Corrugated board adhesion defect detection method and system based on thermal imaging

Abstract

The present application relates to the technical field of nondestructive testing and computer vision, and proposes a corrugated board bonding defect detection method and system based on thermal imaging, which comprises: extracting the temperature field dynamic response characteristics of the temperature attenuation sequence image to generate a transient thermal feature map of the heated corrugated board; inputting the transient thermal feature map into a lightweight residual network of a spatial channel double attention mechanism to extract a joint feature map of the deep texture and heat distribution of the heated corrugated board; combining a conditional random field to perform spatial consistency constraint optimization on the joint feature map to output a pixel-level bonding defect probability distribution map of the heated corrugated board; post-processing the bonding defect probability distribution map to obtain a post-processing probability distribution map, and extracting a defect mask of the post-processing probability distribution map to determine the bonding defect positioning and bonding defect classification of the heated corrugated board. The present application can realize the automatic positioning and reliable classification of the corrugated board delamination defect.

Description

Technical Field

[0001] This invention relates to a method and system for detecting adhesive defects in corrugated cardboard based on thermal imaging, belonging to the field of interdisciplinary technology of nondestructive testing and computer vision. Background Technology

[0002] Corrugated board bonding defect detection refers to the process of using machine vision or sensing technology to automatically identify and judge the bonding status between the face paper and the corrugated core paper in the corrugated board production line in real time. By quantitatively analyzing the bonding quality to replace traditional manual sampling inspection, it can promptly detect and eliminate quality hazards such as delamination, bubbling, and weak bonding, effectively avoiding the risk of logistics damage caused by insufficient packaging strength. It plays a key role in ensuring the structural stability of the finished packaging, reducing the company's waste costs, and improving the level of automation control of the production line.

[0003] Traditional methods for detecting bonding defects in corrugated cardboard often employ visible light machine vision or contact ultrasonic technology. However, these methods cannot simultaneously achieve effective detection of internal defects and high-speed online inspection. Visible light vision can only identify surface flaws and cannot detect hidden internal delamination. While contact ultrasonic technology can measure internal defects, it requires point-by-point physical scanning, which is extremely slow and cannot meet the real-time requirements of high-speed production lines. This results in low reliability of corrugated cardboard bonding defect detection. Summary of the Invention

[0004] This invention provides a method and system for detecting adhesive defects in corrugated cardboard based on thermal imaging. Its main purpose is to achieve automated location and reliable classification of delamination defects in corrugated cardboard.

[0005] To achieve the above objectives, the present invention provides a method for detecting bonding defects in corrugated cardboard based on thermal imaging, comprising: A rapid pulse heating method is used to heat the corrugated cardboard using an active heat source to obtain heated corrugated cardboard. Subsequently, an infrared thermal imager is used to simultaneously acquire temperature decay sequence images of the surface of the heated corrugated cardboard. Extract the dynamic response features of the temperature field from the temperature decay sequence image to generate a transient thermal feature map of the heated corrugated cardboard; The transient thermal feature map is input into a lightweight residual network with a spatial channel dual attention mechanism to extract a joint feature map of the deep texture and thermal distribution of the heated corrugated cardboard. By combining conditional random fields to optimize the spatial consistency of the joint feature map, a pixel-level probability distribution map of adhesive defects of the heated corrugated cardboard is output. The probability distribution map of the bonding defects is post-processed to obtain a post-processed probability distribution map. The defect mask of the post-processed probability distribution map is extracted to determine the location and classification of bonding defects in the heated corrugated cardboard.

[0006] Optionally, the transient thermal feature map is input into a lightweight residual network with a spatial channel dual attention mechanism to extract a joint feature map of the deep texture and thermal distribution of the heated corrugated cardboard, including: The standardized transient thermal feature map is input into the backbone network of the lightweight residual network for preliminary feature extraction to obtain a preliminary feature map; The preliminary feature map is sequentially input into the dual attention mechanism constructed by the channel attention module and the spatial attention module; The channel attention module is used to perform global average pooling and multilayer perceptron network on the preliminary feature map along the channel dimension to generate channel weight coefficients. The channel weight coefficients are then multiplied by the preliminary feature map to output the first weighted feature map. The spatial attention module is used to perform average pooling and max pooling on the first weighted feature map in the spatial dimension to generate spatial weight coefficients. The spatial weight coefficients are then multiplied by the first weighted feature map to output the second weighted feature map. The second weighted feature map is fused with the identity mapping feature map corresponding to the backbone network element by element to output the joint feature map of deep texture and heat distribution.

[0007] Optionally, the standardized transient thermal feature map is input into the backbone network of the lightweight residual network for preliminary feature extraction to obtain a preliminary feature map, including: The standardized transient thermal feature map is input into the first layer of the backbone network's standard convolution to perform initial channel expansion and basic thermal feature mapping, resulting in an identity mapping feature map. Construct a lightweight feature extraction sequence consisting of multiple sequentially connected depthwise separable convolutional layers; The identity mapping feature map is input into the lightweight feature extraction sequence to obtain the convolution output feature; After the convolutional output features are processed by batch normalization layers and nonlinear activation functions, the output size is compressed and the channel dimension is expanded, resulting in the preliminary feature map.

[0008] Optionally, the channel attention module is used to perform global average pooling and a multilayer perceptron network to generate channel weight coefficients on the preliminary feature map along the channel dimension, including: The channel attention module is used to simultaneously perform global average pooling and global max pooling operations on the preliminary feature map along the spatial dimension, generating an average pooling channel descriptor representing the global background heat distribution and a max pooling channel descriptor representing the peak response of local defects, respectively. The average pooling channel descriptor and the max pooling channel descriptor are input into a shared multilayer perceptron network for feature mapping, and the corresponding first saliency feature and second saliency feature are output respectively. The first salient feature and the second salient feature are added element by element, and then mapped to the zero-to-one interval by an activation function to generate the channel weight coefficients.

