Machine vision guided corrugated board defect real-time detection method and system

Abstract

The invention discloses a machine vision guided corrugated board defect real-time detection method and system, and relates to the technical field of corrugated board detection.The method comprises the steps that a gray level image of a target corrugated board production line is collected in real time; obtaining a dynamic statistical feature list, performing statistical feature extraction on the corresponding traversal gray level image, and constructing a primary feature vector according to a statistical feature extraction result; constructing a dynamic threshold model based on historical defect detection records, and performing primary screening on the corrugated boards by taking the dynamic threshold model as a constraint and combining the primary feature vectors; and according to a preliminary screening result, correspondingly extracting a grayscale image, carrying out adaptive grayscale enhancement, obtaining an enhanced grayscale image, and carrying out defect detection based on the enhanced grayscale image. The technical problems of poor adaptability and insufficient accuracy of corrugated board defect detection in the prior art are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of equipment state monitoring, in particular to a machine vision guided corrugated board defect real-time detection method and system. BACKGROUND

[0002] In the field of corrugated board production and manufacturing, product quality control is of great importance, and defect detection is a key link to ensure the qualification of products leaving the factory. In the prior art, machine vision has been widely used in corrugated board surface defect detection, which mainly relies on a static detection model with a preset fixed threshold. However, due to factors such as the variety of corrugated board product specifications, production line speed changes, and environmental light fluctuations, the fixed threshold is difficult to adapt to different working conditions, resulting in high false rejection rate and high false rejection rate; at the same time, the static model lacks the ability to perceive and learn real-time data in the production process, resulting in poor detection accuracy, insufficient defect recognition ability, and high computational complexity, which is difficult to meet the detection requirements of corrugated board production. SUMMARY

[0003] The present application provides a machine vision guided corrugated board defect real-time detection method and system, which is used to solve the technical problems of poor adaptability and insufficient accuracy of corrugated board defect detection in the prior art.

[0004] In view of the above problems, the present application provides a machine vision guided corrugated board defect real-time detection method and system.

[0005] In a first aspect, the present application provides a machine vision guided corrugated board defect real-time detection method, which comprises: real-time acquisition of a gray image of a target corrugated board production line; obtaining a dynamic statistical feature list, extracting statistical features corresponding to traversing the gray image, and constructing a primary feature vector according to the statistical feature extraction result; constructing a dynamic threshold model based on historical defect detection records, and screening corrugated board in combination with the primary feature vector based on the dynamic threshold model as a constraint; According to the preliminary screening result, the gray image is adaptively enhanced to obtain an enhanced gray image, and defect detection is performed based on the enhanced gray image.

[0006] In a second aspect, the present application provides a machine vision guided corrugated board defect real-time detection system, comprising: an image acquisition module for real-time acquisition of a gray image of a target corrugated board production line; a feature extraction module for obtaining a dynamic statistical feature list, extracting statistical features corresponding to traversing the gray image, and constructing a primary feature vector according to the statistical feature extraction result; The feature preliminary screening module is configured to construct a dynamic threshold model based on historical defect detection records, and to perform preliminary screening on the corrugated board based on the dynamic threshold model and the preliminary feature vector. The defect detection module is configured to perform adaptive gray scale enhancement on the gray scale image according to the preliminary screening result, to obtain an enhanced gray scale image, and to perform defect detection based on the enhanced gray scale image.

[0007] The one or more technical solutions provided in the present application have at least the following technical effects or advantages: The present application provides a machine vision guided real-time corrugated board defect detection method and system, which combines dynamic statistical feature extraction and adaptive threshold adjustment, significantly improving the accuracy and adaptability of corrugated board defect detection. Compared with traditional methods, the technical solutions provided by the present application significantly overcome the limitations of traditional methods in dealing with different product specifications and changes in production environment, achieving the technical effect of efficient and accurate detection of corrugated board defects in complex production environments. BRIEF DESCRIPTION OF DRAWINGS

[0008] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0009] Figure 1 A flowchart of a machine vision guided real-time corrugated board defect detection method provided by an embodiment of the present application.

