Carton board surface coating uniformity detection method and system based on multi-sensor fusion

Abstract

This invention discloses a method and system for detecting the uniformity of coatings on cardboard surfaces based on multi-sensor fusion, belonging to the field of industrial vision inspection technology. The method includes: S1, outputting displacement pulses through a rotary encoder to trigger a controller to synchronously drive a linear array visible light camera and an infrared thermal imager to simultaneously acquire matched visible light and infrared images; S2, performing geometric registration and channel stitching on the two images to generate a multi-channel registered image; S3, feeding the image into a multi-task convolutional neural network, sharing the encoder to extract features, and having each decoder output a coating confidence map and a relative thickness map respectively; S4, adaptively binarizing the confidence map to obtain a mask, and multiplying it pixel-by-pixel with the thickness map to generate a corrected thickness map; S5, statistically analyzing the mean of the thickness standard deviation in the corrected thickness map to obtain a quantified result of coating uniformity. This invention combines visible light and infrared multi-sensor imaging to achieve non-contact online detection of transparent adhesive coatings, solving the problem of difficult and reliable online detection of uniformity in this type of coating.

Description

Technical Field

[0001] This invention relates to the field of industrial visual inspection technology, specifically to a method and system for detecting the uniformity of coatings on cardboard surfaces based on multi-sensor fusion. Background Technology

[0002] In the corrugated cardboard box manufacturing process, the uniformity of the adhesive coating directly affects the bonding strength of the cardboard, the compression resistance of the box, and the product qualification rate. The coated adhesive is usually a transparent or semi-transparent material, and its coating thickness spatial distribution needs to be strictly monitored. Existing detection methods mainly include manual weighing, offline thickness measurement, and online detection based on a single sensor. Manual weighing calculates the coating amount by weighing a cut sample, which is a destructive test with poor real-time performance and can only reflect the average value of a region, failing to obtain the spatial distribution of the coating thickness. Offline precision thickness measurement devices, such as laser displacement sensors or eddy current thickness gauges, cannot be used online because the adhesive layer is transparent and its thickness is on the micrometer to sub-millimeter scale. Ordinary optical reflection methods cannot accurately capture the air-coating interface, and contact measurement can contaminate or scratch the undried coating. Visible light machine vision-based methods use area or line scan cameras to acquire grayscale or color images of the cardboard surface, attempting to invert coating thickness from image intensity differences. However, transparent coatings have extremely weak absorption and scattering of visible light, and grayscale changes in the coating area in the image are mainly dominated by external factors such as the texture of the substrate cardboard and uneven ambient lighting, obscuring coating thickness information and making it difficult to establish a reliable grayscale-thickness mapping relationship. Infrared thermographic methods utilize the difference in thermal emissivity between the adhesive and the cardboard substrate or temperature changes in the coating area after coating to identify coating distribution. However, due to the lack of visible light texture reference, the low signal-to-noise ratio of a single infrared image, and the blurring of coating boundaries caused by temperature differences in the cardboard itself, thermal reflection interference, and thermal diffusion effects during transportation, it is impossible to effectively distinguish the coated area from non-coating abnormal areas such as stains and watermarks on the cardboard, resulting in unstable thickness inversion based on infrared images.

[0003] Furthermore, some technologies attempt to fuse visible light and infrared images, but these often employ two independent trigger sources or rely on software timestamp alignment. Due to sensor frame rate fluctuations and transmission delays, rows or pixels in the two images are spatially misaligned, resulting in artifacts or thickness jumps at the coating edges after fusion, severely impacting the accuracy of subsequent uniformity calculations. At the data processing level, traditional image processing workflows first segment the coating region separately, then estimate the thickness within the segmented region. This sequential approach not only accumulates errors but also uses different feature representations for segmentation and thickness estimation, making it difficult to guarantee the completeness of thickness information at boundaries. Recent deep learning models often perform segmentation or regression on a single task, failing to jointly optimize the two related tasks of coating region identification and thickness measurement from the same feature space. This prevents the network from fully utilizing the complementary information between the two tasks, limiting the reliability of uniformity assessment. Therefore, there is an urgent need for a coating uniformity detection method and system that can achieve synchronous triggering at the sensor hardware level, precise pixel registration, and simultaneous coating identification and thickness regression. Summary of the Invention

[0004] In view of the shortcomings of the prior art, the purpose of this invention is to provide a method and system for detecting the uniformity of coatings on cardboard surfaces based on multi-sensor fusion, to solve the problem of unreliable online, non-contact detection of the uniformity of transparent adhesive coatings. This invention uses a rotary encoder to simultaneously trigger a linear array visible light camera and a linear array infrared thermal imager to acquire spatially corresponding images of the cardboard surface line by line. After geometric registration and channel stitching, these images are input into a pre-constructed multi-task convolutional neural network. This network extracts features using a shared encoder, and a mask decoder generates a coating presence confidence map, while a thickness decoder generates a coating relative thickness map. The confidence scores are then analyzed. Figure 2 After value-based scaling, the thickness map is multiplied pixel by pixel to obtain a corrected thickness map containing only the coating area. Then, the mean of the standard deviation of the thickness of each window is statistically analyzed by dividing the map into windows to output the coating uniformity quantification result.