[0009] Optionally, the spatial attention module is used to perform average pooling and max pooling on the first weighted feature map in the spatial dimension to generate spatial weight coefficients, including: Using the spatial attention module, average pooling and max pooling operations are performed along the channel dimension of the first weighted feature map to generate an average pooling spatial description map representing the spatial background heat distribution and a max pooling spatial description map representing the peak response of spatial defects. The average pooling spatial description map and the max pooling spatial description map are concatenated along the channel dimension to generate a dual-channel spatial description feature. The dual-channel spatial description features are convolved to reduce the dimensionality to a single channel, and then mapped to the zero-to-one interval by an activation function to generate the spatial weight coefficients used to characterize the saliency of thermal texture at the location of local defects in corrugated cardboard.

[0010] Optionally, spatial consistency constraint optimization is performed on the joint feature map using a conditional random field to output a pixel-level probability distribution map of adhesive defects in the heated corrugated cardboard, including: The classification prediction probability of the corresponding pixel in the joint feature map is defined as the unary potential of the conditional random field. Based on the spatial distance between adjacent pixels, the cosine similarity of feature vectors in the joint feature map, and the label compatibility function, the pairwise potential energy function of the heated corrugated cardboard is constructed. The global energy function of the heated corrugated cardboard is constructed by combining the univariate potential energy and the paired potential energy function. The global energy function is then iteratively minimized using the mean field approximation algorithm to update the bonding defect probability of each pixel until convergence. This process smooths outlier points within the defect region and sharpens defect edges, outputting a pixel-level bonding defect probability distribution map of the heated corrugated cardboard.

[0011] Optionally, based on the spatial distance between adjacent pixels, the cosine similarity of feature vectors in the joint feature map, and the label compatibility function, a pairwise potential energy function for the heated corrugated cardboard is constructed, including: Based on the spatial distance between the adjacent pixels and the cosine similarity of the feature vectors in the joint feature map, a two-sided Gaussian kernel function is constructed. By combining the bilateral Gaussian kernel function and the label compatibility function, the paired potential energy function of the heated corrugated cardboard is constructed.

[0012] The bilateral Gaussian kernel function refers to an exponential function with the natural constant e as the base.

[0013] Optionally, the dynamic response features of the temperature field in the temperature decay sequence image are extracted to generate a transient thermal feature map of the heated corrugated cardboard, including: The initial peak temperature, the cooling slope within a preset time window, and the thermal decay half-life when the temperature drops to half of the peak temperature are extracted from each pixel in the temperature decay sequence image as time-domain dynamic response features. The relative temperature difference matrix of each pixel in the temperature decay sequence image with respect to the global average temperature is calculated as a spatial dynamic response feature. The time-domain dynamic response features and the spatial-domain dynamic response features are spliced ​​and fused along the channel dimension to generate the multi-channel transient thermal feature map.

[0014] Optionally, after generating the multi-channel transient thermal feature map, the method further includes: The Min-Max normalization algorithm is used to normalize the data of each channel in the transient thermal feature map to the same numerical range, thus obtaining a normalized transient thermal feature map.

[0015] To address the aforementioned problems, the present invention also provides a thermal imaging-based corrugated cardboard bonding defect detection system, the system comprising: The pulse thermal imaging acquisition module is used to rapidly pulse heat the corrugated cardboard with an active heat source to obtain heated corrugated cardboard, and then synchronously acquires temperature decay sequence images of the surface of the heated corrugated cardboard through an infrared thermal imager. The transient feature extraction module is used to extract the dynamic response features of the temperature field from the temperature decay sequence image to generate a transient thermal feature map of the heated corrugated cardboard. The joint feature parsing module is used to input the transient thermal feature map into a lightweight residual network with a spatial channel dual attention mechanism to extract the joint feature map of the deep texture and heat distribution of the heated corrugated cardboard. The spatial consistency optimization module is used to perform spatial consistency constraint optimization on the joint feature map by combining conditional random fields, and output a pixel-level probability distribution map of adhesive defects of the heated corrugated cardboard. The defect location and classification module is used to post-process the adhesive defect probability distribution map to obtain a post-processed probability distribution map, and extract the defect mask from the post-processed probability distribution map to determine the location and classification of adhesive defects in the heated corrugated cardboard.

[0016] Beneficial effects: By effectively amplifying the thermal conduction differences at the bonding defects of corrugated cardboard through active pulse thermal excitation and capturing transient thermal response features using infrared sequence images, this approach combines a lightweight residual network with a spatial channel dual attention mechanism to deeply mine the joint features of deep texture and thermal distribution. Furthermore, a conditional random field is introduced for spatial consistency constraint optimization to eliminate isolated false detection noise. Finally, defect masks are extracted through post-processing to achieve accurate localization and classification. This scheme significantly improves the pixel-level segmentation accuracy and generalization ability of delamination defects in complex backgrounds while ensuring the model's lightweight and easy deployment. It achieves non-destructive, efficient, and highly reliable automated detection, providing strong technical support for real-time online monitoring of corrugated cardboard production quality. Attached Figure Description

[0017] Figure 1 A schematic flowchart of a corrugated cardboard bonding defect detection system based on thermal imaging is provided in an embodiment of the present invention. Figure 2 A schematic diagram of a lightweight residual network structure for a method of detecting adhesive defects in corrugated cardboard based on thermal imaging, provided in an embodiment of the present invention; Figure 3 A functional block diagram of a corrugated cardboard bonding defect detection system based on thermal imaging is provided in an embodiment of the present invention. Figure 4 A schematic diagram of a computer device for detecting adhesive defects in corrugated cardboard based on thermal imaging, according to an embodiment of the present invention. The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0018] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0019] This application provides a method for detecting bonding defects in corrugated cardboard based on thermal imaging. The executing entity of this method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the method for detecting bonding defects in corrugated cardboard based on thermal imaging can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster.

