PVC lamp box fabric surface defect online detection system based on machine vision

Abstract

The invention discloses a PVC lamp-box fabric surface defect online detection system based on machine vision, and relates to the technical field of PVC lamp-box fabric surface detection.The system comprises a data acquisition module, a model training module and a regulation and control output module; according to the technical key points, three-dimensional surface information of the PVC lamp-box fabric is collected, the three-dimensional surface information is imported into a preset defect analysis model, the defect analysis model is established on the basis of a CNN architecture, a loss function of the defect analysis model is updated, and a surface defect category area of the PVC lamp-box fabric is output on the basis of the loss function and the defect analysis model; adjusting and repairing the surface defects of the PVC lamp box fabric according to a preset adjustment strategy based on the surface defect category area; according to the invention, process closed-loop optimization can be realized, the detection efficiency and speed are improved, and the omission ratio is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of PVC light box cloth surface detection, in particular to a PVC light box cloth surface defect online detection system based on machine vision. BACKGROUND

[0002] PVC light box cloth is a special material widely used in outdoor advertising, commercial display and other fields, which has the characteristics of durability, waterproofness, light transmission, etc. The surface quality directly affects the product appearance and service life. With the development of machine vision technology, machine vision technology is also used in the field of industrial production quality detection to detect defects on the surface of PVC light box cloth online. Defects such as uneven coating (spots, color difference or inconsistent gloss), scratches or scratches, bubbles or blisters, pinholes or missed coating, impurity particles, etc. may cause uneven light transmission, ink adhesion failure and other problems. During detection, the influence of the environment (such as temperature and humidity changes, light intensity) on the online detection of PVC light box cloth needs to be considered, which will enhance the complexity of analysis. In the process of PVC light box cloth production, there is a lack of real-time closed-loop feedback mechanism under high-speed production line, which reduces the detection efficiency and speed and increases the missed detection rate. SUMMARY

[0003] (I) Technical problems solved In view of the shortcomings of the prior art, the present application provides a PVC light box cloth surface defect online detection system based on machine vision, which urgently needs an online detection system that integrates three-dimensional image and multi-spectral reflection analysis. By dividing high-risk defect areas or low-risk defect areas, and adjusting and repairing the defect area categories, process closed-loop optimization is realized, and the problems in the background art are solved.

[0004] (II) Technical solutions To achieve the above purpose, the present application realizes the following technical solutions: The present application provides a PVC light box cloth surface defect online detection system based on machine vision, which comprises: An information acquisition module for acquiring three-dimensional surface information of PVC light box cloth, wherein the three-dimensional surface information is obtained based on three-dimensional surface images scanned from PVC light box cloth; A model training module for importing the three-dimensional surface information into a preset defect analysis model, wherein the defect analysis model is established based on CNN architecture, the loss function of the defect analysis model is updated, and the surface defect category area of the PVC light box cloth is output based on the loss function and the defect analysis model, wherein the surface defect category area includes high-risk defect area and low-risk defect area; The regulation output module is configured to adjust and repair the surface defects of the PVC light box cloth according to a preset adjustment strategy based on the surface defect category area.

[0005] Further, the three-dimensional surface information is introduced into a preset defect analysis model, which includes: The preset defect analysis model at least includes an input layer, a recognition layer, a marking layer, and a calculation layer; The three-dimensional surface information is input into the input layer, and the three-dimensional surface features are extracted through the recognition layer. The three-dimensional surface features are spatially labeled through the marking layer to obtain surface labeled features. The surface labeled features are input into the analysis layer to obtain abnormal analysis results, including first, second, third, and fourth analysis results. The abnormal analysis results are fused and input into the calculation layer to obtain updated weight coefficients. The loss function of the defect analysis model is updated according to the updated weight coefficients, and the PVC light box cloth surface defect category area is output based on the updated defect analysis model.

[0006] Further, the three-dimensional surface features are extracted through the recognition layer, which includes: The three-dimensional surface features are point-sorted, and the feature points outside the key areas are removed. The key areas include: joint, coating coverage area, edge contour, and curved transition part. The sorted feature points are classified and summarized to obtain a joint point set, a coating coverage point set, an edge contour point set, and a curved transition point set. Based on the preset direction, the preset interval, and the position information of the feature points, a plurality of line segments are used to connect the classified feature points, and the distribution characteristics of the line segments are analyzed. The results of the analysis constitute the three-dimensional surface features. The line segments at least include any one of straight lines and curves, and the distribution characteristics of the straight lines are the slopes of the straight lines, and the distribution characteristics of the curves are the slopes, curvatures, and curvature radii of the curves.