[0010] Figure 2 A structure diagram of a machine vision guided real-time corrugated board defect detection system provided by an embodiment of the present application.

[0011] In the drawings, the components represented by the numbers are described as follows: Image acquisition module 100, feature extraction module 200, feature preliminary screening module 300, defect detection module 400. DETAILED DESCRIPTION

[0012] The present application provides a machine vision guided real-time corrugated board defect detection method and system, which is used to solve the technical problems of poor adaptability and insufficient accuracy of corrugated board defect detection in the prior art.

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0014] It should be noted that the terms "comprising" and "having" are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to these processes, methods, products, or devices.

[0015] Example 1, as Figure 1 As shown, this application provides a machine vision-guided real-time defect detection method for corrugated cardboard, wherein the method includes: S10: Real-time acquisition of grayscale images of the target corrugated cardboard production line.

[0016] In this embodiment, an industrial camera is used to acquire grayscale images of the target corrugated cardboard production line in real time. The target corrugated cardboard production line refers to the production line that produces the corrugated cardboard to be inspected. A grayscale image is an image where each pixel has only one sampled color; such images are typically displayed as grayscale ranging from the darkest black to the brightest white. Unlike black and white images, grayscale images have many levels of color depth between black and white, which can preserve image details while simplifying the matrix and reducing computational load.

[0017] S20: Obtain a dynamic statistical feature list, traverse the grayscale image to extract statistical features, and construct a primary feature vector based on the statistical feature extraction results.

[0018] Traditional methods often employ fixed feature extraction schemes, which cannot adapt to changes in different product specifications and production conditions. This results in features that lack specificity and are computationally redundant, affecting both detection efficiency and the inability to accurately capture real defect features, thus hindering the improvement of detection accuracy.

[0019] Step S20 in the method provided in this application embodiment includes: Obtain product specification information for the target corrugated cardboard production line; Based on the product specification information, retrieve the historical defect detection image set; Based on the preset prior full statistical feature list, the historical defect detection image set is traversed to perform full feature extraction and obtain the historical defect feature set; Input the historical defect feature set as verification data into the full-inspection defect detector, obtain a verification result, and correspondingly extract a detection contribution rate of each statistical feature to the full-inspection defect detector; Based on the detection contribution rate, perform cumulative calculation, extract the top N statistical features that meet a preset contribution rate threshold, and output the dynamic statistical feature list; Traverse the gray-scale image based on the dynamic statistical feature list to perform feature extraction, and obtain a statistical feature extraction result; Perform non-dimensionalization processing on the statistical feature extraction result, and perform mean normalization based on a non-defective product to obtain a standard feature set; Extract a contribution rate list corresponding to the dynamic statistical feature list, and correspondingly normalize to obtain a feature weight list; According to the feature weight list and the standard feature set, construct the primary feature vector of the gray-scale image.

[0020] In the embodiment of the application, product specification information of the corrugated board currently produced is read from the production management system, including paperboard type, flute type, grammage, etc.

[0021] According to the product specification information, all corrugated board images of the same specification that have been labeled with defects are retrieved from the historical database to obtain a historical defect detection image set, wherein the historical defect detection image set contains corrugated board images with defects and defect labels, such as labels indicating the presence of collapsed flutes and rough surfaces.

[0022] According to a preset prior full statistical feature list, full feature extraction is performed on the historical defect detection image set to obtain a historical defect feature set. The prior full statistical feature list refers to all possible features that may contribute to defect recognition, such as various texture features, gray-scale statistical features, geometric shape features, etc.

[0023] A full-inspection defect detector is obtained, which is a detector model facing the current factory with perfect detection capability. The historical defect feature set is input into the full-inspection defect detector as verification data, a verification result is obtained, and the detection contribution rate of each statistical feature to the full-inspection defect detector is correspondingly extracted. The detection contribution rate refers to the contribution degree of each statistical feature to the final detection result, for example, the contribution degree of the texture feature to the surface roughness defect may be large, and the contribution degree of the geometric shape feature to the surface roughness defect may be small. Exemplarily, the principal component analysis method can be used to extract the detection contribution rate, for example, the principal component contribution rate of each feature item is calculated to obtain the detection contribution rate. Through this process, the relative importance and contribution degree of each statistical feature in defect detection can be quantitatively evaluated, providing a basis for subsequent feature screening.

