[0158] ​The processing of dynamic defects embodies the system innovation value. When the latitude and longitude offset defects are detected, the system starts the motion trajectory tracking engine: based on the Kalman filtering algorithm, the moving path of the defects is predicted, the target entity in the continuous frames is associated through the Hungarian algorithm, the whole process from the initial deformation to the structural instability is recorded completely, and the actual case of the production line shows that the system successfully captures the 0.5° mesh deflection caused by the sudden change of the transmission speed, and the quality warning is given 10 seconds earlier than the traditional detection.

[0159] The management closed loop is formed by the quality decision and the execution link, the evaluation module generates the quality score by comprehensively considering the defect type, size, motion speed and other parameters, the sorting instruction is triggered when the score is lower than 0.92, the coordinate conversion system maps the defect position to the working space of the six-axis mechanical arm, the RRT* algorithm is used to plan the optimal grabbing path, and the high-speed sorting of 150ms / time is realized by cooperating with the vacuum suction cup. After the continuous 72-hour pressure test, the system maintains zero missed detection in 52,000 detections, successfully intercepts 47 critical defects, and avoids 150,000 yuan of potential loss.

[0160] The innovation performance of the embodiment is embodied in three industrial scenes: firstly, the processing advantage of the reflective net cloth is obvious, the oil stain detection accuracy is improved through the cooperation of polarization imaging and multi-spectrum verification, secondly, the quality traceability system of dynamic defects is established, and the space-time evolution law of 12 types of process defects is recorded completely, and thirdly, the self-evolution of the detection standard is realized, the model parameters are automatically optimized once every 500 rolls of net cloth, the new material research and development demand is adapted, and the scheme makes the product qualified rate increase and the sorting efficiency is 17 times of the traditional manual, which saves the quality cost of 3.8 million yuan per year for the automobile filter manufacturer.

[0161] It should be noted that, in the present document, relational terms such as first and second and the like can be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. An element proceeded by "comprises... a" does not, without more constraints, exclude the presence of additional identical elements in the process, method, article, or apparatus that comprises the element.

[0162] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, combinations, and alterations can be made to the embodiments without departing from the principles and spirit of the application, and the scope of the application is defined by the appended claims and their equivalents.

Claims

1. A procedure for wire cloth appearance inspection characterized by: The method comprises the following steps: S1, collecting static image data and dynamic video stream data of the steel wire mesh; S2, performing multi-modal feature fusion processing on the static image data and dynamic video stream data to generate a multi-scale feature matrix of the steel wire mesh surface; S3, performing initial screening of the multi-scale feature matrix based on an adaptive threshold segmentation algorithm to generate a potential defect coordinate set; S4, performing defect type identification on the potential defect coordinate set through a convolutional neural network model to generate defect classification data, the defect types including broken wires, holes, oil stains, and warp and weft offsets; S5, constructing a dynamic defect evolution map, performing trajectory tracking on defects in the dynamic video stream data based on time series analysis to generate defect dynamic behavior data; S6, fusing the defect classification data and defect dynamic behavior data to generate a comprehensive quality evaluation report of the steel wire mesh; S7, triggering an automatic sorting instruction according to the comprehensive quality evaluation report to control a mechanical arm to perform a physical separation operation on unqualified steel wire mesh.

2. A procedure for visual inspection of wire cloth according to claim 1, characterized in that: The S1 comprises: S11, acquire the static image data matrix I of the steel wire mesh cloth surface by a high-resolution linear array camera under constant illumination conditions static ; S12, collect continuous frame video stream by industrial high-speed camera in the process of steel wire mesh cloth transmission, generate dynamic video stream data V dynamic ; S13, reconstruct the three-dimensional surface topology of the steel wire mesh cloth using photometric stereo vision algorithm to generate the surface depth map D map .

3. A procedure for visual inspection of wire cloth according to claim 1, characterized in that: The S2 comprises: S21, to I static Wavelet multi-resolution decomposition is performed to extract low-frequency texture features F texture and high-frequency edge features F edge ; S22, to V dynamic performing optical flow field calculation to generate motion vector field M flow ; S23, F texture , F edge , M flow , and D map are aligned by spatial coordinates, a four-dimensional feature tensor T feature = [F texture , F edge , M flow , D map ] is constructed.