[0007] Further, the surface labeled features are input into the analysis layer to obtain abnormal analysis results, which include: The analysis layer includes primary analysis and secondary analysis. Primary analysis: standardize the three-dimensional surface features under a plurality of line segments and weighted sum to obtain surface parameter values. Each surface parameter value is compared and analyzed with the corresponding preset adaptive standard value: When the surface parameter value is less than the adaptive standard value, it indicates that there is no obvious defect anomaly. When the surface parameter value is greater than or equal to the adaptive standard value, it indicates that there is an obvious defect anomaly, and secondary analysis is performed. Secondary analysis: The preset light source is controlled to move based on a detection device, different waveband reflection information of the PVC light box cloth surface is extracted based on an irradiation angle, an intensity and a reflection line trajectory of the preset light source, the reflection information includes ultraviolet waveband, visible light waveband and near-infrared light waveband, and thus a first reflectivity of the ultraviolet waveband corresponding to the PVC light box cloth, a second reflectivity of the visible light waveband and a third reflectivity of the near-infrared light waveband are obtained respectively; A right-angle coordinate system is established, the first reflectivity / second reflectivity is taken as an X axis, and the first reflectivity / third reflectivity is taken as a Y axis, a trajectory change curve about the waveband is drawn, and the curve mutation coefficient and the curvature gradient coefficient of each key area are determined based on the trajectory change curve; Based on the curve mutation coefficient and the curvature gradient coefficient, the abnormal sequences of each key area are identified, including a first abnormal sequence and a second abnormal sequence, and a defect position coordinate set is generated.

[0008] Further, the curve mutation coefficient and the curvature gradient coefficient of each key area are determined based on the trajectory change curve, including: A preset time sliding window; Based on the trajectory change curve, a peak point set and a valley point set are determined, a difference absolute value between the peak point set and the valley point set is marked as a span value, a standard deviation of each trajectory point in the time sliding window is obtained, a product of the standard deviation of each trajectory point and the corresponding span value is calculated, and the curve mutation coefficient is obtained; Based on the trajectory change curve, a slope set is determined, a fluctuation value of the slope set in the time sliding window and a slope standard deviation of each trajectory point are obtained, a square of a difference value between the slope standard deviation and the fluctuation value is calculated, and the curvature gradient coefficient is obtained.

[0009] Further, based on the curve mutation coefficient and the curvature gradient coefficient, the abnormal sequences of each key area are identified, including The curve mutation coefficient and the curvature gradient coefficient are respectively sorted in descending order, the first n corresponding curve mutation coefficients and the curvature gradient coefficients are selected, and the first abnormal sequence and the second abnormal sequence are respectively mapped to the order arrangement in the trajectory change curve; The Pearson correlation coefficient of the first abnormal sequence and the second abnormal sequence is calculated, the Shannon entropy of all the curve mutation coefficients in the first abnormal sequence is calculated, the ratio of the absolute value of the Pearson correlation coefficient to the Shannon entropy is calculated, and the abnormal overlap index is obtained; the abnormal overlap index is taken as an input of a clustering algorithm, two clustering clusters are obtained, and a high-risk defect area and a low-risk defect area are divided; Based on the defect position coordinate set, the high-risk defect area and the low-risk defect area, an abnormality analysis result is obtained, including: a first analysis result, a second analysis result, a third analysis result and a fourth analysis result; wherein the first analysis result represents a joint defect, the second analysis result represents a coating coverage defect, the third analysis result represents an edge profile defect, and the fourth analysis result represents a curved surface transition defect.

[0010] Further, the abnormality analysis results are fused, including: the first analysis result, the second analysis result, the third analysis result and the fourth analysis result are concatenated and fused.

[0011] Further, the weight coefficient is updated, and the calculation formula is: ; In the formula, α represents the updated weight coefficient, μ represents the basic coefficient, ε represents the control coefficient, K1 represents the position set corresponding to the high-risk defect area in the training set, K2 represents the position set corresponding to the low-risk defect area in the training set, and δ (G) represents the mean value of all abnormal overlap indexes in all high-risk defect areas.

[0012] Further, the PVC light box cloth is adjusted and repaired according to the preset rules, including: The high-risk defect area is immediately repaired; The low-risk defect area is temporarily repaired.

[0013] In a second aspect, the application provides a PVC light box cloth surface defect online detection method based on machine vision, which comprises: Collecting three-dimensional surface information of the PVC light box cloth, wherein the three-dimensional surface information is obtained based on a three-dimensional surface image obtained by scanning the PVC light box cloth; The three-dimensional surface information is imported into a preset defect analysis model, and the defect analysis model is established based on a CNN architecture; the loss function of the defect analysis model is updated, and based on the loss function and the defect analysis model, the surface defect category area of the PVC light box cloth is output, wherein the surface defect category area includes a high-risk defect area and a low-risk defect area; Based on the surface defect category area, the surface defect of the PVC light box cloth is adjusted and repaired according to a preset adjustment strategy.