[0024] The cumulative calculation is performed based on the detection contribution rate, the first N statistical features meeting the preset contribution rate threshold are extracted, and the output is a dynamic statistical feature list. The contribution rate threshold is a threshold set to filter out statistical features with high contribution rates and prevent statistical features with low contribution rates from wasting computing power. It can be set to 0.6, for example.

[0025] With the dynamic statistical feature list as the target, the gray-scale image is traversed for feature extraction, and a statistical feature extraction result is obtained. Specifically, based on a convolutional neural network, a gray-scale image feature extractor is constructed, which can have a 4-layer structure, for example. The input layer is used to receive the gray-scale image, the convolution layer uses 64 3x3 convolution kernels and uses a ReLU function for activation, the pooling layer uses 2x2 max pooling, and the output layer is used to output the extracted gray-scale image features. The labeled historical gray-scale images are used for training, and the labeling content is the features of the gray-scale images. The historical gray-scale images are input into the constructed gray-scale feature extractor, and the predicted error is calculated by the cross-entropy loss function. This error quantifies the inaccuracy under the current model parameters. Using the backpropagation algorithm, the gradient of each parameter in the model is automatically calculated. The gradient indicates the direction and magnitude of each parameter that should be adjusted to reduce the loss. Using the Adam optimizer, all parameters in the network are iteratively updated based on the calculated gradient. Until the loss value no longer changes significantly and the model accuracy tends to be stable, for example, the accuracy reaches 90%, the gray-scale image feature extractor is trained. The gray-scale image feature extractor is used to traverse the gray-scale image for feature extraction with the dynamic statistical feature list as the target, and a feature extraction result is obtained.

[0026] The statistical feature extraction result is dimensionless, and the mean normalization based on the defect-free product is performed to obtain a standard feature set. For example, the Z-score standardization algorithm is used for dimensionless processing and normalization, and the standard feature value = (feature value - mean of all historical values of this feature) ÷ mean of all historical values of this feature.

[0027] The contribution rate list corresponding to the dynamic statistical feature list is extracted and normalized to obtain a feature weight list. The feature weight = contribution rate ÷ sum of all feature contribution rates, and the sum of all feature weights calculated is 1.

[0028] According to the feature weight list and the standard feature set, a primary feature vector of the gray-scale image is constructed.

[0029] The dynamic statistical feature list guides feature extraction, and the primary feature vector is constructed, realizing the intelligentization and precision of feature extraction. It can adaptively select the most effective feature extraction scheme according to the actual production situation, significantly improve the representativeness and computational efficiency of the features, and provide high-quality feature input for subsequent defect recognition.

[0030] S30: constructing a dynamic threshold model based on the historical defect detection records, and performing preliminary screening on the corrugated board in combination with the primary feature vector based on the dynamic threshold model as a constraint.

[0031] The traditional defect detection method is difficult to adapt to the dynamic changes in the production process, is prone to misjudgment and omission, and seriously affects the reliability of detection.

[0032] The step S30 in the method provided in the embodiment of the application comprises: Obtaining defect detection records of the corrugated board of the previous N batches, and obtaining the historical defect detection records, wherein N is greater than or equal to 3; According to the dynamic statistical feature list, traversing the historical defect detection records to extract features, and obtaining a historical defect detection feature set; Analyzing the historical defect detection feature set, determining a feature value of each statistical feature satisfying a preset confidence value, and merging and outputting multiple feature values corresponding to the dynamic statistical feature list as the dynamic threshold model; Wherein, analyzing the historical defect detection feature set, determining a feature value of each statistical feature satisfying a preset confidence value, and merging and outputting multiple feature values corresponding to the dynamic statistical feature list as the dynamic threshold model, comprises: Randomly selecting any statistical feature as a first analysis feature based on the dynamic statistical feature list; Extracting a feature value corresponding to the first analysis feature from the historical defect detection feature set to obtain a first analysis feature value set; Serializing the first analysis feature value set, and determining a feature value at a quantile corresponding to a preset confidence value to obtain a first dynamic threshold; Traversing the dynamic statistical feature list to obtain multiple dynamic thresholds, and merging and outputting as the dynamic threshold model.

