Printing and laminating process quality detection method based on image generation

Abstract

The invention relates to the technical field of image processing, and discloses a printing film process quality detection method based on image generation, which comprises the following steps: S1, carrying out multispectral image acquisition on the surface of a printing film to form a multidimensional image data set; s2, performing high-resolution image generation on the multispectral initial image by adopting a super-resolution generative adversarial network to obtain a super-resolution image; and S3, carrying out noise filtering processing on the super-resolution image, eliminating image noise generated in an acquisition process by combining Gaussian filtering and median filtering methods, and carrying out brightness equalization and contrast enhancement. By introducing a multispectral imaging technology and a material adaptive processing model based on deep learning, different types of coating materials can be identified, image processing parameters are automatically adjusted according to optical characteristics of different materials, and the adaptive capacity of the system under various coating materials is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a printing lamination process quality detection method based on image generation. Background Art

[0002] Printing and lamination is one of the important processes to improve the appearance and durability of printed products. In actual production, the stability of the lamination process quality is affected by many factors. Traditional lamination process quality inspection mainly relies on manual visual inspection and image processing technology, which is difficult to meet the needs of efficient and accurate quality inspection in industrial production.

[0003] The existing lamination process quality detection method based on image processing has made technical improvements in some aspects. For example, the comparative document CN118334026B proposes a technical method to improve the detection accuracy through image acquisition, reflection compensation and multi-scale feature extraction in the image-based printing lamination process quality detection method and system. This method has made progress in improving detection efficiency and accuracy, but there are still significant shortcomings:

[0004] The coating quality detection method mentioned in the comparative document mainly focuses on surface reflection compensation and image feature extraction, but its detection adaptability to different materials has not been effectively solved, resulting in significant differences in reflection compensation and feature extraction effects, which makes it impossible to identify defects on some materials;

[0005] The comparative document proposes to improve the detection of image details by extracting multi-scale features, but it does not solve the problem of highly sensitive detection of tiny defects. In actual production, tiny defects have a significant impact on high-end printed products or products with strict quality requirements;

[0006] The comparative document mentions improving detection efficiency through image processing methods, but its methods mainly focus on offline detection and static analysis, and do not introduce real-time detection and production feedback mechanisms. In modern production lines, real-time quality detection and rapid feedback during the production process are crucial and can avoid the production of large-scale defective products. However, the existing technology lacks dynamic detection and real-time adjustment methods for the production process, resulting in quality problems that cannot be corrected in a timely manner.

[0007] Therefore, those skilled in the art provide a printing and lamination process quality detection method based on image generation to solve the problems raised in the above background technology. Summary of the invention

[0008] In view of the deficiencies in the prior art, the present invention provides a printing and lamination process quality detection method based on image generation to solve the problems raised in the above background technology.

[0009] To achieve the above objectives, the present invention is implemented through the following technical solutions: A printing lamination process quality detection method based on image generation, comprising the following steps: S1, collecting multispectral images of the printed lamination surface to form a multi-dimensional image data set;

[0010] S2, using a super-resolution generative adversarial network to generate a high-resolution image for the multi-spectral initial image to obtain a super-resolution image;

[0011] S3, performing noise filtering on the super-resolution image, combining Gaussian filtering and median filtering methods to eliminate image noise generated during the acquisition process, and performing brightness equalization and contrast enhancement;

[0012] S4. Based on multispectral data, a deep learning material recognition model is used to classify different coating materials, and image processing parameters are generated according to the spectral characteristics of the materials to adjust the image compensation strategy;

[0013] S5. For the image after material recognition, dynamic reflection restoration is performed using the reflection compensation algorithm, using the following formula:

[0014] Among them, I comp (x, y) is the image pixel value after reflection compensation, I orig (x, y) is the initial image pixel value, α is the reflection coefficient, Represents the second-order derivative of the original image;

[0015] S6. Use a multi-path feature pyramid network to extract multi-scale features from the compensated image, extract feature information from details to overall shape, and enhance the features of each layer through a convolutional neural network;

[0016] S7, based on the enhanced feature image, using detail enhancement algorithm to process edges and details;

[0017] S8. Use a real-time camera to capture dynamic images during the lamination process, use edge computing devices to generate processed defect images in real time, and adjust production equipment parameters in real time based on the generated defect data to form a closed-loop control feedback mechanism;

[0018] S9, inputting the processed image data into a preset deep neural network model to perform defect classification and analysis;

[0019] S10. Generate a test report and display the defect image and test results to the operator through a visualization tool.

[0020] Preferably, in step S2, the super-resolution generative adversarial network is adversarial training between a generative network and a discriminative network. The following is the algorithm formula in step S2:

[0021] The perceptual loss: The perceptual loss measures the quality by calculating the feature space distance between the generated high-resolution image and the true high-resolution image. The formula is as follows:

[0022]

[0023] Among them, I HR is a real high-resolution image, G(I LR ) is the high-resolution image generated by the generator, It is the feature map at position (i, j) and channel k in the VGG network. W, H, and C represent the width, height, and number of channels of the feature map.

[0024] The adversarial loss: The adversarial loss minimizes the difference between the distribution of generated images and the distribution of real images through the adversarial process between the generating network and the discriminative network. Its formula is:

[0025] L Gen = -log(D(G(I LR ))), where G(I LR ) is the output of the generator network;

[0026] The content loss: The content loss measures the difference between the generated image and the real image in the pixel space, and its formula is as follows: Where N is the total number of pixels in the image, I HR is a high-resolution real image, G(I LR ) is to generate an image;

[0027] The total loss function: The final generative network loss function is a weighted combination of perceptual loss, adversarial loss, and content loss: L total =L Content +λ1L VGG +λ2L Gen ,

[0028] Among them, λ1 and λ2 are weight coefficients.