[0014] (Three) beneficial effects The application provides a PVC light box cloth surface defect online detection system based on machine vision, which has the following beneficial effects: 1. The application can realize high-precision extraction of the reflection characteristics of the PVC light box cloth surface by expanding the dynamic angle adjustment of the light source, multi-band intensity adaptation and reflection trajectory; in this process, by designing the ratio of ultraviolet band to visible light band and the ratio of ultraviolet band to near infrared wave, the influence of illumination intensity change or temperature and humidity change can be eliminated, and defects on the surface of the PVC light box cloth, such as uneven coating, can be detected through collaborative analysis among ultraviolet, visible light and near infrared light, and the sensitivity of defect identification is high. 2. The application abstracts the three-dimensional surface features into a plurality of line segments, including straight lines and curves, and extracts the corresponding distribution characteristics for feature point classification and summary, to obtain a set of joint points, a set of coating coverage points, a set of edge contour points and a set of curved surface transition points; this facilitates the construction of trajectory change curves in different regions later, supports full-automatic quality inspection of complex curved surfaces (arc-shaped, cylindrical) to generate a corresponding set of defect position coordinates; 3. The application sets a trajectory change curve, identifies a first abnormal sequence based on a curve mutation coefficient, identifies a second abnormal sequence based on a curvature gradient coefficient, calculates an abnormal overlap index based on the first abnormal sequence and the second abnormal sequence, divides high-risk defect areas and low-risk defect areas, and then calculates weight parameters in different defect category areas, which considers the relative change influence characteristics between the high-risk defect area and the low-risk defect area in the training sample of the PVC light box cloth surface defect recognition in a complex scene, improves the accuracy of the defect analysis model (CNN neural network model) for PVC light box cloth surface defect recognition and the efficiency of online detection; 4. The application adjusts and repairs the PVC light box cloth in a targeted manner, including immediate repair of high-risk defect areas and temporary repair of low-risk areas, which avoids greater losses by prioritizing high-risk defects, and reduces unnecessary repair work, saves materials and time, and provides a high-robustness detection benchmark. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 is a module schematic diagram of a PVC light box cloth surface defect online detection system according to an exemplary embodiment; Figure 2 is a flowchart of a model training module according to an exemplary embodiment. DETAILED DESCRIPTION