[0033] In the embodiment of the application, the defect detection records of the corrugated board of the previous N batches are obtained, the historical defect detection records are obtained according to the defect detection records of the previous N batches, wherein N is greater than or equal to 3. Obtaining more defect information of the recent batches is more conducive to extracting commonalities and providing a data basis for subsequent defect analysis.

[0034] According to the dynamic statistical feature list, traversing the historical defect detection records to extract features, and obtaining a historical defect detection feature set. Exemplarily, the historical defect detection feature set comprises one or more of image gray average value, image gray value standard deviation, image contrast and the like.

[0035] The historical defect detection feature set is analyzed to determine a feature value that meets a preset confidence value for each statistical feature, and a plurality of feature values corresponding to the dynamic statistical feature list are merged and output as a dynamic threshold model. The feature value that meets the preset confidence value is used as the dynamic threshold, which can dynamically adjust the judgment standard according to the actual production data, avoid the problem of missed detection caused by a fixed threshold that is too loose, prevent the misjudgment phenomenon caused by a threshold that is too strict, and significantly improve the adaptability and reliability of the defect detection system under different production batches and process parameter changes.

[0036] Specifically, based on the dynamic statistical feature list, any statistical feature is randomly selected as a first analysis feature, which is the current processing object.

[0037] From the historical defect detection feature set, the feature value corresponding to the first analysis feature is extracted to obtain a first analysis feature value set.

[0038] The first analysis feature value set is serialized, and a feature value at a quantile corresponding to a preset confidence value is determined to obtain a first dynamic threshold. For example, the first analysis feature is the image gray value standard deviation, a plurality of feature values of the image gray value standard deviation are extracted from the historical defect detection feature set, and the plurality of feature values are integrated into a set as the first analysis feature value set. The plurality of feature values in the first analysis feature value set are sorted from small to large, for example, there are 10 feature values, and the sorting is “0.1, 0.12, 0.2, 0.3, 0.45, 0.5, 0.52, 0.55, 0.61, 0.8”. When the preset confidence is 0.8, the feature value at the corresponding quantile is 0.55, and the first dynamic threshold is 0.55. This threshold value represents that 80% of the gray value standard deviations in the first analysis feature value set are less than 0.55.

[0039] The dynamic statistical feature list is traversed to obtain a plurality of dynamic thresholds, which are merged and output as a dynamic threshold model. The dynamic threshold model sets a dynamic threshold value based on historical defect data statistical experience for each key feature, and the threshold value set in the model can be used as a standard for rapid preliminary screening of data.

[0040] And with the dynamic threshold model as a constraint, combined with the primary feature vector, the corrugated board is preliminarily screened to obtain a preliminary screening result.

[0041] By constructing a dynamic threshold model based on historical data and combining a primary feature vector for preliminary screening, adaptive optimization of the defect judgment standard is realized, which enables the defect detection to dynamically adjust the judgment standard according to the actual production situation, significantly improves the accuracy and adaptability of the preliminary screening, and provides a reliable preprocessing result for subsequent fine detection.

[0042] S40: performing adaptive gray scale enhancement on the gray scale images corresponding to the screening results to obtain enhanced gray scale images, and performing defect detection based on the enhanced gray scale images.

[0043] After screening, how to accurately identify suspected defects is the final challenge. Traditional methods cannot optimize the processing of defects with different characteristics, which easily leads to difficulty in accurately detecting weak defects or special defects, and high computational complexity, resulting in low detection efficiency.