[0029] Preferably, in the step S4, the material recognition model based on multispectral data and deep learning automatically identifies the coating material and generates specific image processing parameters. The following is an algorithm formula for material adaptive processing:

[0030] The material classification model: The first step of material adaptive processing is to classify different materials through a convolutional neural network. Assuming that the input image is a multispectral image I multi (x, y), its loss function is calculated by cross entropy loss, the formula is as follows:

[0031] Where N is the total number of samples, y i is the actual material category label, p i is the predicted probability of the i-th category output by the convolutional neural network;

[0032] The material characteristic parameter generation: according to the identified material category, an adaptive parameter set for subsequent image processing is generated, assuming that various materials have different reflective properties to define the parameter generation function F material , the input is the classification result C material , the output is the adaptive processing parameter set θ material :

[0033] θ material =F material (C material )=[α reflect ,β contrast ,γ texture ],

[0034] Among them, α reflect is the coefficient used for reflection compensation, β contrast is the coefficient used for image contrast adjustment, γ texture is the coefficient used for texture enhancement.

[0035] Preferably, in the step S5, the reflective effect caused by the coating material is compensated by an adaptive algorithm, and the dynamic reflective compensation and restoration process includes detection of the reflective area, calculation of the reflective compensation coefficient and restoration of the image;

[0036] The reflective area detection is achieved by identifying the local brightness gradient of the image and the reflection highlight points. The calculation formula of the brightness gradient is:

[0037]

[0038] Among them, G(x,y) is the gradient magnitude of image I(x,y) at position (x,y), and is the gradient of the image in the x and y directions;

[0039] The reflection compensation coefficient calculation: Calculate the reflection compensation coefficient α according to the material characteristics reflect (x,y), the calculation formula is as follows:

[0040]

[0041] Among them, α reflect (x,y) is the reflection compensation coefficient at position (x,y), f(θ material ) is a function related to the material, G maxis the maximum gradient value in the image, and G(x,y) is the gradient amplitude of the image at that point;

[0042] Dynamic reflection compensation and image restoration: Once the reflection compensation coefficient is calculated, the following formula is used to compensate and restore the image to reduce the reflection effect:

[0043] I comp (x,y)=I orig (x,y)-α reflect (x,y)·H(x,y),

[0044] Among them, I comp (x, y) is the image after reflection compensation, I orig (x, y) is the pixel value of the original image, α reflect (x, y) is the reflection compensation coefficient at that point, and H(x, y) is the enhancement function of the highlight area;

[0045] The dynamic reflection compensation relies on the compensation of a single pixel, further smoothes and corrects the image through global optimization, and refines the reflection area using a global balancing strategy. The formula is:

[0046]

[0047] Among them, I final (x, y) is the final restored image, I mean is the average brightness value of the image, and λ is the global brightness balance coefficient.

[0048] Preferably, in step S6, the purpose is to extract features of different scales in the image through a multi-path feature pyramid network. The following is the algorithm formula in step S6:

[0049] The multi-scale feature extraction formula: multi-scale feature extraction is performed on the input image through a convolutional neural network. Assuming that the input image is I(x, y)I(x, y)I(x, y), the feature extraction performs convolution operations through different convolution kernels to extract features of different scales. The basic formula of the convolution operation is as follows:

[0050]

[0051] Among them, F l (x, y) is the feature map after the lth layer of convolution, W l (i, j) is the convolution kernel weight, I(x+i, y+j) is the pixel value of the input image at position (x+i, y+j), and b l is the bias term;

[0052] The feature fusion formula: After multi-scale feature extraction, feature maps of different scales are fused through upper and lower path connections. The fusion process formula is as follows:

[0053] F fused (x,y)=α·F low (x,y)+β·F high (x,y),

[0054] Among them, F fused (x,y) is the fused feature map, F low (x, y) is the bottom-up low-level feature map, F high (x, y) is a top-down high-level feature map, α and β are weight coefficients used to balance details and global information;

[0055] The pyramid hierarchical processing: by refining the feature map layer by layer, the hierarchical processing is usually achieved through downsampling and upsampling operations, and the downsampling process is performed through a pooling operation, and its formula is:

[0056] F pooled (x,y)=max{F(x+i,y+j)∣i,j∈[-k,k]},

[0057] Among them, F pooled (x, y) is the feature map after pooling, and the max operation is used to obtain the maximum value in the local area;

[0058] The multi-path characteristic connection formula:

[0059] Among them, F final (x, y) is the final multi-scale feature map, F i (x,y) is the feature map extracted on different paths, ω i is the weight coefficient of each path.

[0060] Preferably, in the step S7, the edges and details of the image are enhanced by a variety of image processing techniques to make the tiny defects more prominent. The following is an algorithm formula in the step S7, covering the processing of contrast enhancement, edge detection and detail amplification;

[0061] The contrast enhancement formula:

[0062] Among them, I enh (x, y) is the pixel value of the image after contrast enhancement, I(x, y) is the pixel value of the original image, and I min and I max are the minimum and maximum pixel values ​​in the image;

[0063] The detail magnification formula is:detail (x,y)=I enh (x,y)+λ·I edge (x,y),

[0064] Among them, I detail (x, y) is the image after detail magnification, I enh (x, y) is the image after contrast enhancement, I edge (x, y) is the edge enhanced image obtained by Laplacian operator or Sobel operator.