[0005] This invention provides a method for detecting the uniformity of coatings on cardboard surfaces based on multi-sensor fusion, comprising: S1: The displacement pulse signal generated by the rotary encoder triggers the synchronous trigger controller, which generates a line trigger signal and sends it to the linear array visible light camera and the linear array infrared thermal imager. It collects the visible light image line signal and the infrared image line signal corresponding to the space, and accumulates them to form the visible light image signal and the infrared image signal. S2: Geometric registration and stitching of visible light image signals and infrared image signals are performed to obtain multi-channel registered image signals; S3: Input the multi-channel registered image signal into the multi-task convolutional neural network, extract features by the shared encoder, generate a coating presence confidence map signal by the mask decoder, and generate a coating relative thickness map signal by the thickness decoder. S4: Binarize the confidence map signal of the coating to obtain a binary mask signal, and multiply it pixel by pixel with the coating relative thickness map signal to obtain the corrected thickness map signal; S5: Calculate the thickness standard deviation by dividing the window according to the corrected thickness map signal, and take the average standard deviation of each window as the result signal of coating uniformity quantification.

[0006] In one embodiment of the present invention, geometric registration in S2 includes: performing distortion correction on the visible light image signal using pre-calibrated intrinsic parameters of a linear array visible light camera, and performing distortion correction on the infrared image signal using pre-calibrated intrinsic parameters of a linear array infrared thermal imager; scaling and translation transformation of the distortion-corrected infrared image signal according to pre-calibrated relative spatial pose parameters between the two sensors, so that each pixel corresponds to the same position on the cardboard surface as the visible light image signal; stitching involves stacking the registered visible light image signal and the registered infrared image signal in the channel dimension, so that each pixel position of the multi-channel registered image signal contains a visible light intensity channel value and at least one infrared radiation channel value. The calculation formula for the geometric transformation mapping of the infrared image signal is as follows: , in, To match the infrared pixel coordinate vectors of the visible light pixel positions after registration, k is the global scale scaling factor of the infrared image, and H is the relative spatial pose homography matrix between the linear array visible light camera and the infrared thermal imager. This is a mapping function for infrared image distortion correction. This represents the original infrared pixel coordinate vector before distortion correction. To calibrate the intrinsic parameter matrix of the infrared thermal imager, This is the infrared distortion coefficient vector. For a two-dimensional vector, the 2-norm operator, The d-th order infrared distortion coefficient, Operators for solving matrix determinants This is the x-axis translation offset. This is the y-axis translation offset. Construct an operator for a two-dimensional column vector, where d is the distortion order traversal variable.

[0007] In one embodiment of the present invention, the multi-task convolutional neural network consists of a shared encoder, a mask decoder, and a thickness decoder. The shared encoder contains multiple serially connected convolutional layers, which extract features from the multi-channel registered image signal step by step, and output multi-scale feature maps with progressively decreasing spatial resolution and progressively increasing channel number. Both the mask decoder and the thickness decoder adopt a top-down progressive upsampling structure. During each upsampling stage, the feature map of the previous stage is upsampled and then concatenated with the feature map of the corresponding scale in the shared encoder through a skip connection to achieve channel dimension, and then feature fusion is achieved through convolution operation. The output layer of the mask decoder generates a coating presence confidence map signal through a single-channel convolutional layer and an sigmoid nonlinear activation function that maps the output value to the zero-to-one interval. The output layer of the thickness decoder generates a coating relative thickness map signal through a single-channel convolutional layer and an identity mapping activation function. The mask decoder and the thickness decoder share the multi-scale features from the shared encoder during each upsampling stage, but each has independent convolution parameters.