[0016] The technical solutions in the embodiments of the application will be described in detail below with reference to the drawings of the embodiments of the application. Obviously, the described embodiments are only a part of the embodiments of the application, not all. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0017] Embodiment 1: The embodiment of the present application provides a PVC light box cloth surface defect online detection system based on machine vision; Figure 1 It is a module schematic diagram of a PVC light box cloth surface defect online detection system according to an exemplary embodiment, Figure 2 It is a flow schematic diagram of a model training module according to an exemplary embodiment; please refer to Figures 1 to 2 The system comprises a data acquisition module, a model training module and a regulation and output module, and the data acquisition module, the model training module and the regulation and output module are in communication connection; The following is the explanation and description of each module: Data acquisition module: is used for collecting three-dimensional surface information of PVC light box cloth, wherein the three-dimensional surface information is obtained based on three-dimensional surface images scanned and acquired from the PVC light box cloth; Wherein, the three-dimensional surface images are acquired by scanning the PVC light box cloth through a detection device, the coverage range of which is centered on the detection device, and the maximum scanning distance is the width of the PVC light box cloth; the detection device can be a 3D line laser scanner or an infrared thermal imaging system; the key areas collected include the joint, the coating coverage area, the edge profile and the curved transition part; in this step, the three-dimensional surface information of the PVC light box cloth is acquired to effectively determine the surface parameter values of each key area; The model training module is used for importing the three-dimensional surface information into a preset defect analysis model, and the defect analysis model is established based on a CNN architecture; the loss function of the defect analysis model is updated, and based on the loss function and the defect analysis model, the surface defect category area of the PVC light box cloth is output, wherein the surface defect category area includes a high-risk defect area and a low-risk defect area; Specifically, the preset defect analysis model comprises an input layer, an identification layer, a marking layer, a calculation layer and an output layer; The three-dimensional surface information is input into the input layer, and the three-dimensional surface features are extracted through the identification layer; the three-dimensional surface features are spatially labeled through the marking layer to obtain surface labeled features; the surface labeled features are input into the analysis layer to obtain abnormal analysis results, including first, second, third and fourth analysis results; the abnormal analysis results are fused and input into the calculation layer to obtain updated weight coefficients; According to the updated weight coefficients, the loss function of the defect analysis model is updated, and based on the updated defect analysis model, the surface defect category area of the PVC light box cloth is output; The three-dimensional surface features are extracted through the identification layer, including: The three-dimensional surface features are point-sorted, and the feature points outside the key areas are removed; wherein the key areas include the joint, the coating coverage area, the edge profile and the curved transition part; The classified and summarized feature points are classified and summarized to obtain a set of joint points, a set of coating covering points, a set of edge contour points, and a set of curved surface transition points; Based on the preset direction, the preset interval, and the position information of the feature points, a plurality of line segments are used to connect the classified feature points, and the distribution characteristics of the line segments are analyzed, and the analysis results constitute the three-dimensional surface features; wherein the line segments at least include any one of a straight line and a curve, and the distribution characteristics of the straight line are the slope of the straight line, and the distribution characteristics of the curve are the slope, the curvature, and the curvature radius of the curve; It should be noted that the preset direction and the preset interval are the scanning direction and the scanning interval of the detection equipment, which can be set according to requirements, and the position information of the feature points is determined based on the three-dimensional surface image; For example, a three-dimensional rectangular coordinate system is constructed with the detection equipment (such as a 3D line laser scanner) as the origin: the width direction of the PVC light box fabric is the X axis, the length direction of the PVC light box fabric is the Y axis, and the vertical to the surface is the Z axis; the scanning radius is equal to the width of the PVC light box fabric (such as 2 meters), so as to ensure that the spatial coordinates of all feature points (joints, coatings, edges, and curved surface transitions) satisfy: X 2 +Y 2 ≤width 2 ; The three-dimensional surface features are spatially labeled by the labeling layer to obtain surface labeled features, and the surface labeled features include position features and content features; For example: Feature segmentation: a semantic segmentation algorithm is used to identify key areas and divide the three-dimensional coordinate boundary position information of each key area; wherein the semantic segmentation algorithm is established based on a U-Net architecture; Label generation: the surface detected in the key area is labeled using machine vision, and coordinate set generation (X, Y, Z) and type label are performed; Parameter labeling: for the surface corresponding to each key area, the position information of the feature points and the distribution characteristics of the line segments are extracted, and a quantization operation (such as forming a quantized heat map) is performed; The surface labeled features are input into the analysis layer to obtain abnormal analysis results, including: The analysis layer includes primary analysis and secondary analysis; Primary analysis: The three-dimensional surface features under a plurality of line segments are standardized and weighted summed to obtain a surface parameter value; When it is a straight line, the slopes under each straight line are standardized and weighted summed to obtain a surface parameter value; When it is a curve, the slope, the curvature, and the curvature radius under each curve are standardized and weighted summed to obtain a surface parameter value; The weight values of the weighted sum are obtained by at least one of the following manners: Manner one: the weight values are determined by using a coefficient of variation method, the coefficient of variation method is a method for weighting each index according to the variation degree of the current value and the target value of each index, and specifically includes: if the numerical difference of an index is large, the index can clearly distinguish each evaluated object, and the index has rich distinguishing