[0065] Preferably, the edge enhancement formula is processed by a Laplace operator or a Sobel operator: the Laplace operator is an edge detection method based on second-order derivatives, and the formula is as follows:

[0066] Among them, I lap (x,y) is the edge enhanced image calculated by the Laplacian operator, and is the second-order derivative of the image in the x and y directions;

[0067] The Sobel operator enhances the edge by calculating the first-order derivative of the image. The formula is:

[0068]

[0069] Among them, G x and G y is the gradient of the image in the x and y directions, W x and W y is the convolution kernel for the Sobel operator, and G(x,y) is the edge enhanced image after combining the horizontal and vertical gradients.

[0070] Preferably, in the step S8, real-time image processing and edge computing technology are used to timely feedback the production equipment for adjustment. The following is the algorithm formula in the step S8;

[0071] The real-time dynamic image capture formula: Among them, I t (x, y) represents the image frame at time t, T is the total number of frames captured, and f is the number of frames captured per second;

[0072] The defect detection formula: P defect (x,y)=CNN(I t (x,y)),

[0073] Among them, P defect (x,y) is the defect probability at position (x,y), CNN(I t(x,y)) is the defect classification model of the pre-trained convolutional neural network for the frame image;

[0074] The defect tracking formula: t (x,y)=I t-1 (x+Δx,y+Δy),

[0075] Among them, I t (x, y) is the pixel value of the current frame, I t-1 (x+Δx, y+Δy) is the pixel value after displacement in the previous frame, Δx and Δy are the displacement of the defect area between each other frames;

[0076] The feedback control formula: F t =γ·∑ (x,y) P defect (x,y)·A(x,y), where F t is the feedback signal generated at time t, γ is the feedback control parameter, P defect (x,y) is the probability of a defect at position (x,y), and A(x,y) is the weight or area of ​​that position.

[0077] Preferably, the real-time feedback mechanism: θ new =θ prev +λ·(F t -F setpoint ),

[0078] Among them, θ new is the adjusted production parameter, θ prev is the current production parameter value, λ is the adjustment step of feedback control, F setpoint is the preset target feedback value.

[0079] Preferably, the defect trend prediction formula is:

[0080] Among them, F pred (t+1) is the feedback signal for predicting the next moment, F t 、F t-1 、F t-2 are the feedback signals of the current and previous moments, and w1, w2, and w3 are weight coefficients.

[0081] The present invention provides a method for detecting the quality of printing and laminating process based on image generation.

[0082] Beneficial effects:

[0083] 1. The present invention introduces multispectral imaging technology and a material adaptive processing model based on deep learning, which can identify different types of coating materials and automatically adjust image processing parameters according to the optical properties of different materials, significantly improving the system's adaptability to a variety of coating materials.

[0084] 2. The present invention improves the detection sensitivity of tiny defects by combining super-resolution generative adversarial networks and detail enhancement technology. The super-resolution technology can generate high-resolution images from low-resolution images, further enhancing the details of tiny defects and making them more prominent.

[0085] 3. By introducing long short-term memory networks, the present invention can predict defect trends, detect quality problems in the production process in advance, and actively adjust production equipment parameters to prevent the expansion of defects, thereby enhancing the intelligence and foresight of the system, enabling it to identify current defects and predict future defect trends. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0087] In order to make the technical personnel in the technical field understand the scheme of the present invention, the technical scheme in the embodiment of the present invention will be clearly and completely described below in combination with the drawings in the embodiment of the present invention. Obviously, the described embodiment is a partial embodiment of the present invention, not a complete embodiment. Based on the embodiment of the present invention, other embodiments obtained by ordinary technicians in the field without creative work should fall within the scope of protection of the present invention.

[0088] The present invention is described in detail below in conjunction with the accompanying drawings:

[0089] Example:

[0090] Please see attached Figure 1 , an embodiment of the present invention provides a printing lamination process quality detection method based on image generation, comprising the following steps: S1, collecting multispectral images of the printed lamination surface to form a multi-dimensional image data set;

[0091] S2, using a super-resolution generative adversarial network to generate a high-resolution image for the multi-spectral initial image to obtain a super-resolution image;

[0092] S3, performing noise filtering on the super-resolution image, combining Gaussian filtering and median filtering methods to eliminate image noise generated during the acquisition process, and performing brightness equalization and contrast enhancement;

[0093] S4. Based on multispectral data, a deep learning material recognition model is used to classify different coating materials, and image processing parameters are generated according to the spectral characteristics of the materials to adjust the image compensation strategy;

[0094] S5. For the image after material recognition, dynamic reflection restoration is performed using the reflection compensation algorithm, using the following formula:

[0095] Among them, I comp (x, y) is the image pixel value after reflection compensation, I orig (x, y) is the initial image pixel value, α is the reflection coefficient, Represents the second-order derivative of the original image;

[0096] S6. Use a multi-path feature pyramid network to extract multi-scale features from the compensated image, extract feature information from details to overall shape, and enhance the features of each layer through a convolutional neural network;

[0097] S7, based on the enhanced feature image, using detail enhancement algorithm to process edges and details;

[0098] S8. Use a real-time camera to capture dynamic images during the lamination process, use edge computing devices to generate processed defect images in real time, and adjust production equipment parameters in real time based on the generated defect data to form a closed-loop control feedback mechanism;

[0099] S9, inputting the processed image data into a preset deep neural network model to perform defect classification and analysis;

[0100] S10. Generate a test report and display the defect image and test results to the operator through a visualization tool.