[0008] In one embodiment of the present invention, the multi-task convolutional neural network employs a joint loss function for parameter optimization during training. The joint loss function is defined as a weighted sum of the mask branch loss term and the thickness branch loss term. The mask branch loss term is obtained by calculating the binary cross-entropy loss based on the coating existence confidence map signal and the pre-acquired true binary mask annotations. The thickness branch loss term is obtained by calculating the smoothed absolute deviation loss based on the coating relative thickness map signal and the pre-acquired true relative thickness map annotations. This smoothed absolute deviation loss takes the squared loss when the absolute value of the difference between the predicted value and the labeled value of a single pixel is less than a preset switching threshold, and takes the linear loss when it is greater than or equal to the switching threshold. The weighting coefficients of the joint loss function are set according to the order-of-magnitude ratio of the mask branch loss and the thickness branch loss in the early stages of training, ensuring that the gradient magnitudes of the two losses match during backpropagation. During training, all learnable parameters in the shared encoder, mask decoder, and thickness decoder are synchronously updated using the backpropagation algorithm. The calculation formula for the joint loss function is as follows: , , , in, Here, λ represents the overall joint loss value of the multi-task convolutional neural network, λ is the weighting coefficient of the mask branch loss, and N is the total number of pixels in the multi-channel registered image. The single-pixel binary cross-entropy loss function is... The actual coating binary mask annotation value for the i-th pixel. Let α be the confidence prediction value for the existence of the coating at pixel i, and let α be the weighting coefficient for the thickness branch loss. The single-pixel smoothing absolute deviation loss function is... This is the predicted relative thickness of the coating for the i-th pixel. This represents the actual relative thickness value of the i-th pixel. This is a conditional indicator function.

[0009] In one embodiment of the present invention, the value of each pixel in the coating relative thickness map signal output by the thickness decoder has a positive linear mapping relationship with the physical thickness of the adhesive coating on the cardboard surface at that pixel location. This positive linear mapping relationship is determined through a pre-calibration process, which includes: acquiring a calibration cardboard with a known continuously varying coating thickness region; acquiring its multi-channel registration image signal on a detection system and inputting it into a multi-task convolutional neural network to obtain the coating relative thickness map signal; simultaneously acquiring the actual coating thickness spatial distribution map of the calibration cardboard using a high-precision thickness measurement device; spatially aligning the actual coating thickness spatial distribution map with the coating relative thickness map signal; taking several sampling points with the coating relative thickness value as the independent variable and the actual coating thickness value as the dependent variable to perform linear regression, obtaining the slope and intercept parameters of the linear transformation and storing them; during online detection, multiplying the pixel value in the coating relative thickness map signal by the slope and adding the intercept to obtain the coating thickness map signal characterizing the absolute coating thickness; and calculating the corrected thickness map signal based on the coating thickness map signal.

[0010] In one embodiment of the present invention, the binarization processing of the coating presence confidence map signal in S4 specifically involves: statistically analyzing the grayscale histogram of the coating presence confidence map signal; using the maximum inter-class variance criterion to traverse possible grayscale values ​​as candidate thresholds; calculating the inter-class variance between the foreground pixel class and the background pixel class under each candidate threshold; selecting the candidate threshold that maximizes the inter-class variance as the adaptive segmentation threshold; determining pixels in the coating presence confidence map signal whose pixel values ​​are greater than or equal to the adaptive segmentation threshold as coating foreground pixels and assigning them a first logical value in the binary mask; determining pixels whose pixel values ​​are less than the adaptive segmentation threshold as background pixels and assigning them a second logical value in the binary mask to form an initial binary mask; and then sequentially performing morphological closing and morphological opening operations on the initial binary mask to fill the holes inside the coating area caused by confidence fluctuations and remove isolated false detection points in the background area to obtain the binary mask signal.

[0011] In one embodiment of the present invention, the process of obtaining the corrected thickness map signal in S4 includes: performing pixel-by-pixel multiplication of the binary mask signal and the coating relative thickness map signal, so that the thickness value at the position corresponding to the background pixel in the binary mask signal is set to zero, while the thickness value at the position corresponding to the foreground pixel of the coating is retained to generate a preliminary thickness map containing only the coating region; performing connected component analysis on the preliminary thickness map, marking each connected region and calculating the pixel area of ​​each connected region, setting the thickness value of all pixels in the connected region whose pixel area is less than the preset minimum coating region area to zero, the preset minimum coating region area being determined according to the lower limit of the spatial scale required for coating defect identification, thereby eliminating sporadic interference responses and obtaining the corrected thickness map signal.