information, and thus the index should be given a larger weight; on the contrary, if the numerical difference of each evaluated object on an index is small, the index has weak ability to distinguish each evaluated object, and thus the index should be given a smaller weight; this method directly uses the information contained in each index, and the weight of the index is obtained by calculation, and thus the method has objectivity; Manner two: the weight values are determined by using a high-dimensional target evolution algorithm: A reference vector set Omega uniformly distributed in an S-dimensional target search space is generated to provide a direction for population evolution; An initial population is randomly generated, the iteration number is set as tmax, the population is composed of m particles, and each particle has a unique position vector and a velocity vector; A binary bidding tournament selection strategy is used to select particles from the population into a mating pool in turn until the number of particles in the mating pool reaches m, two particles are randomly selected from the mating pool each time, simulated binary crossover and polynomial mutation operations are performed on the two particles to generate two new particles into a child population, and new individuals are repeatedly generated until the number of individuals in the child population reaches m; The parent population and the child population are combined to obtain a hybrid population, the fitness values of all particles in the hybrid population are calculated, the ideal point Fmin of the hybrid population is calculated, and the minimum value of each target function value at the ideal point Fmin is taken as the minimum value of the hybrid population on the target to update the velocity vector and the position vector of the particle; wherein, the position of each particle is a possible solution, the fitness value of the particle can be calculated by bringing the particle into a target function, and the solution is better and the position is better as the fitness value is smaller; The ideal point Fmin drives the weight distribution to ensure that the key target is optimized first, that is, the weight value is regarded as the inertia weight in the particle swarm algorithm to control the updating direction of the parameter; The surface parameter value is compared and analyzed with the corresponding preset adaptive standard value: When the surface parameter value is less than the adaptive standard value, it indicates that there is no obvious defect anomaly; When the surface parameter value is greater than or equal to the adaptive standard value, it indicates that there is an obvious defect anomaly, and secondary analysis is performed; Secondary analysis: The preset light source is controlled to move based on a detection device, the irradiation angle (dividing the width of the PVC light box fabric into fan-shaped areas, adapting to curved surface detection, and performing area segmentation scanning), intensity (integrating three independently controllable waveband light sources of ultraviolet (365-400 nm), visible light (400-700 nm), and near-infrared (700-850 nm)), and reflection line trajectory (recording reflection line coordinates (X, Y) and reflectivity distribution) of the preset light source are controlled; Wherein, the power of the controllable waveband light source is dynamically adjusted through the material gram weight of the PVC light box fabric: ; In the formula, represents the adjusted power corresponding to the controllable waveband, I0 represents the reference power of the controllable waveband (for example: the reference power of the ultraviolet waveband is 100 mW / cm², the reference intensity of the visible light is 5000 lux, and the reference power of the near-infrared light source is 200 mW / cm²), and Weight represents the gram weight of the PVC light box fabric; The reflection information of the PVC light box fabric surface in different wavebands is extracted, including the ultraviolet waveband, the visible light waveband, and the near-infrared light waveband, so as to obtain the first reflectivity of the ultraviolet waveband, the second reflectivity of the visible light waveband, and the third reflectivity of the near-infrared waveband corresponding to the PVC light box fabric, respectively; A rectangular coordinate system is established, the first reflectivity / second reflectivity is taken as the X-axis, and the first reflectivity / third reflectivity is taken as the Y-axis, to draw a trajectory change curve about the waveband, and to determine the curve mutation coefficient and the curvature gradient coefficient of each key area based on the trajectory change curve; Through the dynamic angle adjustment, multi-waveband intensity adaptation, and reflection trajectory of the extended light source, high-precision extraction of the reflection characteristics of the PVC light box fabric surface can be realized. In actual life, the reflection characteristic information of the PVC light box fabric surface is affected by changes in light intensity and changes in temperature and humidity. Specifically, the PVC film is originally a kind of heat-sensitive material, and its softness and hardness will change with the change of air temperature (for example, high air temperature will soften, and low air temperature will harden). In this embodiment, the ratio of the ultraviolet waveband to the visible light waveband and the ratio of the ultraviolet waveband to the near-infrared wave are designed to eliminate the influence of changes in light intensity or changes in temperature and humidity. The ultraviolet waveband may be sensitive to the aging or micro-defects of the coating, the visible light reflects the color and surface texture, and the near-infrared may penetrate the surface layer to detect the substrate structure. The surface characteristics of the material are identified by the difference in reflection characteristics of different wavebands. The collaborative analysis of ultraviolet-visible light and ultraviolet-near-infrared can detect uneven coating thickness and substrate micro-cracks, and has high sensitivity to defect identification. The curve mutation coefficient and the curvature gradient coefficient of each key area are determined based on the trajectory change curve, including: A time sliding window [kΔt, kΔt+T] is preset, and k=1, 2, …, n, n is an integer, and T represents time sequence, and the time sliding window is moved on the trajectory change curve; Based on the trajectory change curve, a peak point set and a valley point set are determined, a difference absolute value between the peak point set and the valley point set is marked as a span value, a standard deviation of each trajectory point in the time sliding window is obtained, a product of the standard deviation of each trajectory point and the corresponding span value is calculated, and a curve mutation coefficient is obtained; the span value reflects variability or dynamic range of the trajectory change curve, and the greater the value is, the more obvious the anomaly of the trajectory is; Based on the first order derivative of the trajectory change curve, a slope set is determined, a fluctuation value of the slope set in the time sliding window and a slope standard deviation of each trajectory point are obtained, a square of a difference between the slope standard deviation and the fluctuation value is calculated, and a curvature gradient