[0101] The benefits of S1 can help the system accurately distinguish the optical properties of coated materials, especially significantly improving the detection effect on different materials or reflective characteristics;

[0102] The benefits of S2 improve the system's ability to detect tiny defects and reduce missed detections;

[0103] The benefits of S3 are that it can further highlight the details in the image, especially in the case of unfavorable lighting conditions or over-brightness, and enhance the stability and accuracy of defect detection;

[0104] The benefits of S4 are that it can perform targeted image compensation based on the optical characteristics of different materials, improving the system's adaptability when processing a variety of coated materials, allowing the system to efficiently handle surface defects of highly reflective or matte materials;

[0105] Benefits of S5 By adjusting the reflectivity, the system can accurately restore image details and improve the ability to detect defects on reflective materials;

[0106] Benefits of S6 Through enhanced processing of convolutional neural networks, the system can improve the robustness of detection and the accuracy of detail extraction, ensuring that complex surface defects can be identified;

[0107] Benefits of S7: After multi-level processing, the system can capture detailed information more accurately and improve the accuracy of the test results;

[0108] The benefits of S8 are that it can analyze defects in the production process in real time, generate feedback signals immediately, adjust the parameters of production equipment, and quickly correct problems in production through a closed-loop feedback mechanism, reduce defective products, and ensure efficient quality control;

[0109] Benefits of S9 Improve the automation of defect detection, reduce errors caused by manual intervention, and make quality inspection in large-scale production more efficient and accurate;

[0110] The benefits of S10 are that it displays defect images directly to operators through visualization tools, facilitating quick understanding and taking corresponding measures, simplifying manual inspection steps, and providing data support for subsequent process optimization.

[0111] In step S2, the super-resolution generative adversarial network is an adversarial training between the generative network and the discriminative network. The following is the algorithm formula in step S2:

[0112] Perceptual loss: Perceptual loss measures the quality by calculating the feature space distance between the generated high-resolution image and the true high-resolution image. The formula is as follows:

[0113]

[0114] Among them, I HR is a real high-resolution image, G(I LR ) is the high-resolution image generated by the generator, It is the feature map at position (i, j) and channel k in the VGG network. W, H, and C represent the width, height, and number of channels of the feature map.

[0115] Adversarial loss: Adversarial loss minimizes the difference between the distribution of generated images and the distribution of real images through the adversarial process between the generating network and the discriminative network. Its formula is:

[0116] L Gen = -log(D(G(I LR ))), where G(I LR ) is the output of the generator network;

[0117] Content loss: Content loss measures the difference between the generated image and the real image in pixel space, and its formula is as follows: Where N is the total number of pixels in the image, I HR is a high-resolution real image, G(I LR ) is to generate an image;

[0118] Total loss function: The final generative network loss function is a weighted combination of perceptual loss, adversarial loss, and content loss: L total =L Content +λ1L VGG +λ2L Gen ,

[0119] Among them, λ1 and λ2 are weight coefficients.

[0120] The role of the perceptual loss formula is to ensure that the generated image is close to the real image at the pixel level and can be consistent with the higher-level perceptual features by calculating the difference between the generated high-resolution image and the real high-resolution image in the feature space. The perceptual features are extracted through the pre-trained VGG network to measure the high-level information in the image, such as texture, edge, structure, etc.

[0121] The benefits of the perceptual loss formula are that compared to simple pixel differences, the perceptual loss can preserve the global structure and details of the image, and the generated image is more realistic and natural, effectively avoiding the problem of blurry images generated by traditional pixel difference methods, especially when generating high-resolution images;

[0122] The role of the adversarial loss formula is the core of the generative adversarial network. Through adversarial training of the generative network and the discriminative network, the distribution difference between the generated high-resolution images and the real images is minimized. The generative network tries to deceive the discriminative network to make the generated images as realistic as possible, while the discriminative network tries to distinguish between real images and generated images.

[0123] The benefit of the adversarial loss formula is that the generated images are highly similar to the real images in distribution, making the generated high-resolution images sharper and clearer;

[0124] The content loss formula measures the proximity between the generated image and the real image at the original pixel level by calculating the pixel-by-pixel difference between the generated image and the real image in the pixel space. It is a classic mean square error loss and is used to maintain the local consistency of the image.

[0125] The benefit of the content loss formula is that it ensures that the details of the generated image are consistent with those of the real image in the pixel space, especially for the details such as edges and contours, which improves the overall accuracy of the generated image.

[0126] The total loss function formula combines the pixel-level, feature-level, and generative adversarial losses to enable the generative network to generate high-quality, high-resolution images.

[0127] The benefit of the total loss function formula is that it ensures that the generated images are close to the real images at multiple levels, while maintaining the accuracy and details of the generated images, making the generated high-resolution images sharper and more natural.

[0128] In step S4, the material recognition model based on multispectral data and deep learning automatically identifies the coating material and generates specific image processing parameters. The following is the algorithm formula for material adaptive processing:

[0129] Material classification model: The first step of material adaptive processing is to classify different materials through convolutional neural networks. Assuming that the input image is a multispectral image I multi (x, y), its loss function is calculated by cross entropy loss, the formula is as follows:

[0130] Where N is the total number of samples, y i is the actual material category label, p i is the predicted probability of the i-th category output by the convolutional neural network;

[0131] Material characteristic parameter generation: Generate an adaptive parameter set for subsequent image processing based on the identified material category, assuming that various materials have different reflective properties to define the parameter generation function F material , the input is the classification result C material , the output is the adaptive processing parameter set θ material :

[0132] θ material =F material (C material )=[α reflect ,β contrast ,γ texture ],

[0133] Among them, α reflect is the coefficient used for reflection compensation, β contrast is the coefficient used for image contrast adjustment, γ texture is the coefficient used for texture enhancement.