[0012] In one embodiment of the present invention, S5, which involves dividing the window, calculating the thickness standard deviation, and taking the average, specifically involves: uniformly dividing the pixel area covered by the corrected thickness map signal into several equal-sized, non-overlapping rectangular windows along the row and column directions; for each rectangular window, counting the thickness values ​​of all pixels falling within the window, and calculating the arithmetic mean and standard deviation of the thickness values ​​within the window; after filtering out invalid windows whose total number of pixels is less than a preset effective pixel count threshold or whose standard deviation is zero, calculating the arithmetic mean of the standard deviations of all valid windows, and outputting this arithmetic mean as the coating uniformity measurement result signal. The formula for calculating the arithmetic mean is as follows: , , , , Where U represents the final uniformity measurement result of the coating, and M represents the total number of valid rectangular windows after filtering. W represents the set of all valid windows, where W is the set of all rectangular windows used to partition the image. Let w be the standard deviation of the coating thickness within the w-th window. The number of effective pixels of the coating within the w-th window. Correct the absolute coating thickness value for the j-th pixel in the w-th window. The arithmetic mean of the coating thickness within the w-th window. This is a function to indicate the validity of a window. The preset effective pixel count threshold is defined, w is the window traversal variable, and j is the pixel traversal variable within the window.

[0013] In one embodiment of the present invention, the size of the rectangular window is set according to the typical spatial scale range of the uniformity defect of the cardboard adhesive coating, so that the side length of a single rectangular window covers several times the width of the expected local coating defect, so as to retain the standard deviation information reflecting the regional coating thickness fluctuation while suppressing random noise; when dividing the window, the adjacent windows can also be partially overlapped, and the adjacent windows slide in both the row direction and the column direction with a preset step size, which is smaller than the side length of the rectangular window, so as to increase the local statistical density.

[0014] This invention also provides a multi-sensor fusion-based system for detecting the uniformity of coatings on cardboard surfaces, comprising: The synchronous triggering and image acquisition module includes a rotary encoder, a synchronous triggering controller, a linear array visible light camera, and a linear array infrared thermal imager. The rotary encoder is mechanically coupled to the conveyor line and outputs displacement pulse signals. The synchronous triggering controller generates row trigger signals based on the displacement pulse signals and sends them to the two cameras simultaneously. This enables the two cameras to acquire visible light image row signals and infrared image row signals of the same physical row on the surface of the carton line one by one, and accumulate them to form spatially corresponding visible light image signals and infrared image signals. The multispectral image preprocessing module is used to perform geometric registration of visible light image signals and infrared image signals to eliminate parallax and scale differences, and to stack the registered images in the channel dimension to generate multi-channel registered image signals. The multi-task neural network inference module contains a pre-trained multi-task convolutional neural network. This network includes a shared encoder, a mask decoder, and a thickness decoder. The shared encoder extracts multi-scale features from the multi-channel registered image signal. The mask decoder generates a coating presence confidence map signal based on the multi-scale features. The thickness decoder generates a coating relative thickness map signal based on the multi-scale features. The mask generation and thickness correction module is used to perform adaptive threshold binarization and morphological filtering on the coating existence confidence map signal to obtain a binary mask signal, and multiply the binary mask signal with the coating relative thickness map signal pixel by pixel to obtain a corrected thickness map signal; The windowed statistics and uniformity output module is used to divide the corrected thickness map signal into multiple non-overlapping rectangular windows, calculate the standard deviation of the coating thickness value in each window, and take the arithmetic mean of the standard deviations of all windows as the coating uniformity quantification result signal output.

[0015] The invention provides a method and system for detecting the uniformity of coatings on cardboard surfaces based on multi-sensor fusion. This method uses a rotary encoder to simultaneously trigger a linear array visible light camera and a linear array infrared thermal imager to acquire spatially corresponding images of the cardboard surface line by line. After geometric registration and channel stitching, these images are input into a pre-constructed multi-task convolutional neural network. This network extracts features using a shared encoder, and a mask decoder generates a coating presence confidence map, while a thickness decoder generates a coating relative thickness map. The confidence scores are then analyzed. Figure 2 After value-based scaling, the thickness map is multiplied pixel by pixel to obtain a corrected thickness map containing only the coating area. Then, the mean of the standard deviation of the thickness of each window is statistically analyzed by dividing the map into windows to output the coating uniformity quantification result. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of a method for detecting the uniformity of coatings on the surface of cardboard based on multi-sensor fusion. Figure 2 This is a flowchart illustrating the synchronous triggering and image acquisition process; Figure 3 The flowchart shows the process of multispectral preprocessing and multi-task neural network inference. Figure 4 Flowchart for mask generation correction and window uniformity quantization; Figure 5 This is a system architecture diagram of a cardboard surface coating uniformity detection system based on multi-sensor fusion. Detailed Implementation