coefficient is obtained; the fluctuation value is an average value of a difference between a maximum value and a minimum value of the trajectory change curve in a local time sliding window; Based on the curve mutation coefficient and the curvature gradient coefficient, abnormal sequences of the key regions are identified, including a first abnormal sequence and a second abnormal sequence, and a defect position coordinate set is generated; Based on the curve mutation coefficient and the curvature gradient coefficient, abnormal sequences of the key regions are identified, including: The curve mutation coefficient and the curvature gradient coefficient are sorted in descending order respectively, the first n corresponding curve mutation coefficients and curvature gradient coefficients are selected, and the first abnormal sequence and the second abnormal sequence are formed by mapping the curve mutation coefficients and the curvature gradient coefficients to the trajectory change curve in order respectively; The Pearson correlation coefficient of the first abnormal sequence and the second abnormal sequence is calculated, the Shannon entropy of all the curve mutation coefficients in the first abnormal sequence is calculated, the ratio of the absolute value of the Pearson correlation coefficient to the Shannon entropy is calculated, and is marked as an abnormal overlap index; the abnormal overlap index is used as an input of a clustering algorithm, two clustering clusters are obtained, and a high-risk defect region and a low-risk defect region are divided; The clustering algorithm uses K-means clustering, and the steps are as follows: Input parameters: the abnormal overlap index is input as a sample; Algorithm process: The number of clusters is set to 2, and the cluster centers are initialized; The distance between each sample and each cluster center is calculated, so that the sum of the distances from each sample to the cluster center is minimized; The 3D coordinates of the samples in the cluster are projected onto the position information of the feature points, and a defect position coordinate set is generated; The sizes of the abnormal overlap indexes of the center points of the two clustering clusters are compared, the position region corresponding to the cluster with the maximum abnormal overlap index is marked as a high-risk defect region, and the position region corresponding to the cluster with the minimum abnormal overlap index is marked as a low-risk defect region; Based on the set of defect position coordinates, the high-risk defect area and the low-risk defect area, an anomaly analysis result is obtained, and the anomaly analysis result includes: a first analysis result, a second analysis result, a third analysis result and a fourth analysis result; wherein the first analysis result represents a joint defect, the second analysis result represents a coating coverage defect, the third analysis result represents an edge profile defect, and the fourth analysis result represents a curved surface transition defect; The following is an explanation of the terms involved: Joint defect: abnormality caused by poor adhesion or processing error at the joint of the light box cloth; for example: joint misalignment, adhesive failure and light leakage; Coating coverage defect: surface anomaly caused by uneven coating or process defect; for example: pinholes, bubbles and impurity particles; Edge profile defect: edge anomaly caused by poor cutting or edging process; for example: burrs, edge warping and coating accumulation; Curved surface transition defect: coating or substrate failure caused by stress concentration in the curved area; for example: coating cracking, substrate peeling and wrinkling, etc. For the high-risk defect area, the more obvious the PVC light box cloth surface defect is, the higher the abnormal overlap index increases; in addition, the curvature gradient coefficient and the curve gradient coefficient also increase accordingly; thus, the curvature gradient coefficient, the curve gradient coefficient and the abnormal overlap index of the high-risk area and the low-risk area of the training data set are obtained; in the early stage of the convolutional neural network, the smaller the weight coefficient is, the better the effect of distinguishing the high-risk defect area and the low-risk defect area is; Fusion of anomaly analysis results, input to the calculation layer, obtain updated weight coefficient, including: Updated weight coefficient: ; In the formula, α represents the updated weight coefficient, μ represents the basic coefficient, and the value range is 0.05, which is used to avoid the updated weight coefficient being too small, ε represents the control coefficient, and the value range is 0.1, which is used to control the value range of the updated weight coefficient, exp represents the exponential function with natural constant as the base number, K1 represents the position set corresponding to the high-risk defect area in the training set, K2 represents the position set corresponding to the low-risk defect area in the training set, and δ(G) represents the mean value of all abnormal overlap indexes in all high-risk defect areas. Further, update the loss function of the defect analysis model based on the updated weight coefficient to recalculate the loss and perform forward propagation to update the model parameters, the specific steps are known and will not be described. Specifically, the defect analysis model is based on the CNN neural network architecture (convolutional neural network). The input to the computation layer is the fused anomaly analysis results [F1, F2, F3, F4], including the first analysis result F1, the second analysis result F2, the third analysis result F3, and the fourth analysis result F4, which correspond to seam defects, coating coverage defects, edge contour defects, and curved surface transition defects, respectively. An ADAM optimizer (Adaptive Moment Estimator) of size 32 is used with a learning rate of 0.001 to balance convergence speed and stability and avoid gradient oscillations. Training is then performed by substituting the updated weight coefficients α into the loss function of the calculated layer as the loss function for the CNN neural network. Simultaneously, the updated weight coefficients α are varied with the number of iterations to balance the model's ability to distinguish between broad categories and finer-grained categories. Based on the calculated updated weight coefficients, model training begins. In the first 12 training rounds, the main focus is on optimizing the large-class loss function to distinguish between high-risk and low-risk defect regions, while keeping the weight coefficient α constant. In the subsequent training rounds, the weight of the small-class loss function is gradually increased until it approaches 0.5. This is to ensure that the defect classification model can also learn the differences between subclasses while maintaining partial error in the large-class classification. The CNN method is a well-known technique, and the specific process will not be elaborated here. The output is processed through the output layer, and the output results can include major categories and subcategories, such as: high-risk defect areas - seam defects - light leakage, low-risk defect areas - coating coverage defects - coating pinholes, etc. This completes the identification of defect categories on the surface of the PVC lightbox fabric.