[0020] Reference Figure 1 The diagram shown is a flowchart illustrating a method for detecting bonding defects in corrugated cardboard based on thermal imaging, according to an embodiment of the present invention. In this embodiment, the method for detecting bonding defects in corrugated cardboard based on thermal imaging includes: S1. The corrugated cardboard is rapidly pulse-heated using an active heat source to obtain heated corrugated cardboard, and then the temperature decay sequence image of the surface of the heated corrugated cardboard is synchronously acquired by an infrared thermal imager.

[0021] It should be explained that the active heat source refers to a pulse heating device capable of instantly releasing high-energy heat flow to stimulate differences in heat conduction within the cardboard, such as a high-power halogen lamp array or an infrared strobe lamp; the corrugated cardboard refers to a multi-layered packaging material made by bonding and pressing face paper, core paper, and liner paper together using a gluing machine; the heated corrugated cardboard refers to corrugated cardboard whose internal heat distribution is in a non-steady-state conduction process after absorbing the aforementioned pulse heat flow; the infrared thermal imager refers to an infrared focal plane detector with high frame rate non-contact temperature measurement capability, used to capture minute changes in infrared radiation on the cardboard surface; and the temperature decay sequence image refers to a collection of multiple frames of infrared thermal images continuously acquired after heating stops, reflecting the decrease in surface temperature of the cardboard over time.

[0022] S2. Extract the dynamic response features of the temperature field from the temperature decay sequence image to generate a transient thermal feature map of the heated corrugated cardboard.

[0023] The present invention extracts the dynamic response features of the temperature field from the temperature decay sequence image to generate a transient thermal feature map of the heated corrugated cardboard. This effectively amplifies the transient difference between the heat retention effect caused by air insulation at the degummed defect and the rapid cooling of the normal area, and significantly improves the contrast and signal-to-noise ratio of weak internal defects against a complex background.

[0024] Specifically, the step of extracting the dynamic response features of the temperature field from the temperature decay sequence image to generate a transient thermal feature map of the heated corrugated cardboard includes: The initial peak temperature, the cooling slope within a preset time window, and the thermal decay half-life when the temperature drops to half of the peak temperature are extracted from each pixel in the temperature decay sequence image as time-domain dynamic response features. The relative temperature difference matrix of each pixel in the temperature decay sequence image with respect to the global average temperature is calculated as a spatial dynamic response feature. The time-domain dynamic response features and the spatial-domain dynamic response features are spliced ​​and fused along the channel dimension to generate the multi-channel transient thermal feature map.

[0025] The initial peak temperature refers to the highest temperature reached by the pixel on the cardboard surface at the instant the pulse heating stops. For example, the initial peak temperature is significantly higher in the debonded area because heat is difficult to conduct to the interior. The cooling slope within the preset time window refers to the rate at which the pixel temperature decreases over time within a specific period after heating ends. For example, the heat in the normally bonded area is rapidly absorbed by the core paper, resulting in a larger absolute value of the cooling slope. The thermal decay half-life when the temperature drops to half of the peak temperature refers to the time required for the pixel temperature to cool from the initial peak temperature to half, used to quantify the persistence of heat retention. The time-domain dynamic response characteristic refers to the change in temperature of a single pixel over time, which is composed of the initial peak temperature, cooling slope, and thermal decay half-life. The system comprises a set of changing thermodynamic parameters; the global average temperature refers to the overall arithmetic mean of the temperatures of all pixels at a certain moment in the temperature decay sequence image, reflecting the current overall cooling baseline of the cardboard; the relative temperature difference matrix refers to the two-dimensional spatial difference array formed by subtracting the global average temperature from the temperature of each pixel, highlighting local abnormal hot spots; the spatial dynamic response feature refers to the spatial comparison information of the difference between the heat distribution of the pixel and its surrounding environment as represented by the relative temperature difference matrix; and the transient thermal feature map refers to the multi-dimensional data tensor formed by superimposing and splicing the temporal dynamic response feature and the spatial dynamic response feature in the channel dimension, such as a three-dimensional feature map containing a temperature peak channel, a cooling slope channel, and a spatial temperature difference channel.

[0026] Furthermore, after generating the multi-channel transient thermal feature map, the method further includes: The Min-Max normalization algorithm is used to normalize the data of each channel in the transient thermal feature map to the same numerical range, thus obtaining a normalized transient thermal feature map.

[0027] The Min-Max normalization algorithm is a linear data scaling method used to eliminate differences in physical dimensions, thereby eliminating the impact of dimensional differences between different thermal property parameters on subsequent feature extraction by the network.

[0028] S3. Input the transient thermal feature map into a lightweight residual network with a spatial channel dual attention mechanism to extract the joint feature map of the deep texture and thermal distribution of the heated corrugated cardboard.

[0029] This invention inputs the transient thermal feature map into a lightweight residual network with a spatial channel dual attention mechanism to extract the joint feature map of the deep texture and thermal distribution of the heated corrugated cardboard. The spatial channel dual attention mechanism focuses on the tiny thermal textures at the defect edges and suppresses background thermal noise. At the same time, the lightweight network significantly reduces computational overhead, thereby efficiently extracting the joint feature map that is highly coupled with the physical thermal field and micro-texture, significantly improving the recognition accuracy and robustness of tiny adhesive defects.

[0030] See Figure 2 The diagram illustrates a lightweight residual network structure for a thermal imaging-based method for detecting adhesive defects in corrugated cardboard, according to an embodiment of the present invention. The lightweight residual network takes a transient thermal feature map as input. First, it extracts features through a standard convolutional layer (3×3 kernel). Then, it enters a lightweight feature extraction sequence composed of multiple depthwise separable convolutional layers stacked sequentially. Next, it performs feature normalization and nonlinear mapping through batch normalization layers and nonlinear activation function layers to obtain a preliminary feature map. This preliminary feature map is then fed into two parallel attention modules: a channel attention module (containing global average pooling, global max pooling, and a multilayer perceptron network) and a spatial attention module (containing average pooling, max pooling, and convolutional dimensionality reduction). These two modules weight the features from the channel dimension and the spatial dimension, respectively. Finally, the outputs of the two attention modules are fused through an element-wise addition fusion node to obtain the final joint feature map, which is used for subsequent defect detection. Furthermore, the dashed arrows in the diagram represent identity mapping feature maps, reflecting the design concept of skip connections in the residual network.