[0018] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0019] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0020] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0021] Please see Figures 1-5 The figure shows the method and system for detecting the uniformity of coating on the surface of cardboard based on multi-sensor fusion according to the present invention. The method for detecting the uniformity of coating on the surface of cardboard based on multi-sensor fusion according to the present invention includes: S1: using the displacement pulse signal generated by the rotary encoder to trigger the synchronous trigger controller, generating a row trigger signal and simultaneously sending it to the linear array visible light camera and the linear array infrared thermal imager, acquiring the visible light image row signal and the infrared image row signal corresponding to the space, and accumulating them to form visible light image signal and infrared image signal; S2: performing geometric registration and stitching on the visible light image signal and infrared image signal to obtain a multi-channel registered image signal; S3: inputting the multi-channel registered image signal into a multi-task convolutional neural network, extracting features by the shared encoder, generating a coating presence confidence map signal by the mask decoder, and generating a coating relative thickness map signal by the thickness decoder; S4: binarizing the coating presence confidence map signal to obtain a binary mask signal, multiplying it pixel by pixel with the coating relative thickness map signal to obtain a corrected thickness map signal; S5: dividing the corrected thickness map signal into windows to calculate the thickness standard deviation, and taking the average of the standard deviations of each window as the coating uniformity measurement result signal.

[0022] like Figure 1As shown, the core idea of ​​the cardboard adhesive coating uniformity detection method described in this invention lies in establishing a strict spatial correspondence between visible light images and infrared images at the signal acquisition source through a hardware-level row synchronous triggering mechanism, thereby providing a precise physical basis for subsequent multimodal fusion. Specifically, a rotary encoder is installed on the conveyor line of the detection system. The encoder's shaft is directly mechanically coupled to the drive roller or belt of the conveyor line. The encoder rotates with the movement of the conveyor line, outputting a series of pulse signals. Each pulse corresponds to a fixed, small displacement of the conveyor line. The synchronous triggering controller continuously receives this displacement pulse signal and immediately generates a row trigger signal with a fixed width and a steep rising edge when the internal pulse count reaches a preset value. This row trigger signal is simultaneously sent to the external trigger input ports of the linear array visible light camera and the linear array infrared thermal imager. Since the row exposure start time of both sensors is determined by the edge of the same trigger signal, the consistency of their exposure times is guaranteed by the hardware circuit and is not affected by software scheduling or transmission delays. At each trigger, the linear array visible light camera acquires the visible light brightness information of the currently detected row on the cardboard surface, while the linear array infrared thermal imager acquires the infrared radiation information of the same row position. The cardboard moves continuously and uniformly on the conveyor line, and the rotary encoder generates pulses based on displacement increments rather than time increments. Therefore, even with slight fluctuations in the conveyor speed, the spatial distance the cardboard moves at each row trigger moment remains fixed. After row-by-row cumulative reconstruction, the image rows with the same row number in the visible light image signal and the infrared image signal have strictly consistent physical row positions on the cardboard surface. This row-by-row spatial synchronization mechanism fundamentally eliminates the inter-row misalignment problem that may be caused by frame rate jitter or transmission delay, making the two heterogeneous images naturally aligned in the spatial dimension. This significantly reduces the difficulty of subsequent registration algorithms and ensures that pixel information in the coating edge area of ​​the fused image does not appear ghosted or shifted.

[0023] like Figure 2As shown in the flowchart, this flowchart fully presents the entire process of synchronous image acquisition at the front end of the cardboard adhesive coating detection. It incorporates parallel acquisition branches and continuous feeding loop logic. All computations and hardware units within the process are as follows: cardboard moves with the conveyor line; rotary encoder outputs displacement pulse signals; synchronous trigger controller generates line trigger signals by shaping pulses; linear array visible light camera acquires visible light image line signals; linear array infrared thermal imager acquires infrared image line signals; accumulated line signals generate complete visible light and infrared image signals; a node determining whether the conveyor line is continuously feeding; and outputting dual-channel raw images to the preprocessing stage. All units form a closed-loop linkage based on the conveyor line's movement sequence. The movement of the cardboard with the conveyor line is the starting point of the entire acquisition process. The cardboard is placed on the surface of the uniformly rotating conveyor rollers, which continuously transport the material forward, providing a continuous moving carrier for line-by-line image acquisition. The material's movement speed directly determines the line acquisition interval, which is the physical basis of the entire synchronous acquisition sequence.