[0134] The material classification model formula measures the difference between the actual material category and the material category probability predicted by the convolutional neural network;

[0135] The benefits of the material classification model formula are that it can effectively distinguish different materials, especially capturing the optical differences of various materials in multispectral images, which can better guide the convergence process of the deep learning model, so that the material classification model can maintain a high recognition accuracy in complex and diverse coating materials;

[0136] The material characteristic parameter generation formula indicates that the material recognition model generates specific image processing parameters based on the classification results for subsequent image compensation and processing;

[0137] The benefits of the material characteristic parameter generation formula are that it ensures that the image processing strategy can be automatically adjusted according to different materials, optimizes processing effects such as reflection compensation, contrast adjustment and texture enhancement, ensures that stable and high-quality image processing effects can be obtained even on complex material surfaces, improves the accuracy of complex material surface defect detection, and reduces errors caused by different materials.

[0138] In step S5, the reflective effect caused by the coating material is compensated by an adaptive algorithm. The dynamic reflective compensation and restoration process includes the detection of the reflective area, the calculation of the reflective compensation coefficient and the restoration of the image.

[0139] Reflective area detection is achieved by identifying the local brightness gradient of the image and the reflection highlight points. The calculation formula of the brightness gradient is:

[0140]

[0141] Among them, G(x,y) is the gradient magnitude of image I(x,y) at position (x,y), and is the gradient of the image in the x and y directions;

[0142] Reflection compensation coefficient calculation: Calculate the reflection compensation coefficient α according to the material characteristics reflect (x,y), the calculation formula is as follows:

[0143]

[0144] Among them, α reflect (x,y) is the reflection compensation coefficient at position (x,y), f(θ material ) is a function related to the material, G max is the maximum gradient value in the image, and G(x,y) is the gradient amplitude of the image at that point;

[0145] Dynamic reflection compensation and image restoration: Once the reflection compensation coefficient is calculated, the following formula is used to compensate and restore the image to reduce the reflection effect:

[0146] I comp (x,y)=I orig (x,y)-αreflect (x,y)·H(x,y),

[0147] Among them, I comp (x, y) is the image after reflection compensation, I orig (x, y) is the pixel value of the original image, α reflect (x, y) is the reflection compensation coefficient at that point, and H(x, y) is the enhancement function of the highlight area;

[0148] Dynamic reflection compensation relies on the compensation of a single pixel, further smoothes and corrects the image through global optimization, and uses a global balancing strategy to refine the reflection area. The formula is:

[0149]

[0150] Among them, I final (x, y) is the final restored image, I mean is the average brightness value of the image, and λ is the global brightness balance coefficient.

[0151] The function of the reflective area detection formula is to detect the brightness change of the reflective area in the image by calculating the local brightness gradient of the image in the x and y directions;

[0152] The benefits of the reflective area detection formula are that it can efficiently identify the highlight areas with strong reflections in the image, provide a basis for subsequent reflection compensation, distinguish between highlight areas caused by reflections and other normal image brightness changes, and reduce erroneous compensation for non-reflective areas;

[0153] The function of the reflection compensation coefficient calculation formula is to calculate according to the material characteristic function and the local gradient amplitude so that the compensation coefficient is obtained in the area with strong reflection;

[0154] The benefit of the reflection compensation coefficient calculation formula is that it can adaptively adjust the reflection compensation intensity, ensuring that the reflection problem of different materials is effectively handled, avoiding over-compensation for areas with small brightness changes, and maintaining the naturalness of the image;

[0155] The function of dynamic reflection compensation and image restoration formula is to perform dynamic reflection compensation and restoration on the image;

[0156] The benefits of dynamic reflection compensation and image restoration formula can reduce the high light interference caused by film reflection in the image, restore the real details of the image, ensure that the brightness and color of other non-reflective areas in the image are not unnecessarily affected, and ensure the overall quality of the image;

[0157] The function of the global equalization strategy formula is to perform global brightness equalization and smoothing on the image after local compensation;

[0158] The benefit of the global balancing strategy formula is that it can solve the local reflection problem, maintain the brightness consistency of the entire image, and prevent image distortion or unevenness caused by excessive processing. The global balancing strategy ensures that the reflective areas in the image are smoothly transitioned while retaining the brightness of other areas, making the final image look more natural and realistic.

[0159] In step S6, the purpose is to extract features of different scales in the image through a multi-path feature pyramid network. The following is the algorithm formula in step S6:

[0160] Multi-scale feature extraction formula: Multi-scale feature extraction is performed on the input image through a convolutional neural network. Assuming that the input image is I(x,y)I(x,y)I(x,y), feature extraction performs convolution operations through different convolution kernels to extract features of different scales. The basic formula of the convolution operation is as follows:

[0161]

[0162] Among them, F l (x, y) is the feature map after the lth layer of convolution, W l (i, j) is the convolution kernel weight, I(x+i, y+j) is the pixel value of the input image at position (x+i, y+j), and b l is the bias term;

[0163] Feature fusion formula: After multi-scale feature extraction, feature maps of different scales are fused through upper and lower path connections. The fusion process formula is as follows:

[0164] F fused (x,y)=α·F low (x,y)+β·F high (x,y),

[0165] Among them, F fused (x,y) is the fused feature map, F low (x, y) is the bottom-up low-level feature map, F high (x, y) is a top-down high-level feature map, α and β are weight coefficients used to balance details and global information;

[0166] Pyramid hierarchical processing: By refining the feature map layer by layer, hierarchical processing is usually achieved through downsampling and upsampling operations. The downsampling process is performed through pooling operations, and its formula is:

[0167] F pooled (x,y)=max{F(x+i,y+j)∣i,j∈[-k,k]},

[0168] Among them, F pooled(x, y) is the feature map after pooling, and the max operation is used to obtain the maximum value in the local area;

[0169] Multipath feature connection formula: Among them, F final (x, y) is the final multi-scale feature map, F i (x,y) is the feature map extracted on different paths, ω i is the weight coefficient of each path.