[0044] The step S40 in the method provided in the embodiments of the present application includes: According to the screening results, the gray scale images are extracted to perform adaptive gray scale enhancement to obtain enhanced gray scale images, and defect detection is performed based on the enhanced gray scale images, which includes: Performing image segmentation based on the historical defect detection image set to obtain a historical defect image set; Performing image gray scale value-oriented statistical analysis on the historical defect image set to obtain defect image gray scale distribution characteristics, and performing function fitting based on the defect image gray scale distribution characteristics to output a function fitting result as an interval gray scale enhancement function; Combining the interval gray scale enhancement function, an adaptive gray scale enhancement function is constructed, wherein the adaptive gray scale enhancement function is a piecewise function; According to the screening results, the gray scale images are extracted to perform adaptive gray scale enhancement to obtain enhanced gray scale images, and defect detection is performed based on the enhanced gray scale images, which includes: Using the screening results as an index, the gray scale images are extracted to output a set of images to be detected; According to the adaptive gray scale enhancement function, the set of images to be detected is subjected to gray scale enhancement processing to generate the enhanced gray scale images; Based on a knowledge distillation method, the historical defect detection feature set and the historical defect detection record are combined to perform model compression on the full detection defect detector to obtain a light defect detector, and the enhanced gray scale images are input into the light defect detector for defect detection to obtain a defect detection result; The defect detection result at least includes a defect position and a defect category.

[0045] In the embodiments of the present application, image segmentation is performed based on a historical defect detection image set to obtain a historical defect image set. For example, a gray scale threshold-based segmentation method is used to separate the defect area from the background in the image to obtain the historical defect image set, which contains images of various defects.

[0046] The gray scale values of all defect pixels are statistically analyzed by traversing the historical defect image set, for example, a gray scale histogram is drawn to obtain the gray scale distribution characteristics of the defect image. Through statistical analysis, the gray scale values of the defect region are usually concentrated in a certain interval. Based on the gray scale distribution characteristics of the defect image, an interval gray scale enhancement function is constructed by using a function fitting method such as using a piecewise linear function. The core function of the function is to stretch the high contrast for the gray scale interval in the defect feature set, and to compress the non-key interval, so as to highlight the defects.

[0047] Combined with the interval gray scale enhancement function, an adaptive gray scale enhancement function is constructed. The adaptive gray scale enhancement function is a piecewise function, and its basic logic is: if the gray scale value of the input pixel is in the defect feature interval, the interval gray scale enhancement function is applied for significant enhancement; if it is outside the interval, a gentle mapping function is applied to maintain the overall naturalness of the image. That is, a function that can adaptively and selectively enhance the defect region is obtained.

[0048] The gray scale image is extracted based on the preliminary screening result as the index, and the image set to be detected is output.

[0049] According to the adaptive gray scale enhancement function, the gray scale enhancement processing is performed on the image set to be detected to generate an enhanced gray scale image.

[0050] Based on the knowledge distillation method, the historical defect detection feature set and the historical defect detection record are combined to compress the full detection defect detector to obtain a light defect detector. Specifically, the full detection defect detector with complex structure but high precision is used as a teacher model, a network with lighter structure such as a 3-layer structure is constructed, wherein the input layer is used to receive the gray scale image, the convolutional layer contains 32 3x3 convolutional kernels, and the output layer is used to output the convolutional neural network of the defect detection result as the light defect detector as the student model. Using the historical defect detection feature set and the historical defect detection record, the light defect detector not only learns how to predict the real defect label, but also learns to imitate the probability distribution output by the full detection defect detector. This probability distribution contains rich knowledge of the teacher model, so that the light defect detector can still maintain high precision after compression. The mean square error is used as the loss function for training, and the Adam optimizer is used.

[0051] The enhanced gray scale image is input into the light defect detector for defect detection to obtain a defect detection result. The defect detection result at least includes the defect position and the defect category.

[0052] Through adaptive gray scale enhancement and defect detection based on the enhanced image, accurate identification of suspected defects is realized, individualized image enhancement processing can be performed for different defect characteristics, the detection capability for various defects is significantly improved, the detection efficiency is improved on the premise of ensuring the detection accuracy, and finally efficient and accurate defect detection is realized.