[0024] The rotary encoder output displacement pulse signal unit is rigidly mechanically coupled to the conveyor roller. During roller rotation, it drives the encoder grating to rotate synchronously. The grating cuts through the built-in light source and outputs continuous, equally spaced electrical pulses. The number of pulses corresponds one-to-one with the carton's movement distance. This pulse signal serves as a unified timing reference for synchronous acquisition by the two imaging devices, compensating for imaging misalignment caused by conveyor speed fluctuations. The synchronous trigger controller's shaping pulse generation line trigger signal unit receives the original displacement pulses. Its internal circuitry performs pulse noise reduction and waveform shaping, filtering out noise interference from the transmission line. Then, according to the preset line acquisition interval, it generates a dual-channel synchronous line trigger command, simultaneously outputting completely synchronized acquisition signals to both imaging devices. This ensures that each line of image acquired by both cameras corresponds to the same physical position on the carton. A parallel acquisition branch is formed by a linear array visible light camera acquiring visible light image line signals and a linear array infrared thermal imager acquiring infrared image line signals. Both devices synchronously receive the line trigger signal and complete single-line exposure respectively. The visible light camera captures the visible light reflectance grayscale information of the cardboard and adhesive, while the infrared thermal imager captures temperature imaging data based on the thermal radiation difference between the coating and the cardboard. Both types of single-line image data flow synchronously downwards. The accumulated line signals generate complete visible light and infrared image signals. The unit continuously caches the single-line image data obtained from each round of parallel acquisition, stitching the line segments into two complete, spatially corresponding two-dimensional images, eliminating the fragmentation problem of single-line data. The conveyor line continuous feeding determination node receives the complete image cache results. After completing one round of material imaging, it automatically determines whether there is a cardboard board to be inspected on the conveyor line. If there is, it jumps to the cardboard board moving with the conveyor line unit to start a new round of cyclic acquisition, achieving uninterrupted continuous inspection of the production line. If there is no new material on the conveyor line, it directly outputs dual-channel raw images to the preprocessing unit, transmitting the paired registered basic raw images to the back-end image processing link. The entire process does not have an independent, single linear flow path. Parallel acquisition branches enable synchronous imaging from dual sensors, and cyclic judgment nodes adapt to continuous production conditions on industrial production lines. Each unit performs its own function, and signals are transmitted step by step. The absence of any hardware unit will cause image timing misalignment and missing imaging data, making it impossible to carry out subsequent multi-channel image fusion processing. It fully covers all hardware working logic of front-end image acquisition.

[0025] After obtaining spatially identical visible light and infrared image signals, geometric registration is performed on them. Although the two sensors achieved spatial synchronization in the row direction during acquisition, differences in lens parameters, sensor pixel size, installation position, and optical axis direction between the linear array visible light camera and the linear array infrared thermal imager can lead to sub-pixel-level scale differences in the column direction, as well as parallax distortion caused by non-parallel optical axes. Therefore, pre-calibrated camera intrinsic and relative extrinsic parameters are used to perform distortion correction and spatial transformation on the visible light and infrared images, respectively, so that every pixel in the registered images points to the same spatial point on the cardboard surface. Subsequently, based on pixel-level alignment, the registered visible light and infrared images are stacked in the channel dimension to form a multi-channel registered image signal. Each spatial location carries both visible light texture information and infrared radiation information, thus providing rich and physically meaningful input data for subsequent convolutional neural networks.

[0026] Multi-channel registered image signals are fed into a pre-built and trained multi-task convolutional neural network. Structurally, this network consists of a shared encoder, a mask decoder, and a thickness decoder. The shared encoder contains multiple sequentially stacked convolutional layers. Each layer progressively reduces the spatial resolution of the feature maps through stride convolution or pooling operations, while gradually increasing the number of channels in the feature maps, thus transforming the input image from pixel space to an abstract feature space with increasing semantic hierarchy. In this process, shallow feature maps retain more spatial detail, while deep feature maps capture more semantic contextual information. These feature maps at different scales collectively constitute a multi-scale feature representation, which is then used by the two subsequent decoders. Both the mask decoder and the thickness decoder employ a top-down, progressively upsampling structure. During decoding, each layer first upsamples the low-resolution feature map from the previous layer, doubling its resolution. Then, it concatenates this upsampled feature map with the same spatial resolution in the shared encoder along the channel dimension, and finally, cross-layer feature fusion is achieved through several convolutional operations. This skip-layer connection design allows the decoder to continuously obtain fine-grained spatial cues from the encoder during the spatial resolution recovery process, which is beneficial for the accurate reconstruction of coating region boundaries and subtle thickness variations. The last layer of the mask decoder is a single-channel convolutional layer, followed by an sigmoid nonlinear activation function that restricts the output value to the interval between zero and one. Its output is the coating presence confidence map signal, where the value of each pixel represents the probability that the location belongs to the coating region. The last layer of the thickness decoder is also a single-channel convolutional layer, but it uses an identity mapping activation function to maintain the continuous unbounded nature of the output value. Its output is the coating relative thickness map signal, where the value of each pixel in the signal has a positive linear mapping relationship with the physical thickness of the adhesive coating at the corresponding location, that is, the larger the value, the thicker the coating.