[0018] Controlled output module: Used to adjust and repair surface defects of PVC lightbox fabric based on surface defect category areas and according to preset adjustment strategies; High-risk defect areas refer to areas with significant defects that may affect functionality or appearance, requiring immediate repair, and an "immediate repair instruction" is issued. Low-risk defect areas refer to areas with minor defects that meet process tolerances, allowing for delayed repair, and a "delayed repair instruction" is issued. Professional personnel or related equipment (e.g., PVC lightbox fabric high-speed production line) take different actions based on different instructions to ensure full-parameter repair of high-risk areas, while low-risk areas only require monitoring and predictive maintenance. This achieves a real-time closed-loop feedback mechanism on the high-speed production line, improving detection efficiency and reducing surface defects in PVC lightbox fabric.

[0019] Example 2: This invention provides an online detection method for surface defects of PVC lightbox fabric based on machine vision, including the following steps: The three-dimensional surface information of the PVC light box cloth is collected, wherein the three-dimensional surface information is obtained based on a three-dimensional surface image scanned and acquired from the PVC light box cloth; The three-dimensional surface information is introduced into a preset defect analysis model, the defect analysis model is established based on a CNN architecture, a loss function of the defect analysis model is updated, and a surface defect category area of the PVC light box cloth is output based on the loss function and the defect analysis model, wherein the surface defect category area includes a high-risk defect area and a low-risk defect area. Based on the surface defect category area, the surface defects of the PVC light box cloth are adjusted and repaired according to a preset adjustment strategy.