[0053] Example 2, as Figure 2 As shown, based on the same inventive concept as the machine vision-guided real-time defect detection method for corrugated cardboard provided in Embodiment 1, this embodiment of the invention also provides a machine vision-guided real-time defect detection system for corrugated cardboard, comprising: Image acquisition module 100 is used to acquire grayscale images of the target corrugated cardboard production line in real time; The feature extraction module 200 is used to obtain a dynamic statistical feature list, perform statistical feature extraction by traversing the grayscale image accordingly, and construct a primary feature vector based on the statistical feature extraction results. The feature screening module 300 is used to construct a dynamic threshold model based on historical defect detection records, and to perform initial screening of corrugated cardboard by combining the dynamic threshold model with the primary feature vector; The defect detection module 400 is used to extract the grayscale image according to the initial screening result, perform adaptive grayscale enhancement, obtain an enhanced grayscale image, and perform defect detection based on the enhanced grayscale image.

[0054] In one embodiment, the feature extraction module 200 is further configured to: Obtain product specification information for the target corrugated cardboard production line; Based on the product specification information, retrieve the historical defect detection image set; Based on the preset prior full statistical feature list, the historical defect detection image set is traversed to perform full feature extraction and obtain the historical defect feature set; The historical defect feature set is used as verification data and input into the full inspection defect detector to obtain the verification results. The contribution rate of each statistical feature to the detection of the full inspection defect detector is extracted accordingly. Based on the detection contribution rate, cumulative calculation is performed, and the top N statistical features that meet the preset contribution rate threshold are extracted and output as the dynamic statistical feature list. Using the dynamic statistical feature list as the target, the grayscale image is traversed to extract features and obtain the statistical feature extraction results; The statistical feature extraction results are dimensionless and normalized based on the mean of defect-free products to obtain a standard feature set. Extract the contribution rate list corresponding to the dynamic statistical feature list, and normalize it accordingly to obtain the feature weight list; Based on the feature weight list and the standard feature set, the primary feature vector of the grayscale image is constructed.

[0055] In one embodiment, the feature screening module 300 is further configured to: Obtaining defect detection records of the previous N batches of corrugated board, obtaining the historical defect detection records, wherein N is greater than or equal to 3; According to the dynamic statistical feature list, traversing the historical defect detection records for feature extraction, and obtaining a historical defect detection feature set; Analyzing the historical defect detection feature set, determining a feature value of each statistical feature satisfying a preset confidence value, and merging and outputting multiple feature values corresponding to the dynamic statistical feature list as the dynamic threshold model; Wherein, analyzing the historical defect detection feature set, determining a feature value of each statistical feature satisfying a preset confidence value, and merging and outputting multiple feature values corresponding to the dynamic statistical feature list as the dynamic threshold model, comprising: Randomly selecting any statistical feature as a first analysis feature based on the dynamic statistical feature list; Extracting the feature value corresponding to the first analysis feature from the historical defect detection feature set, and obtaining a first analysis feature value set; Serializing the first analysis feature value set, and determining a feature value at a quantile corresponding to a preset confidence value, to obtain a first dynamic threshold; Traversing the dynamic statistical feature list to obtain multiple dynamic thresholds, and merging and outputting as the dynamic threshold model.

[0056] In one embodiment, the defect detection module 400 is further configured to: According to the preliminary screening result, extracting the gray image for adaptive gray scale enhancement, obtaining an enhanced gray image, and performing defect detection based on the enhanced gray image, and previously comprising: Based on the historical defect image set, image segmentation is performed to obtain a historical defect image set; Traversing the historical defect image set for image gray value-oriented statistical analysis, obtaining defect image gray distribution features, and based on the defect image gray distribution features, function fitting is performed to output a function fitting result as an interval gray scale enhancement function; Combined with the interval gray scale enhancement function, an adaptive gray scale enhancement function is constructed, wherein the adaptive gray scale enhancement function is a piecewise function; According to the preliminary screening result, extracting the gray image for adaptive gray scale enhancement, obtaining an enhanced gray image, and performing defect detection based on the enhanced gray image, comprising: Taking the preliminary screening result as an index, the gray image is extracted to output a set of images to be detected; According to the adaptive gray scale enhancement function, the set of images to be detected is subjected to gray scale enhancement processing to generate the enhanced gray image; Based on a knowledge distillation method, the historical defect detection feature set and the historical defect detection record are combined to perform model compression on the full defect detector, to obtain a light defect detector, and the enhanced gray image is input into the light defect detector for defect detection, to obtain a defect detection result. The defect detection result at least includes a defect position and a defect category.