[0027] like Figure 3 As shown in the flowchart, this flowchart illustrates the complete computational chain from raw dual-channel image input to coating prediction image output. It employs a dual-decoder parallel branch structure with no purely linear flow. The process includes receiving raw visible light and infrared image signals, visible light image distortion correction, infrared image distortion correction, infrared image scaling and translational geometric registration, channel stacking to generate multi-channel registered image signals, shared encoder extracting multi-scale fusion features, mask decoder generating coating presence confidence map signals, thickness decoder generating coating relative thickness map signals, and simultaneous output of both prediction images. All computational units are represented, with each unit corresponding to the core computational steps of image preprocessing and deep learning inference. The data flow logic of each unit is complete and clear. The unit receiving raw visible light and infrared image signals receives two complete sets of two-dimensional images transmitted from the front-end acquisition process, uniformly buffering the pixel matrices of both images to provide the raw data source for subsequent correction calculations. Simultaneously, it retrieves camera intrinsic parameters and sensor pose parameters stored in the previous equipment calibration as auxiliary data for correction calculations. The visible light image distortion correction unit retrieves the exclusive calibration parameters of the linear visible light camera and corrects the image coordinates pixel-by-pixel for radial and tangential optical distortion of the lens, eliminating imaging distortions such as edge stretching and local distortion. The infrared image distortion correction unit simultaneously calls the calibration parameters of the infrared thermal imager to perform the same logical distortion correction on the raw infrared image, unifying the distortion elimination standards for both types of images. The two correction operations are performed simultaneously in parallel, shortening preprocessing time. The infrared image scaling and translation geometric registration unit reads the relative spatial pose calibration parameters of the two sensors and performs global scaling and two-dimensional coordinate translation transformation on the corrected infrared image, aligning the physical coordinates of the cardboard corresponding to the infrared and visible light images pixel-by-pixel, completely eliminating parallax offset caused by different installation positions of the two devices, achieving pixel-level spatial matching. The channel stacking generation multi-channel registration image signal unit receives two perfectly aligned corrected images and performs data stacking in the image channel dimension. Each pixel simultaneously stores the visible light reflectance intensity value and the infrared radiation value, generating a composite image with multi-dimensional information, which serves as dedicated input data for the neural network.

[0028] The shared encoder extracts multi-scale fusion feature units that receive multi-channel registered images. It extracts image features level by level through multiple serial convolutional layers. Shallow convolutions capture detailed features such as coating edges and cardboard textures, while deep convolutions extract overall features of large-area coating regions. It outputs multiple sets of multi-scale feature maps with progressively decreasing resolution and increasing channel count, retaining features from each level throughout the process for subsequent skip-layer fusion calculations. The mask decoder generates a coating presence confidence map signal, and the thickness decoder generates a coating relative thickness map signal, forming parallel prediction branches. Both decoders employ a top-down, progressively upsampling structure. After each upsampling stage, the corresponding scale feature map from the shared encoder is retrieved to complete skip-layer stitching and fusion. At the end of the mask decoder, a coating confidence image with a value range of 0 to 1 is output through single-channel convolution combined with an sigmoid activation function to distinguish between areas with and without coating. The thickness decoder outputs a relative thickness image without value range constraints through single-channel convolution combined with identity mapping activation to characterize coating thickness variations. The two decoders share encoder features but their convolution parameters are independent, simultaneously completing two tasks: region segmentation and thickness regression. The two-way prediction image synchronous output unit summarizes the confidence map and relative thickness map generated by the parallel branch, and transmits the two sets of prediction images synchronously to the subsequent mask generation and thickness correction process. The entire process relies on parallel branches to realize image correction and dual-task synchronous inference. Each operation unit is connected layer by layer, fully covering all algorithm logic of multi-sensor image fusion and multi-task deep learning inference.

[0029] The parameter learning of the aforementioned multi-task convolutional neural network during the training phase is achieved by minimizing a joint loss function. This joint loss function is a weighted combination of the loss terms from two branches. The mask branch uses a binary cross-entropy loss function, which compares the network-predicted coating existence confidence map with the real binary mask obtained by manual annotation or offline precise measurement pixel by pixel, measuring the difference between the two probability distributions. The thickness branch uses a smooth absolute bias loss function. That is, when the absolute value of the difference between the predicted thickness value and the labeled true thickness value of a single pixel is less than a preset smooth transition interval width, a squared loss is used to ensure a smooth gradient transition, which is differentiable near zero; when the difference is greater than or equal to this width, a linear loss is used to reduce the sensitivity to outliers or annotation noise. The weighting coefficients of the two loss terms are reasonably set according to the numerical magnitude of their respective losses in the early stage of training, so that during backpropagation, the contributions of the encoder and the two decoders to the parameter update gradients from the two tasks are roughly equal, avoiding one task dominating the training process and causing the other task to converge insufficiently. By simultaneously receiving gradient backpropagation from supervision signals from two tasks via a shared encoder, the network can spontaneously extract common features that are helpful for both coating presence identification and coating thickness regression during the learning process, thereby achieving knowledge sharing and mutual promotion among multiple tasks.