[0170] The benefits of the multi-scale feature extraction formula are that the system can extract features of different scales from the image. Small-scale convolution kernels capture detail information, and large-scale convolution kernels capture global information. The system can extract high-level features of the image and better understand the global and local features of the image, which helps to detect defects more accurately.

[0171] Benefits of feature fusion formula The system can retain both local details and global structures in the image, making defect detection more comprehensive and accurate. The system can dynamically adjust the feature fusion method according to different tasks to improve the detection accuracy of different types of defects.

[0172] The benefits of the pyramid layered processing formula are that it reduces the dimension of the feature map and significantly reduces the computational overhead, allowing the system to efficiently process large-scale images. The system can retain the most significant features in the feature map, ensuring that key information is not lost during the downsampling process.

[0173] The benefits of the multi-path feature connection formula are that the system can integrate features of different scales, ensuring that both the details and overall information of the image can be reflected in the final feature map, improving the ability to detect diverse defects, and can adaptively adjust according to specific tasks, so that features on certain paths contribute more information in specific scenarios, thereby optimizing feature expression capabilities.

[0174] In step S7, multiple image processing techniques are used to enhance the edges and details of the image, making tiny defects more prominent. The following is the algorithm formula in step S7, covering the processing of contrast enhancement, edge detection and detail amplification;

[0175] Contrast enhancement formula:

[0176] Among them, I enh (x, y) is the pixel value of the image after contrast enhancement, I(x, y) is the pixel value of the original image, and I min and I max are the minimum and maximum pixel values ​​in the image;

[0177] Detail magnification formula: I detail (x,y)=I enh (x,y)+λ·I edge(x,y),

[0178] Among them, I detail (x, y) is the image after detail magnification, I enh (x, y) is the image after contrast enhancement, I edge (x, y) is the edge-enhanced image obtained by Laplacian operator or Sobel operator;

[0179] The edge enhancement formula is processed by the Laplace operator or the Sobel operator: The Laplace operator is an edge detection method based on the second-order derivative, and the formula is as follows:

[0180] Among them, I lap (x,y) is the edge enhanced image calculated by the Laplacian operator, and is the second-order derivative of the image in the x and y directions;

[0181] The Sobel operator enhances the edge by calculating the first-order derivative of the image. The formula is:

[0182]

[0183] Among them, G x and G y is the gradient of the image in the x and y directions, W x and W y is the convolution kernel for the Sobel operator, and G(x,y) is the edge enhanced image after combining the horizontal and vertical gradients.

[0184] The benefits of the contrast enhancement formula make the difference between the bright and dark parts of the image more obvious, highlighting the important information and detail features in the image. The system can better identify and magnify minor defects in the image, such as slight scratches, small dents, etc.;

[0185] Benefits of the detail amplification formula: The edges and subtle textures in the image are further enhanced, making tiny defects more noticeable and easier to detect and process. The system can adaptively adjust according to different scenarios.

[0186] The advantage of the Laplace operator edge enhancement formula is that it can capture the global edge features in the image, and is especially suitable for detecting complex edges and contours in the image. For defects such as scratches and cracks that appear in the printing and laminating process, the Laplace operator can enhance the edges of the defects, making them more prominent in the image and easier to detect.

[0187] The advantage of the Sobel operator edge enhancement formula is that it can perform edge detection in a specific direction, and is particularly suitable for detecting edge information in the horizontal and vertical directions. It can effectively reduce the impact of noise and perform stable detection of edge positions, making the system's edge detection in complex scenes more robust.

[0188] In step S8, real-time image processing and edge computing technology are used to provide timely feedback to production equipment for adjustment. The following is the algorithm formula in step S8;

[0189] Real-time dynamic image capture formula: Among them, I t (x, y) represents the image frame at time t, T is the total number of frames captured, and f is the number of frames captured per second;

[0190] Defect detection formula: P defect (x,y)=CNN(I t (x,y)),

[0191] Among them, P defect (x,y) is the defect probability at position (x,y), CNN(I t (x,y)) is the defect classification model of the pre-trained convolutional neural network for the frame image;

[0192] Defect Tracking Formula: I t (x,y)=I t-1 (x+Δx,y+Δy),

[0193] Among them, I t (x, y) is the pixel value of the current frame, I t-1 (x+Δx, y+Δy) is the pixel value after displacement in the previous frame, Δx and Δy are the displacement of the defect area between each other frames;

[0194] Feedback control formula: F t =γ·∑ (x,y) P defect (x,y)·A(x,y), where F t is the feedback signal generated at time t, γ is the feedback control parameter, P defect (x,y) is the probability of a defect at position (x,y), and A(x,y) is the weight or area of ​​that position;

[0195] Real-time feedback mechanism: θ new =θ prev +λ·(F t -F setpoint ),

[0196] Among them, θ new is the adjusted production parameter, θprev is the current production parameter value, λ is the adjustment step of feedback control, F setpoint is the preset target feedback value;

[0197] Defect trend prediction formula:

[0198] Among them, F pred (t+1) is the feedback signal for predicting the next moment, F t 、F t-1 、F t-2 are the feedback signals of the current and previous moments, and w1, w2, and w3 are weight coefficients.