4. A procedure for visual inspection of wire cloth according to claim 1, characterized in that: The S3 comprises S31, calculate T using Otsu algorithm featuref Dynamic segmentation threshold of each channel in the middle θ k = μ k + α · σ k , k e {1, 2, 3, 4} where μk k is the mean value of channel k, σk k is the standard deviation, and α is an adaptive adjustment factor; S32, mark the pixel point satisfying T feature (x, y, z, k) > θ k as a potential defect point, and generate a binary mask B mask ; S33, extracting B by connected domain analysis mask the minimum circumscribed rectangle coordinates of the defect area in the middle, generating a potential defect coordinate set 5. A procedure for visual inspection of wire cloth according to claim 1, characterized in that: After the S3, the following steps are further included: S3A, performing multi-spectral verification analysis: Acquiring a reflection spectrum curve R(λ) of the defect area by a near-infrared spectrometer, the wavelength range being 1300nm; Calculating the feature absorption peak matching degree: where n is the number of absorption peaks, is the measured absorption peak position, is the corresponding peak position in the standard defect spectrum library, and σ is the standard deviation parameter of the Gaussian function, which is used to control the sensitivity of the matching degree calculation. When M match <0.85, the defect type identified for S4 is corrected, generating post-verification defect classification data 6. A procedure for visual inspection of wire cloth according to claim 1, characterized in that: The S4 comprises: S41, constructing a double-branch convolutional neural network, a first branch adopting a ResNet-34 architecture to process static image blocks, and a second branch adopting a 3D-CNN architecture to process dynamic video cubes; S42, to C defect from each coordinate region, I static extracting image patches P static from V dynamic extracting spatio-temporal cubes C cube ; S43, merging the double-branch outputs through a feature fusion layer to generate a defect probability vector: [p break ,p hole ,p stain ,p offset ] = Softmax(W f · [F static , F dynamic ]) wherein p break is the probability of a broken filament defect, p hole is the probability of a hole defect, p stain is the probability of an oil stain defect, p offset is the probability of a warp and weft offset defect, W f is a weight matrix of the feature fusion layer, F static is a feature vector of the static image branch, F dynamic is a feature vector of the dynamic video branch; S44, defects with a maximum probability value exceeding 0.95 are marked as confirmed defects, and defect classification data D is generated class .

7. A procedure for visual inspection of wire cloth according to claim 1, characterized in that: The S5 comprises: S51, to D class Kallman filter tracking model is constructed for each type of defect: X t = A · X t-1 + B · u t + w t where X t is the defect state vector at time t, containing position, velocity and size information, A is the state transition matrix, X t-1 is the defect state vector at time t-1, B is the control input matrix, u t is the control input vector, w t is the process noise; S52, generate a defect track chain L by matching the defect instances in the continuous frames through the Hungarian algorithm track = {T1, T2, …, T m}; S53, calculate the space-time feature quantity of each trajectory, including average moving speed v avg , area change rate ΔS and life cycle t life , generate defect dynamic behavior data D dynamic .

8. A procedure for visual inspection of wire cloth according to claim 1, characterized in that: The S6 comprises: S61, constructing a defect severity evaluation function: wherein s size is the area ratio of defects, β i is the weight coefficient, v avg is the average moving speed, ΔS is the area change rate, t is time, is the derivative of the area change rate with respect to time; S62, according to D class and S severity generate a quality score matrix Q score ; S63, when Q score <η when marked as non-conforming product, generate a comprehensive quality assessment report R report .

9. A procedure for visual inspection of wire cloth according to claim 1, characterized in that: After the S6, the following steps are further included: S6A, implementing detection parameter adaptive optimization: Constructing a detection effect evaluation function: E val = ω1·Recall + ω2·Precision - ω3·T process where Recall is the defect recall rate, Precision is the detection precision, T process is the single-frame processing time, ω1, ω2, ω3 are weight coefficients. Dynamically adjusting a key parameter set Θ={α,β,γ,τ} through a Bayesian optimization algorithm: where k is a regularization coefficient, KL is the KL divergence, E val (Θ) is the evaluation function value of the detection effect when the parameter is Θ, p(Θ) is the probability distribution of the parameter Θ, p prior (Θ) is the prior probability distribution; When 10 consecutive batches of E val Trigger the convolutional neural network fine-tuning when the lift rate is less than 1%: wherein is the validation set loss function for the difficult samples, μ is the momentum factor, η is the learning rate, D val is the validation data set, W is the network weights, W prev is the previous weight update; The optimized parameter set Θ new and network weights W new are updated to the online detection system.

10. A procedure for visual inspection of wire cloth according to claim 1, characterized in that: The S7 comprises: S71, R report converts the spatial coordinates of the non-conforming product into pose data in the mechanical arm coordinate system; S72, generating a collision-free grabbing trajectory through a path planning algorithm; S73, controlling a six-axis mechanical arm to perform a sorting operation and synchronously updating a product batch database.

Images