[0031] Specifically, the step of inputting the transient thermal feature map into a lightweight residual network with a spatial channel dual attention mechanism to extract a joint feature map of the deep texture and thermal distribution of the heated corrugated cardboard includes: The standardized transient thermal feature map is input into the backbone network of the lightweight residual network for preliminary feature extraction to obtain a preliminary feature map; The preliminary feature map is sequentially input into the dual attention mechanism constructed by the channel attention module and the spatial attention module; The channel attention module is used to perform global average pooling and multilayer perceptron network on the preliminary feature map along the channel dimension to generate channel weight coefficients. The channel weight coefficients are then multiplied by the preliminary feature map to output the first weighted feature map. The spatial attention module is used to perform average pooling and max pooling on the first weighted feature map in the spatial dimension to generate spatial weight coefficients. The spatial weight coefficients are then multiplied by the first weighted feature map to output the second weighted feature map. The second weighted feature map is fused with the identity mapping feature map corresponding to the backbone network element by element to output the joint feature map of deep texture and heat distribution.

[0032] The lightweight residual network refers to a compact neural network that uses depthwise separable convolutions instead of standard convolutions to significantly reduce the number of parameters and introduces cross-layer connections to prevent gradient vanishing, such as a lightweight feature extractor suitable for industrial edge computing devices. The backbone network refers to the sequence of convolutional layers in this lightweight network responsible for progressively extracting basic visual information from the image, used to capture the basic temperature gradient and contour of the corrugated cardboard surface. The preliminary feature map refers to the feature matrix containing primary thermal distribution and texture information output after the standardized transient thermal feature map has undergone convolutional processing by the backbone network. The dual attention mechanism refers to a feature adaptive weighting module that combines channel feature filtering with spatial location focusing. The channel attention module refers to a sub-network that evaluates the importance of each thermal property parameter along the channel dimension of the feature map, for example, automatically evaluating the cooling slope channel relative to the initial temperature channel through global average pooling and fully connected layers. The channel attention module is more capable of detecting defects. The channel weight coefficient is a normalized numerical vector output by the channel attention module to measure the importance of each channel. The first weighted feature map is a feature matrix that suppresses invalid background thermal noise and strengthens key thermodynamic channels after multiplying the preliminary feature map with the channel weight coefficient. The spatial attention module is a sub-network that evaluates the importance of local pixel regions along the spatial dimension of the feature map to accurately locate the physical location of delamination or bubbles on the cardboard surface. The spatial weight coefficient is a two-dimensional weight matrix output by the spatial attention module to measure the importance of each spatial location in the feature map. The second weighted feature map is a feature matrix that highlights the subtle texture changes of the defect edge in the spatial dimension after multiplying the first weighted feature map with the spatial weight coefficient. The joint feature map is a deep fusion representation output after element-wise addition of the second weighted feature map and the identity mapping feature map.

[0033] Further, the step of inputting the standardized transient thermal feature map into the backbone network of the lightweight residual network for preliminary feature extraction to obtain a preliminary feature map includes: The standardized transient thermal feature map is input into the first layer of the backbone network's standard convolution to perform initial channel expansion and basic thermal feature mapping, resulting in an identity mapping feature map. Construct a lightweight feature extraction sequence consisting of multiple sequentially connected depthwise separable convolutional layers; The identity mapping feature map is input into the lightweight feature extraction sequence to obtain the convolution output feature; After the convolutional output features are processed by batch normalization layers and nonlinear activation functions, the output size is compressed and the channel dimension is expanded, resulting in the preliminary feature map.

[0034] The first-layer standard convolution refers to a convolutional layer that uses a conventional two-dimensional convolutional kernel to perform the first full-parameter feature mapping on the input transient thermal feature map. For example, a 3x3 convolutional kernel can be used to increase the number of channels in the feature map containing the basic temperature gradient to thirty-two dimensions. The identity mapping feature map refers to the feature matrix directly output by the first-layer standard convolution and spanning all subsequent deep network structures. The depthwise separable convolutional layer refers to a lightweight convolutional unit that splits the standard convolution operation into two steps: channel-wise spatial feature extraction and point-wise channel linear combination. The lightweight feature extraction sequence refers to a feature dimensionality reduction and mapping module composed of multiple depthwise separable convolutional layers stacked sequentially. The batch normalization layer refers to a processing layer that standardizes the mean and variance of the convolutional output features trained in each batch. The nonlinear activation function refers to a function that introduces nonlinear mapping capabilities to enhance the network's defect discrimination ability, such as a linear rectified function that filters and truncates the thermal feature response.

[0035] Optionally, the identity mapping feature map is input into the lightweight feature extraction sequence and spatial sliding extraction is performed sequentially using channel-wise convolution kernels, and linear combination of channel dimensions is performed using point-wise convolution kernels to obtain convolution output features.

[0036] Further, the step of using the channel attention module to perform global average pooling and multilayer perceptron network generation on the preliminary feature map along the channel dimension includes: The channel attention module is used to simultaneously perform global average pooling and global max pooling operations on the preliminary feature map along the spatial dimension, generating an average pooling channel descriptor representing the global background heat distribution and a max pooling channel descriptor representing the peak response of local defects, respectively. The average pooling channel descriptor and the max pooling channel descriptor are input into a shared multilayer perceptron network for feature mapping, and the corresponding first saliency feature and second saliency feature are output respectively. The first salient feature and the second salient feature are added element by element, and then mapped to the zero-to-one interval by an activation function to generate the channel weight coefficients.