[0199] The benefits of real-time dynamic image capture formula can track image changes at each moment in the production process in real time, ensuring that defect detection can be carried out in a timely manner to avoid missing any potential defects or production problems;

[0200] Benefits of the defect detection formula The system can efficiently and automatically detect defects in images without manual intervention, reducing the errors and efficiency problems of manual detection. It can accurately identify various defects based on training data and improve detection accuracy, especially suitable for real-time detection of tiny defects;

[0201] The benefits of the defect tracking formula are that it can effectively track defects that move or change during the production process, ensure continuous tracking and monitoring of defects, avoid defects being ignored or missed, monitor the evolution of defects throughout the process, and further analyze whether defects are expanding or worsening;

[0202] The benefits of the feedback control formula are to avoid further expansion or deterioration of the defect problem, dynamically adjust the intensity of the feedback signal according to the severity of the defect, and ensure the flexibility and timeliness of production adjustments;

[0203] The benefits of the real-time feedback mechanism formula are that the parameters of the production equipment can be adjusted in real time according to the detected defect signals, forming a closed-loop control mechanism to ensure that defect detection and equipment adjustment are carried out simultaneously, effectively preventing the defect problem from further expanding, and reducing the generation of defective products and material waste;

[0204] The advantage of the defect trend prediction formula is that it can make production adjustments in advance to prevent defects from further developing or expanding, improve the stability and efficiency of the production line, and the system can make decisions in advance to avoid the production of large-scale unqualified products and improve the overall quality control level of production.

[0205] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A printing lamination process quality detection method based on image generation, characterized in that: The method comprises the following steps: S1, collecting multispectral images of the printed film surface to form a multi-dimensional image data set; S2, using a super-resolution generative adversarial network to generate a high-resolution image for the multi-spectral initial image to obtain a super-resolution image; S3, performing noise filtering on the super-resolution image, combining Gaussian filtering and median filtering methods to eliminate image noise generated during the acquisition process, and performing brightness equalization and contrast enhancement; S4. Based on multispectral data, a deep learning material recognition model is used to classify different coating materials, and image processing parameters are generated according to the spectral characteristics of the materials to adjust the image compensation strategy; S5. For the image after material recognition, dynamic reflection restoration is performed using the reflection compensation algorithm, using the following formula: Among them, I comp (x, y) is the image pixel value after reflection compensation, I orig (x, y) is the initial image pixel value, α is the reflection coefficient, Represents the second-order derivative of the original image; S6. Use a multi-path feature pyramid network to extract multi-scale features from the compensated image, extract feature information from details to overall shape, and enhance the features of each layer through a convolutional neural network; S7, based on the enhanced feature image, using detail enhancement algorithm to process edges and details; S8. Use a real-time camera to capture dynamic images during the lamination process, use edge computing devices to generate processed defect images in real time, and adjust production equipment parameters in real time based on the generated defect data to form a closed-loop control feedback mechanism; S9, inputting the processed image data into a preset deep neural network model to perform defect classification and analysis; S10. Generate a test report and display the defect image and test results to the operator through a visualization tool.

2. The method for detecting the quality of printing and laminating process based on image generation according to claim 1, characterized in that: In the step S2, the super-resolution generative adversarial network is an adversarial training between the generative network and the discriminative network. The following is the algorithm formula in the step S2: The perceptual loss: The perceptual loss measures the quality by calculating the feature space distance between the generated high-resolution image and the true high-resolution image. The formula is as follows: Among them, I HR is a real high-resolution image, G(I LR ) is the high-resolution image generated by the generator, It is the feature map at position (i, j) and channel k in the VGG network. W, H, and C represent the width, height, and number of channels of the feature map. The adversarial loss: The adversarial loss minimizes the difference between the distribution of generated images and the distribution of real images through the adversarial process between the generating network and the discriminative network. Its formula is: L Gen = -log(D(G(I LR ))), where G(I LR ) is the output of the generator network; The content loss: The content loss measures the difference between the generated image and the real image in the pixel space, and its formula is as follows: Where N is the total number of pixels in the image, I HR is a high-resolution real image, G(I LR ) is to generate an image; The total loss function: The final generative network loss function is a weighted combination of perceptual loss, adversarial loss, and content loss: L total =L Content +λ1L VGG +λ2L Gen , Among them, λ1 and λ2 are weight coefficients.

3. The method for detecting the quality of printing and laminating process based on image generation according to claim 1, characterized in that: In the S4 step, the material recognition model based on multispectral data and deep learning automatically identifies the coating material and generates specific image processing parameters. The following is the algorithm formula for material adaptive processing: The material classification model: The first step of material adaptive processing is to classify different materials through a convolutional neural network. Assuming that the input image is a multispectral image I multi (x, y), its loss function is calculated by cross entropy loss, the formula is as follows: Where N is the total number of samples, y i is the actual material category label, p i is the predicted probability of the i-th category output by the convolutional neural network; The material characteristic parameter generation: according to the identified material category, an adaptive parameter set for subsequent image processing is generated, assuming that various materials have different reflective properties to define the parameter generation function F material , the input is the classification result C material , the output is the adaptive processing parameter set θ material : i material =F material (C material )=[a reflect ,b contrast ,c texture ], Among them, α reflect is the coefficient used for reflection compensation, β contrast is the coefficient used for image contrast adjustment, γ texture is the coefficient used for texture enhancement.

4. The method for detecting the quality of printing and laminating process based on image generation according to claim 1, characterized in that: In the step S5, the reflective effect caused by the coating material is compensated by an adaptive algorithm, and the dynamic reflective compensation and restoration process includes the detection of the reflective area, the calculation of the reflective compensation coefficient and the restoration of the image; The reflective area detection is achieved by identifying the local brightness gradient of the image and the reflection highlight points. The calculation formula of the brightness gradient is: Among them, G(x,y) is the gradient magnitude of image I(x,y) at position (x,y), and is the gradient of the image in the x and y directions; The reflection compensation coefficient calculation: Calculate the reflection compensation coefficient α according to the material characteristics reflect (x,y), the calculation formula is as follows: Among them, α reflect (x,y) is the reflection compensation coefficient at position (x,y), f(θ material ) is a function related to the material, G max is the maximum gradient value in the image, and G(x,y) is the gradient amplitude of the image at that point; Dynamic reflection compensation and image restoration: Once the reflection compensation coefficient is calculated, the following formula is used to compensate and restore the image to reduce the reflection effect: I comp (x,y)=I orig (x,y)-α reflect (x,y)·H(x,y), Among them, I comp (x, y) is the image after reflection compensation, I orig (x, y) is the pixel value of the original image, α reflect (x, y) is the reflection compensation coefficient at that point, and H(x, y) is the enhancement function of the highlight area; The dynamic reflection compensation relies on the compensation of a single pixel, further smoothes and corrects the image through global optimization, and refines the reflection area using a global balancing strategy. The formula is: Among them, I final (x, y) is the final restored image, I mean is the average brightness value of the image, and λ is the global brightness balance coefficient.