[0057] In summary, the embodiments of the present application have at least the following technical effects: The present application provides a machine vision guided corrugated board defect real-time detection method and system, which significantly improves the accuracy and adaptability of corrugated board defect detection by combining dynamic statistical feature extraction and adaptive threshold adjustment. Specifically, by constructing a dynamic statistical feature list and a dynamic threshold model based on historical data, adaptive adjustment of production conditions is realized, effectively reducing the false detection rate and the missed detection rate. By introducing an adaptive gray scale enhancement mechanism, precise image enhancement processing can be performed on suspected defect areas, improving the recognition ability of weak defects and variant defects. At the same time, the lightweight defect detector realized by using the knowledge distillation technology significantly improves the detection speed while ensuring the detection accuracy. Compared with the traditional method, the technical scheme provided by the present application significantly overcomes the limitations of the traditional method in dealing with different product specifications and changes in production environment, and achieves the technical effect of efficient and accurate detection of corrugated board defects in complex production environment.

[0058] It should be noted that the above-mentioned sequence of the embodiments of the present application is only for description, and does not represent the advantages and disadvantages of the embodiments. Moreover, the above describes specific embodiments of the present application. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multi-task processing and parallel processing are possible or may be advantageous.

[0059] The above only describes the preferred embodiments of the present application, and does not limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

[0060] The present application is only an exemplary description of the present application, and should be considered to cover any and all modifications, changes, combinations or equivalents within the scope of the present application. Obviously, those skilled in the art can make various modifications and changes to the present application without departing from the scope of the present application. Thus, if these modifications and changes of the present application belong to the scope of the present application and its equivalent technology, the present application intends to include these modifications and changes.

Claims

1. A machine vision-guided real-time defect detection method for corrugated cardboard, characterized in that, include: Real-time acquisition of grayscale images of the target corrugated cardboard production line; Obtain a dynamic statistical feature list, traverse the grayscale image accordingly to extract statistical features, and construct a primary feature vector based on the statistical feature extraction results; A dynamic threshold model is constructed based on historical defect detection records, and the corrugated cardboard is initially screened using the dynamic threshold model as a constraint and combined with the primary feature vector. Based on the initial screening results, the corresponding grayscale image is extracted and adaptively enhanced to obtain an enhanced grayscale image, and defect detection is performed based on the enhanced grayscale image.

2. The machine vision-guided real-time defect detection method for corrugated cardboard as described in claim 1, characterized in that, Obtain a list of dynamic statistical features, including: Obtain product specification information for the target corrugated cardboard production line; Based on the product specification information, retrieve the historical defect detection image set; Based on the preset prior full statistical feature list, the historical defect detection image set is traversed to perform full feature extraction and obtain the historical defect feature set; The historical defect feature set is used as verification data and input into the full inspection defect detector to obtain the verification results. The contribution rate of each statistical feature to the detection of the full inspection defect detector is extracted accordingly. Based on the detection contribution rate, cumulative calculations are performed to extract the top N statistical features that meet the preset contribution rate threshold, and the output is the dynamic statistical feature list.

3. The machine vision-guided real-time defect detection method for corrugated cardboard as described in claim 2, characterized in that, Statistical features are extracted by traversing the grayscale image, and a primary feature vector is constructed based on the results of the statistical feature extraction, including: Using the dynamic statistical feature list as the target, the grayscale image is traversed to extract features and obtain the statistical feature extraction results; The statistical feature extraction results are dimensionless and normalized based on the mean of defect-free products to obtain a standard feature set. Extract the contribution rate list corresponding to the dynamic statistical feature list, and normalize it accordingly to obtain the feature weight list; Based on the feature weight list and the standard feature set, the primary feature vector of the grayscale image is constructed.