[0020] In the application, the several formulas involved are all dimensionless values, and the formulas are obtained by software simulation based on a large amount of data to obtain a formula of the nearest real situation, and the parameters in the formula are set by a person skilled in the art according to the actual situation.

[0021] The above embodiments can be realized wholly or partially by software, hardware, firmware or any other combination. When realized by software, the above embodiments can be realized in the form of a computer program product wholly or partially. A person of ordinary skill in the art can realize that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized by hardware or software depends on the specific application and design constraints of the technical solutions.

[0022] The units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, which can be located in one place or distributed on multiple network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiments.

[0023] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the present application, which should be covered within the protection scope of the present application.

Claims

1. A machine vision-based on-line detection system for surface defects of PVC light box cloth, characterized in that, The system includes: The information acquisition module is used to acquire the three-dimensional surface information of PVC lightbox fabric, wherein the three-dimensional surface information is extracted based on the three-dimensional surface image obtained by scanning the PVC lightbox fabric; The model training module is used to import the three-dimensional surface information into a preset defect analysis model. The defect analysis model is built based on a CNN architecture. The module updates the loss function of the defect analysis model and outputs the surface defect category regions of the PVC lightbox fabric based on the loss function and the defect analysis model. The surface defect category regions include high-risk defect regions and low-risk defect regions. The control output module is used to adjust and repair the surface defects of the PVC lightbox fabric according to a preset adjustment strategy based on the surface defect category area. 2.The machine vision based PVC lightbox cloth surface defect online detection system according to claim 1, characterized in that, The step of importing the three-dimensional surface information into a preset defect analysis model includes: The pre-defined defect analysis model includes at least an input layer, an identification layer, a labeling layer, and a calculation layer; The three-dimensional surface information is input into the input layer, and the three-dimensional surface features are extracted by the recognition layer. The three-dimensional surface features are spatially labeled by the labeling layer to obtain surface label features. The surface label features are input into the analysis layer to obtain anomaly analysis results, including the first analysis result, the second analysis result, the third analysis result, and the fourth analysis result. The anomaly analysis results are fused and input into the calculation layer to obtain updated weight coefficients. The loss function of the defect analysis model is updated based on the updated weight coefficients, and the defect category region of the PVC lightbox fabric surface is output based on the updated defect analysis model. 3.The machine vision based PVC lightbox cloth surface defect online detection system according to claim 2, characterized in that, The extraction of three-dimensional surface features through the recognition layer includes: The three-dimensional surface features are screened to remove feature points outside the key areas; the key areas include: seams, coating areas, edge contours and curved transition areas. The sieved feature points are classified and summarized to obtain the seam point set, coating coverage point set, edge contour point set, and curved surface transition point set; Based on preset direction, preset interval and feature point location information, several types of line segments are used to connect the classified feature points, and the distribution characteristics of the line segments are analyzed. The analysis results constitute three-dimensional surface features. Among them, the line segments include at least one type of straight line and curve, and the distribution characteristics of straight line are the slope of the straight line, and the distribution characteristics of curve are the slope, curvature and radius of curvature of the curve. 4.The machine vision based online detection system for surface defects of PVC light box cloth according to claim 2, characterized in that, The step of inputting surface annotation features into the analysis layer to obtain anomaly analysis results includes: The analysis layer includes primary analysis and secondary analysis; One-step analysis: The three-dimensional surface features under several line segments are standardized and weighted summation is performed to obtain the surface parameter values; The surface parameter values ​​were compared and analyzed with the corresponding preset adaptation standard values: When the surface parameter value is less than the adaptation standard value, it indicates that there are no obvious defects or anomalies. If the surface parameter value is greater than or equal to the adaptation standard value, it indicates that there is an obvious defect or anomaly, and a secondary analysis is required. Secondary analysis: The preset light source is controlled to move based on a detection device, and different waveband reflection information of the surface of the PVC light box cloth is extracted based on an irradiation angle, an intensity, and a reflection line trajectory of the preset light source. The reflection information includes ultraviolet waveband, visible light waveband, and near-infrared light waveband. Thus, a first reflectivity corresponding to the ultraviolet waveband, a second reflectivity corresponding to the visible light waveband, and a third reflectivity corresponding to the near-infrared light waveband of the PVC light box cloth are obtained respectively. A rectangular coordinate system is established, the first reflectivity / second reflectivity is taken as an X-axis, and the first reflectivity / third reflectivity is taken as a Y-axis. A trajectory change curve about the waveband is drawn, and a curve mutation coefficient and a curvature gradient coefficient of each key area are determined based on the trajectory change curve. Based on the curve mutation coefficient and the curvature gradient coefficient, abnormal sequences of each key area are identified, including a first abnormal sequence and a second abnormal sequence, and a defect position coordinate set is generated. 5.The machine vision-based PVC lightbox cloth surface defect online detection system according to claim 4, characterized in that, The determination of the curve mutation coefficient and the curvature gradient coefficient of each key area based on the trajectory change curve includes: a preset time sliding window; a peak point set and a valley point set are determined based on the trajectory change curve, a difference absolute value between the peak point set and the valley point set is marked as a span value, a standard deviation of each trajectory point in the time sliding window is obtained, a product of the standard deviation of each trajectory point and the corresponding span value is calculated, and the curve mutation coefficient is obtained; a slope set is determined based on the trajectory change curve, a fluctuation value of the slope set in the time sliding window and a slope standard deviation of each trajectory point are obtained, a square of a difference between the slope standard deviation and the fluctuation value is calculated, and the curvature gradient coefficient is obtained. 6.The machine vision based online detection system for surface defects of PVC light box cloth according to claim 4, characterized in that, The identification of the abnormal sequences of each key area based on the curve mutation coefficient and the curvature gradient coefficient includes: the curve mutation coefficient and the curvature gradient coefficient are sorted in descending order respectively, the first n corresponding curve mutation coefficients and curvature gradient coefficients are selected, and the first abnormal sequence and the second abnormal sequence are formed by mapping the selected curve mutation coefficients and curvature gradient coefficients to the trajectory change curve in order. a Pearson correlation coefficient of the first abnormal sequence and the second abnormal sequence is calculated, a Shannon entropy of all the curve mutation coefficients in the first abnormal sequence is calculated, a ratio of an absolute value of the Pearson correlation coefficient to the Shannon entropy is calculated, and an abnormal overlap index is obtained; the abnormal overlap index is taken as an input of a clustering algorithm, two clustering clusters are obtained, and a high-risk defect area and a low-risk defect area are divided; based on the defect position coordinate set, the high-risk defect area, and the low-risk defect area, an abnormal analysis result is obtained, including a first analysis result, a second analysis result, a third analysis result, and a fourth analysis result; the first analysis result represents a joint defect, the second analysis result represents a coating coverage defect, the third analysis result represents an edge contour defect, and the fourth analysis result represents a curved surface transition defect. 7.The machine vision based online detection system for surface defects of PVC display banner according to claim 2, wherein, The fusion of the abnormal analysis result includes a series fusion of the first analysis result, the second analysis result, the third analysis result, and the fourth analysis result. 8.The machine vision based online detection system for surface defects of PVC display cases according to claim 2, wherein, The updating of the weight coefficient is calculated according to the following formula: ； In the formula, α represents an update weight coefficient, μ represents a basic coefficient, ε represents a control coefficient, K1 represents a position set corresponding to a high-risk defect area in a training set, K2 represents a position set corresponding to a low-risk defect area in the training set, and δ(G) represents a mean value of all abnormal overlapping indexes in all high-risk defect areas. 9.The machine vision based online detection system for surface defects of PVC display banner according to claim 1, wherein, The surface defect of the PVC light box cloth is adjusted and repaired according to a preset adjustment strategy, including: immediately repairing the high-risk defect area; temporarily repairing the low-risk defect area.

10. A method for on-line detection of surface defects of PVC light box cloth based on machine vision, characterized in that, The method comprises the following steps: acquiring three-dimensional surface information of the PVC light box cloth, wherein the three-dimensional surface information is obtained based on a three-dimensional surface image obtained by scanning the PVC light box cloth; inputting the three-dimensional surface information into a preset defect analysis model, updating a loss function of the defect analysis model, and outputting a surface defect category area of the PVC light box cloth based on the loss function and the defect analysis model, wherein the surface defect category area comprises a high-risk defect area and a low-risk defect area; adjusting and repairing the surface defect of the PVC light box cloth according to a preset adjustment strategy based on the surface defect category area.

Images