5. The method for detecting the quality of printing and laminating process based on image generation according to claim 1, characterized in that: The S6 step aims to extract features of different scales in the image through a multi-path feature pyramid network. The following is the algorithm formula in the S6 step: The multi-scale feature extraction formula: multi-scale feature extraction is performed on the input image through a convolutional neural network. Assuming that the input image is I(x, y)I(x, y)I(x, y), the feature extraction performs convolution operations through different convolution kernels to extract features of different scales. The basic formula of the convolution operation is as follows: Among them, F l (x, y) is the feature map after the lth layer of convolution, W l (i, j) is the convolution kernel weight, I(x+i, y+j) is the pixel value of the input image at position (x+i, y+j), and b l is the bias term; The feature fusion formula: After multi-scale feature extraction, feature maps of different scales are fused through upper and lower path connections. The fusion process formula is as follows: F fused (x,y)=α·F low (x,y)+β·F high (x,y), Among them, F fused (x,y) is the fused feature map, F low (x, y) is the bottom-up low-level feature map, F high (x, y) is a top-down high-level feature map, α and β are weight coefficients used to balance details and global information; The pyramid hierarchical processing: by refining the feature map layer by layer, the hierarchical processing is usually achieved through downsampling and upsampling operations, and the downsampling process is performed through a pooling operation, and its formula is: F pooled (x,y)=max{F(x+i,y+j)∣i,j∈[-k,k]}, Among them, F pooled (x, y) is the feature map after pooling, and the max operation is used to obtain the maximum value in the local area; The multi-path characteristic connection formula: Among them, F final (x, y) is the final multi-scale feature map, F i (x,y) is the feature map extracted on different paths, ω i is the weight coefficient of each path.

6. The method for detecting the quality of printing and laminating process based on image generation according to claim 1, characterized in that: In the step S7, the edges and details of the image are enhanced by various image processing techniques to make the tiny defects more prominent. The following is the algorithm formula in the step S7, covering the processing of contrast enhancement, edge detection and detail amplification; The contrast enhancement formula: Among them, I enh (x, y) is the pixel value of the image after contrast enhancement, I(x, y) is the pixel value of the original image, and I min and I max are the minimum and maximum pixel values ​​in the image; The detail magnification formula is: detail (x,y)=I enh (x,y)+λ·I edge (x,y), Among them, I detail (x, y) is the image after detail magnification, I enh (x, y) is the image after contrast enhancement, I edge (x, y) is the edge enhanced image obtained by Laplacian operator or Sobel operator.

7. A printing lamination process quality detection method based on image generation according to claim 6, characterized in that: The edge enhancement formula is processed by the Laplace operator or the Sobel operator: the Laplace operator is an edge detection method based on the second-order derivative, and the formula is as follows: Among them, I lap (x,y) is the edge enhanced image calculated by the Laplacian operator, and is the second-order derivative of the image in the x and y directions; The Sobel operator enhances the edge by calculating the first-order derivative of the image. Its formula is: Among them, G x and G y is the gradient of the image in the x and y directions, W x and W y is the convolution kernel for the Sobel operator, and G(x,y) is the edge enhanced image after combining the horizontal and vertical gradients.

8. The method for detecting the quality of printing and laminating process based on image generation according to claim 1, characterized in that: In the step S8, real-time image processing and edge computing technology are used to provide timely feedback to production equipment for adjustment. The following is the algorithm formula in step S8; The real-time dynamic image capture formula: Among them, I t (x, y) represents the image frame at time t, T is the total number of frames captured, and f is the number of frames captured per second; The defect detection formula: P defect (x,y)=CNN(I t (x,y)), Among them, P defect (x,y) is the defect probability at position (x,y), CNN(I t (x,y)) is the defect classification model of the pre-trained convolutional neural network for the frame image; The defect tracking formula: t (x,y)=I t-1 (x+Δx,y+Δy), Among them, I t (x, y) is the pixel value of the current frame, I t-1 (x+Δx, y+Δy) is the pixel value after displacement in the previous frame, Δx and Δy are the displacement of the defect area between each other frames; The feedback control formula: F t =γ·∑ (x,y) P defect (x,y)·A(x,y), where F t is the feedback signal generated at time t, γ is the feedback control parameter, P defect (x,y) is the probability of a defect at position (x,y), and A(x,y) is the weight or area of ​​that position.

9. The method for detecting the quality of printing and laminating process based on image generation according to claim 8, characterized in that: The real-time feedback mechanism: new =θ prev +λ·(F t -F setpoint ), Among them, θ new is the adjusted production parameter, θ prev is the current production parameter value, λ is the adjustment step of feedback control, F setpoint is the preset target feedback value.

10. The method for detecting the quality of printing and laminating process based on image generation according to claim 8, characterized in that: The defect trend prediction formula: Among them, F pred (t+1) is the feedback signal for predicting the next moment, F t 、F t-1 、F t-2 are the feedback signals of the current and previous moments, and w1, w2, and w3 are weight coefficients.

Images