4. The machine vision-guided real-time defect detection method for corrugated cardboard as described in claim 3, characterized in that, A dynamic threshold model is constructed based on historical defect detection records, including: Obtain the defect detection records of the first N batches of corrugated cardboard, and obtain the historical defect detection records, where N is greater than or equal to 3; Based on the dynamic statistical feature list, the historical defect detection records are traversed to extract features and obtain the historical defect detection feature set. Analyze the historical defect detection feature set, determine the feature value of each statistical feature that satisfies the preset confidence value, and merge and output multiple feature values ​​corresponding to the dynamic statistical feature list as the dynamic threshold model.

5. The machine vision-guided real-time defect detection method for corrugated cardboard as described in claim 4, characterized in that, Analyze the historical defect detection feature set, determine the feature value of each statistical feature that satisfies a preset confidence value, and merge and output multiple feature values ​​corresponding to the dynamic statistical feature list as the dynamic threshold model, including: Based on the dynamic statistical feature list, any one of the statistical features is randomly selected as the first analytical feature; Extract the feature value corresponding to the first analysis feature from the historical defect detection feature set to obtain the first analysis feature value set; Serialize the first set of analytical features and determine the feature value at the quantile corresponding to the preset confidence value to obtain the first dynamic threshold; The dynamic statistical feature list is traversed to obtain multiple dynamic thresholds, which are then merged and output as the dynamic threshold model.

6. The machine vision-guided real-time defect detection method for corrugated cardboard as described in claim 5, characterized in that, Based on the initial screening results, the corresponding grayscale image is extracted and adaptively enhanced to obtain an enhanced grayscale image. Defect detection is then performed based on the enhanced grayscale image. Prior to this, the process includes: Image segmentation is performed based on the historical defect detection image set to obtain the historical defect image set; The historical defect image set is traversed to perform statistical analysis on image grayscale values ​​to obtain the grayscale distribution characteristics of the defect images. Based on the grayscale distribution characteristics of the defect images, a function is fitted, and the function fitting result is output as an interval grayscale enhancement function. By combining the aforementioned interval grayscale enhancement function, an adaptive grayscale enhancement function is constructed, wherein the adaptive grayscale enhancement function is a piecewise function.

7. The machine vision-guided real-time defect detection method for corrugated cardboard as described in claim 6, characterized in that, Based on the initial screening results, the corresponding grayscale image is extracted and adaptively enhanced to obtain an enhanced grayscale image. Defect detection is then performed based on the enhanced grayscale image, including: Using the initial screening results as an index, the grayscale image is extracted, and the set of images to be detected is output. The image set to be detected is subjected to grayscale enhancement processing according to the adaptive grayscale enhancement function to generate the enhanced grayscale image; Based on the knowledge distillation method, the model of the full inspection defect detector is compressed by combining the historical defect detection feature set and the historical defect detection record to obtain a minor defect detector. The enhanced grayscale image is then input into the minor defect detector for defect detection to obtain the defect detection result. The defect detection results include at least the defect location and the defect category.

8. A machine vision-guided real-time defect detection system for corrugated cardboard, characterized in that, For implementing a machine vision-guided real-time defect detection method for corrugated cardboard according to any one of claims 1 to 7, the system comprises: The image acquisition module is used to acquire grayscale images of the target corrugated cardboard production line in real time. The feature extraction module is used to obtain a dynamic statistical feature list, perform statistical feature extraction on the corresponding grayscale image, and construct a primary feature vector based on the statistical feature extraction results. The feature screening module is used to construct a dynamic threshold model based on historical defect detection records, and to perform initial screening of corrugated cardboard by combining the dynamic threshold model with the primary feature vector; The defect detection module is used to extract the grayscale image according to the initial screening results, perform adaptive grayscale enhancement to obtain the enhanced grayscale image, and perform defect detection based on the enhanced grayscale image.

Images