[0037] Wherein, the global background heat distribution refers to the macroscopic attenuation trend of the overall surface temperature field of corrugated cardboard during heating and the substrate thermal radiation state of a large area of ​​defect-free region; the average pooling channel descriptor refers to a one-dimensional vector obtained by averaging the pixel values ​​at all spatial locations of the feature map, which suppresses local high-frequency noise and extracts the overall background heat energy benchmark value of each feature channel; the local defect peak response refers to the phenomenon of a significant local temperature surge in the de-adhesion or bubble area of ​​the cardboard surface, which is significantly higher than the surrounding normal area due to changes in thermal resistance; the maximum pooling channel descriptor refers to a one-dimensional vector formed by extracting the maximum pixel value in each channel of the feature map. The vector captures the strongest thermal anomaly response signal at the defect location; the multilayer perceptron network refers to a fully connected neural network structure containing an input layer, a hidden layer, and an output layer, such as a dimension-reduced hidden layer containing sixteen neurons, which performs cross-feature reconstruction on two pooling descriptors by sharing weights; the first saliency feature refers to the feature vector output by the average pooling descriptor after dimensionality reduction and dimensionality increase mapping by the network, representing the contribution weight distribution of each channel under the background thermal field; the second saliency feature refers to the feature vector output by the max pooling descriptor after processing by the same network, representing the sensitivity and response intensity of each channel to the local defect thermal peak.

[0038] Further, the step of using the spatial attention module to perform average pooling and max pooling on the first weighted feature map in the spatial dimension to generate spatial weight coefficients includes: Using the spatial attention module, average pooling and max pooling operations are performed along the channel dimension of the first weighted feature map to generate an average pooling spatial description map representing the spatial background heat distribution and a max pooling spatial description map representing the peak response of spatial defects. The average pooling spatial description map and the max pooling spatial description map are concatenated along the channel dimension to generate a dual-channel spatial description feature. The dual-channel spatial description features are convolved to reduce the dimensionality to a single channel, and then mapped to the zero-to-one interval by an activation function to generate the spatial weight coefficients used to characterize the saliency of thermal texture at the location of local defects in corrugated cardboard.

[0039] The spatial background thermal distribution refers to the base temperature field distribution of a large, defect-free area on the surface of corrugated cardboard in terms of length and width. The average pooling spatial description map refers to a single-channel matrix generated by averaging across the feature map, which suppresses the specific thermal noise between channels and extracts the macroscopic background thermal energy benchmark of the cardboard surface in spatial dimensions. The spatial defect peak response refers to the local spatial temperature surge caused by delamination on the surface of the corrugated cardboard or by obstructed heat conduction at the location of internal air bubbles. The max-pooling spatial description map refers to a single-channel matrix generated by maximizing across the feature map, which captures the strongest thermal anomaly surge signal at each spatial location in all channels. The dual-channel spatial description feature refers to a two-dimensional feature tensor generated by superimposing the two single-channel matrices generated by the average pooling and max-pooling in depth, for example, by splicing the background description map and the peak description map with dimensions of single channel multiplied by height multiplied by width into a joint tensor of dual channel multiplied by height multiplied by width. The local defect location thermal texture refers to the specific spatial geometric patterns and micro-topological textures formed by abrupt temperature gradient changes at the edges of delamination or air bubbles.

[0040] S4. Combine the conditional random field to optimize the spatial consistency constraint of the joint feature map, and output the pixel-level adhesive defect probability distribution map of the heated corrugated cardboard.

[0041] This invention combines conditional random fields to optimize the spatial consistency of the joint feature map, and outputs a pixel-level probability distribution map of adhesive defects of the heated corrugated cardboard. This effectively suppresses the salt-and-pepper noise and edge blurring caused by independent pixel classification, thereby significantly improving the boundary coherence and positioning accuracy of the pixel-level probability distribution map of adhesive defects of the corrugated cardboard.

[0042] In detail, the combined conditional random field is used to optimize the spatial consistency of the joint feature map, outputting a pixel-level probability distribution map of adhesive defects in the heated corrugated cardboard, including: The classification prediction probability of the corresponding pixel in the joint feature map is defined as the unary potential of the conditional random field. Based on the spatial distance between adjacent pixels, the cosine similarity of feature vectors in the joint feature map, and the label compatibility function, the pairwise potential energy function of the heated corrugated cardboard is constructed. The global energy function of the heated corrugated cardboard is constructed by combining the univariate potential energy and the paired potential energy function. The global energy function is then iteratively minimized using the mean field approximation algorithm to update the bonding defect probability of each pixel until convergence. This process smooths outlier points within the defect region and sharpens defect edges, outputting a pixel-level bonding defect probability distribution map of the heated corrugated cardboard.

[0043] Wherein, the classification prediction probability refers to the initial confidence value of the front-end network for each pixel of the corrugated cardboard to belong to the delamination or normal category; the conditional random field refers to the probabilistic graphical model that fuses the single-pixel prediction with the thermal correlation of adjacent pixels of the cardboard; the univariate potential energy refers to the energy term that measures the confidence of the cardboard's single-pixel defect classification; the spatial distance refers to the Euclidean geometric distance between two pixels of the corrugated cardboard on the joint feature map; the feature vector refers to the high-dimensional data in the joint feature map that characterizes the local deep thermodynamics and texture of the cardboard; the label compatibility function refers to the weight term that penalizes adjacent pixels for assigning different defect labels, for example, the same label is penalized with zero and different labels are positive; the pairwise potential energy function refers to the energy term that calculates the correlation between adjacent pixels based on the spatial distance of the cardboard and the similarity of the joint features; the mean field approximation algorithm refers to the method of approximating the global distribution through independent edge probability iteration to optimize the inference efficiency of cardboard thermal features; the probability outlier refers to the isolated defect pixel in the large background of the cardboard or the isolated background pixel in the defect area; the adhesive defect probability distribution map refers to the optimized output two-dimensional matrix that characterizes the final probability of delamination or air bubbles at each coordinate point of the cardboard.

[0044] Optionally, the tag compatibility function typically employs the Potts model.

[0045] Furthermore, the construction of the pairwise potential energy function for the heated corrugated cardboard based on the spatial distance between adjacent pixels, the cosine similarity of the feature vectors in the joint feature map, and the label compatibility function includes: Based on the spatial distance between the adjacent pixels and the cosine similarity of the feature vectors in the joint feature map, a two-sided Gaussian kernel function is constructed. By combining the bilateral Gaussian kernel function and the label compatibility function, the paired potential energy function of the heated corrugated cardboard is constructed.

[0046] The bilateral Gaussian kernel function refers to an exponential function with the natural constant e as the base.

[0047] As another optional embodiment of the present invention, the expression of the bilateral Gaussian kernel function is:

[0048] in, Represents the two-sided Gaussian kernel function. Represents pixels Two-dimensional spatial coordinates on the joint feature map, Represents pixels Two-dimensional spatial coordinates on the joint feature map, Represents pixels in the joint feature map eigenvectors, Represents pixels in the joint feature map eigenvectors, Represents pixels on the joint feature map and pixels Spatial distance, Represents pixels on the joint feature map and pixels The cosine similarity of the feature vectors. Represents spatial scale hyperparameters. This represents the feature scale hyperparameter.

[0049] It should be noted that the spatial scale hyperparameter is used to control the range of action of the spatial distance constraint. Its value can be empirically set based on the resolution of the infrared thermal imager and the physical size of typical delamination defects in corrugated cardboard. The feature scale hyperparameter is used to control the sensitivity decay rate of the conversion of cosine similarity differences into energy penalties. The smaller the spatial distance and the higher the feature similarity, the larger the output value of the kernel function, which in turn makes the penalty effect of the pairwise potential function on label inconsistency stronger. The value of the feature scale hyperparameter can be empirically set based on the statistical distribution characteristics of the joint feature vector during the training phase.

[0050] S5. Post-process the bonding defect probability distribution map to obtain a post-processed probability distribution map, and extract the defect mask from the post-processed probability distribution map to determine the location and classification of bonding defects in the heated corrugated cardboard.

[0051] It should be explained that the post-processing probability distribution map refers to a smooth and continuous probability matrix that accurately represents the final state of each pixel of the heated corrugated cardboard belonging to the real adhesive defect after noise suppression and morphological correction; in detail, the post-processing specifically includes threshold binarization, connected component denoising and morphological closing operation steps.

[0052] Finally, by extracting the defect mask from the post-processing probability distribution map, the present invention achieves precise spatial isolation and boundary contour delineation from pixel-level probability distribution to physical defect areas in order to determine the location and classification of bonding defects in the heated corrugated cardboard. This provides structured data support for the automated and accurate location and reliable classification of bonding defects in heated corrugated cardboard.

[0053] Specifically, the step of extracting the defect mask from the post-processing probability distribution map to determine the location and classification of bonding defects in the heated corrugated cardboard includes: The post-processed probability distribution map is compared with a preset probability threshold, and pixels that are greater than or equal to the preset probability threshold are marked as defect pixels to generate a binarized defect mask. The location of the bonding defects in the heated corrugated cardboard is determined based on the coordinate distribution of the defect pixels in the defect mask. Based on the connected domain geometric features of the defect mask, the adhesive defects of the heated corrugated cardboard are classified.

[0054] The preset probability threshold refers to the critical probability value used to distinguish between defects and normal background, for example, set to 0.5. When the pixel prediction probability exceeds this value, it is determined that there is a risk of delamination. The defect pixel refers to the pixel point determined to belong to the adhesion abnormal area. The defect mask refers to the binary image matrix composed of defect pixels, where defect area pixels are assigned a value of 1 and normal background pixels are assigned a value of 0, which is used to transform the complex probability distribution into a pure defect spatial topology. The adhesion defect location refers to the specific spatial coordinate position and circumscribed rectangle outline of the delamination area on the entire heated corrugated cardboard. The adhesion defect classification refers to the specific defect morphology of the heated corrugated cardboard, for example, using the area size of the connected region to distinguish between micro-bubble delamination and large-area poor glue adhesion, or using the aspect ratio to distinguish between strip-shaped crack delamination and block-shaped spot delamination, thereby determining the category attribute of the specific adhesion defect.

[0055] Optionally, the preset probability threshold is empirically set based on the tolerance requirements for false detection rate and false negative rate in actual heated corrugated cardboard testing.

[0056] like Figure 3 The diagram shown is a functional block diagram of a corrugated cardboard bonding defect detection system based on thermal imaging according to the present invention.

[0057] The thermal imaging-based corrugated cardboard bonding defect detection system 300 described in this invention can be installed in an electronic device. Depending on the functions implemented, the thermal imaging-based corrugated cardboard bonding defect detection system includes a pulse thermal imaging acquisition module 301, a transient feature extraction module 302, a joint feature parsing module 303, a spatial consistency optimization module 304, and a defect localization and classification module 305. The modules described in this invention can also be referred to as units, which are a series of computer program segments that can be executed by the processor of an electronic device and perform a fixed function, stored in the memory of the electronic device.

[0058] In this embodiment of the invention, the functions of each module / unit are as follows: The pulse thermal imaging acquisition module 301 is used to rapidly pulse heat the corrugated cardboard with an active heat source to obtain heated corrugated cardboard, and then synchronously acquire temperature decay sequence images of the surface of the heated corrugated cardboard through an infrared thermal imager. The transient feature extraction module 302 is used to extract the dynamic response features of the temperature field of the temperature decay sequence image to generate the transient thermal feature map of the heated corrugated cardboard. The joint feature parsing module 303 is used to input the transient thermal feature map into a lightweight residual network with a spatial channel dual attention mechanism to extract the joint feature map of the deep texture and heat distribution of the heated corrugated cardboard. The spatial consistency optimization module 304 is used to perform spatial consistency constraint optimization on the joint feature map by combining a conditional random field, and output a pixel-level probability distribution map of adhesive defects of the heated corrugated cardboard. The defect location and classification module 305 is used to post-process the adhesive defect probability distribution map to obtain a post-processed probability distribution map, and extract the defect mask from the post-processed probability distribution map to determine the location and classification of adhesive defects in the heated corrugated cardboard.

[0059] In detail, the modules in the thermal imaging-based corrugated cardboard bonding defect detection system 300 described in this embodiment of the invention employ the same methods as described above during use. Figure 1 The method used is the same as the thermal imaging-based method for detecting adhesive defects in corrugated cardboard described above, and it can produce the same technical effect, so it will not be repeated here.

[0060] In one embodiment, a computer device is provided, which may be a server or a client, and its internal structure diagram may be as follows: Figure 4 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used for communication with external clients via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a thermal imaging-based corrugated cardboard bonding defect detection method on the server or client side.

[0061] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: The pulse thermal imaging acquisition module is used to rapidly pulse heat the corrugated cardboard with an active heat source to obtain heated corrugated cardboard, and then synchronously acquires temperature decay sequence images of the surface of the heated corrugated cardboard through an infrared thermal imager. The transient feature extraction module is used to extract the dynamic response features of the temperature field from the temperature decay sequence image to generate a transient thermal feature map of the heated corrugated cardboard. The joint feature parsing module is used to input the transient thermal feature map into a lightweight residual network with a spatial channel dual attention mechanism to extract the joint feature map of the deep texture and heat distribution of the heated corrugated cardboard. The spatial consistency optimization module is used to perform spatial consistency constraint optimization on the joint feature map by combining conditional random fields, and output a pixel-level probability distribution map of adhesive defects of the heated corrugated cardboard. The defect location and classification module is used to post-process the adhesive defect probability distribution map to obtain a post-processed probability distribution map, and extract the defect mask from the post-processed probability distribution map to determine the location and classification of adhesive defects in the heated corrugated cardboard.

[0062] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor: The pulse thermal imaging acquisition module is used to rapidly pulse heat the corrugated cardboard with an active heat source to obtain heated corrugated cardboard, and then synchronously acquires temperature decay sequence images of the surface of the heated corrugated cardboard through an infrared thermal imager. The transient feature extraction module is used to extract the dynamic response features of the temperature field from the temperature decay sequence image to generate a transient thermal feature map of the heated corrugated cardboard. The joint feature parsing module is used to input the transient thermal feature map into a lightweight residual network with a spatial channel dual attention mechanism to extract the joint feature map of the deep texture and heat distribution of the heated corrugated cardboard. The spatial consistency optimization module is used to perform spatial consistency constraint optimization on the joint feature map by combining conditional random fields, and output a pixel-level probability distribution map of adhesive defects of the heated corrugated cardboard. The defect location and classification module is used to post-process the adhesive defect probability distribution map to obtain a post-processed probability distribution map, and extract the defect mask from the post-processed probability distribution map to determine the location and classification of adhesive defects in the heated corrugated cardboard.

[0063] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0064] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0065] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0066] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0067] Finally, it should be noted that in the above embodiments, each embodiment can be combined with each other or independent. Deleting any one of them will not affect the technical implementation of other embodiments. The above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for detecting bonding defects in corrugated cardboard based on thermal imaging, characterized in that, The method includes: A rapid pulse heating method is used to heat the corrugated cardboard using an active heat source to obtain heated corrugated cardboard. Subsequently, an infrared thermal imager is used to simultaneously acquire temperature decay sequence images of the surface of the heated corrugated cardboard. Extract the dynamic response features of the temperature field from the temperature decay sequence image to generate a transient thermal feature map of the heated corrugated cardboard; The transient thermal feature map is input into a lightweight residual network with a spatial channel dual attention mechanism to extract a joint feature map of the deep texture and thermal distribution of the heated corrugated cardboard. By combining conditional random fields to optimize the spatial consistency of the joint feature map, a pixel-level probability distribution map of adhesive defects of the heated corrugated cardboard is output. The probability distribution map of the bonding defects is post-processed to obtain a post-processed probability distribution map. The defect mask of the post-processed probability distribution map is extracted to determine the location and classification of bonding defects in the heated corrugated cardboard.

2. The method for detecting bonding defects in corrugated cardboard based on thermal imaging as described in claim 1, characterized in that, The transient thermal feature map is input into a lightweight residual network with a spatial channel dual attention mechanism to extract a joint feature map of the deep texture and thermal distribution of the heated corrugated cardboard, including: The standardized transient thermal feature map is input into the backbone network of the lightweight residual network for preliminary feature extraction to obtain a preliminary feature map; The preliminary feature map is sequentially input into the dual attention mechanism constructed by the channel attention module and the spatial attention module; The channel attention module is used to perform global average pooling and multilayer perceptron network on the preliminary feature map along the channel dimension to generate channel weight coefficients. The channel weight coefficients are then multiplied by the preliminary feature map to output the first weighted feature map. The spatial attention module is used to perform average pooling and max pooling on the first weighted feature map in the spatial dimension to generate spatial weight coefficients. The spatial weight coefficients are then multiplied by the first weighted feature map to output the second weighted feature map. The second weighted feature map is fused with the identity mapping feature map corresponding to the backbone network element by element to output the joint feature map of deep texture and heat distribution.

3. The method for detecting bonding defects in corrugated cardboard based on thermal imaging as described in claim 2, characterized in that, The standardized transient thermal feature map is input into the backbone network of the lightweight residual network for preliminary feature extraction to obtain a preliminary feature map, including: The standardized transient thermal feature map is input into the first layer of the backbone network's standard convolution to perform initial channel expansion and basic thermal feature mapping, resulting in an identity mapping feature map. Construct a lightweight feature extraction sequence consisting of multiple sequentially connected depthwise separable convolutional layers; The identity mapping feature map is input into the lightweight feature extraction sequence to obtain the convolution output feature; After the convolutional output features are processed by batch normalization layers and nonlinear activation functions, the output size is compressed and the channel dimension is expanded, resulting in the preliminary feature map.

4. The method for detecting bonding defects in corrugated cardboard based on thermal imaging as described in claim 2, characterized in that, The channel attention module performs global average pooling and a multilayer perceptron network to generate channel weight coefficients on the preliminary feature map along the channel dimension, including: The channel attention module is used to simultaneously perform global average pooling and global max pooling operations on the preliminary feature map along the spatial dimension, generating an average pooling channel descriptor representing the global background heat distribution and a max pooling channel descriptor representing the peak response of local defects, respectively. The average pooling channel descriptor and the max pooling channel descriptor are input into a shared multilayer perceptron network for feature mapping, and the corresponding first saliency feature and second saliency feature are output respectively. The first salient feature and the second salient feature are added element by element, and then mapped to the zero-to-one interval by an activation function to generate the channel weight coefficients.

5. The method for detecting bonding defects in corrugated cardboard based on thermal imaging as described in claim 2, characterized in that, The spatial attention module is used to perform average pooling and max pooling on the first weighted feature map in the spatial dimension to generate spatial weight coefficients, including: Using the spatial attention module, average pooling and max pooling operations are performed along the channel dimension of the first weighted feature map to generate an average pooling spatial description map representing the spatial background heat distribution and a max pooling spatial description map representing the peak response of spatial defects. The average pooling spatial description map and the max pooling spatial description map are concatenated along the channel dimension to generate a dual-channel spatial description feature. The dual-channel spatial description features are convolved to reduce the dimensionality to a single channel, and then mapped to the zero-to-one interval by an activation function to generate the spatial weight coefficients used to characterize the saliency of thermal texture at the location of local defects in corrugated cardboard.

6. The method for detecting bonding defects in corrugated cardboard based on thermal imaging as described in claim 1, characterized in that, By combining conditional random fields with spatial consistency constraint optimization of the joint feature map, a pixel-level probability distribution map of bonding defects in the heated corrugated cardboard is output, including: The classification prediction probability of the corresponding pixel in the joint feature map is defined as the unary potential of the conditional random field. Based on the spatial distance between adjacent pixels, the cosine similarity of feature vectors in the joint feature map, and the label compatibility function, the pairwise potential energy function of the heated corrugated cardboard is constructed. The global energy function of the heated corrugated cardboard is constructed by combining the univariate potential energy and the paired potential energy function. The global energy function is then iteratively minimized using the mean field approximation algorithm to update the bonding defect probability of each pixel until convergence. This process smooths outlier points within the defect region and sharpens defect edges, outputting a pixel-level bonding defect probability distribution map of the heated corrugated cardboard.

7. The method for detecting bonding defects in corrugated cardboard based on thermal imaging as described in claim 6, characterized in that, Based on the spatial distance between adjacent pixels, the cosine similarity of feature vectors in the joint feature map, and the label compatibility function, a pairwise potential energy function for the heated corrugated cardboard is constructed, including: Based on the spatial distance between the adjacent pixels and the cosine similarity of the feature vectors in the joint feature map, a two-sided Gaussian kernel function is constructed. By combining the bilateral Gaussian kernel function and the label compatibility function, the paired potential energy function of the heated corrugated cardboard is constructed. The bilateral Gaussian kernel function refers to an exponential function with the natural constant e as the base.

8. The method for detecting bonding defects in corrugated cardboard based on thermal imaging as described in claim 1, characterized in that, Extracting the dynamic response features of the temperature field from the temperature decay sequence image to generate a transient thermal feature map of the heated corrugated cardboard includes: The initial peak temperature, the cooling slope within a preset time window, and the thermal decay half-life when the temperature drops to half of the peak temperature are extracted from each pixel in the temperature decay sequence image as time-domain dynamic response features. The relative temperature difference matrix of each pixel in the temperature decay sequence image with respect to the global average temperature is calculated as a spatial dynamic response feature. The time-domain dynamic response features and the spatial-domain dynamic response features are spliced ​​and fused along the channel dimension to generate the multi-channel transient thermal feature map.

9. The method for detecting bonding defects in corrugated cardboard based on thermal imaging as described in claim 8, characterized in that, Following the generation of the multi-channel transient thermal feature map, the following is also included: The Min-Max normalization algorithm is used to normalize the data of each channel in the transient thermal feature map to the same numerical range, thus obtaining a normalized transient thermal feature map.

10. A corrugated cardboard bonding defect detection system based on thermal imaging, characterized in that, For performing a method for detecting adhesive defects in corrugated cardboard based on thermal imaging as described in any one of claims 1-9, the system comprises: The pulse thermal imaging acquisition module is used to rapidly pulse heat the corrugated cardboard with an active heat source to obtain heated corrugated cardboard, and then synchronously acquires temperature decay sequence images of the surface of the heated corrugated cardboard through an infrared thermal imager. The transient feature extraction module is used to extract the dynamic response features of the temperature field from the temperature decay sequence image to generate a transient thermal feature map of the heated corrugated cardboard. The joint feature parsing module is used to input the transient thermal feature map into a lightweight residual network with a spatial channel dual attention mechanism to extract the joint feature map of the deep texture and heat distribution of the heated corrugated cardboard. The spatial consistency optimization module is used to perform spatial consistency constraint optimization on the joint feature map by combining conditional random fields, and output a pixel-level probability distribution map of adhesive defects of the heated corrugated cardboard. The defect location and classification module is used to post-process the adhesive defect probability distribution map to obtain a post-processed probability distribution map, and extract the defect mask from the post-processed probability distribution map to determine the location and classification of adhesive defects in the heated corrugated cardboard.

Images