[0030] The conversion from the relative thickness map signal to the absolute physical thickness value is achieved through linear mapping parameters determined in a pre-calibration process. During calibration, a calibration cardboard board with a known continuously varying coating thickness is selected. Its multi-channel registration image signal is acquired on the detection system and fed into a network that has undergone preliminary training or fine-tuning, resulting in the relative thickness map of the calibration board. Simultaneously, using a high-precision contact or non-contact thickness measurement device, the actual coating thickness is measured at several sampling points on the calibration board to obtain the true spatial distribution map of the coating thickness. After spatially aligning the true thickness map with the relative thickness map output by the network, a linear regression is performed using the relative thickness value at the sampling point as the independent variable and the true thickness value as the dependent variable, employing the least squares method to obtain the slope and intercept parameters. During online detection, simply multiplying each pixel value in the relative thickness map signal output by the network by the slope and adding the intercept yields the coating thickness map signal representing the absolute coating thickness.

[0031] After obtaining the coating presence confidence map signal and the coating relative thickness map signal, a binary mask needs to be generated based on the confidence map to distinguish between foreground and background pixels of the coating. The method for determining the binarization threshold is as follows: First, the grayscale histogram of the entire coating presence confidence map signal is statistically analyzed. Then, by traversing all possible grayscale values ​​as candidate thresholds, the inter-class variance of the foreground pixel set and background pixel set under each candidate threshold is calculated. Based on the maximum inter-class variance criterion, the optimal threshold that maximizes the distinction between the foreground and background is automatically selected as the adaptive segmentation threshold. Pixels greater than or equal to the threshold are considered foreground pixels of the coating and assigned a logical true value; pixels less than the threshold are considered background pixels and assigned a logical false value, thus obtaining the initial binary mask. Next, morphological closing and opening operations are sequentially performed on the initial binary mask. The closing operation fills the tiny holes inside the coating area caused by local fluctuations in confidence through a dilation-erosion operation. The opening operation removes isolated false detection points caused by random noise in the background area through an erosion-dilation operation, resulting in a spatially continuous and smooth-boundary final binary mask signal.

[0032] like Figure 4As shown in the flowchart, this process implements full numerical calculation from the predicted coating image to the uniformity quantification index. It incorporates a window validity determination branch and a window sliding traversal loop. The entire process unit sequentially receives the confidence map and the coating relative thickness map signal, generates a binary mask signal through adaptive thresholding and morphological processing, obtains a corrected thickness map signal through pixel-by-pixel multiplication and connected component filtering, divides the image into rectangular statistical windows, calculates the mean and standard deviation of the thickness for each window, determines whether the window meets the valid statistical conditions, stores the standard deviation of the valid window, slides the window through the entire image region, determines whether all windows have been traversed, and calculates the average of the valid standard deviations to output the coating uniformity result. Each unit corresponds to key algorithm steps such as coating region selection, interference removal, and windowed statistical quantization. The unit receiving the confidence map and the coating relative thickness map signal receives two sets of predicted images output from the neural network inference process, synchronously caching the entire pixel value matrix as the basis for coating region selection and thickness correction calculations, completely preserving all predicted information on coating distribution and thickness variation within the image. The adaptive thresholding and morphological processing method generates a binary mask signal unit. First, it calculates the gray-level histogram of the confidence image. Then, it traverses all candidate thresholds based on the maximum inter-class variance criterion and selects the adaptive segmentation threshold corresponding to the maximum inter-class variance. According to the threshold, it divides the foreground pixels of the coating and the background pixels of the cardboard to generate an initial mask. Then, it sequentially performs morphological closing operations to fill the small holes inside the coating and morphological opening operations to remove isolated false detection pixels in the background. The output is a standard binary mask without noise or holes, which accurately delineates the boundary of the effective coating area.

[0033] The unit that obtains the corrected thickness map signal by pixel-by-pixel multiplication and connected component filtering multiplies the binary mask with the relative thickness image pixel by pixel, directly setting the thickness values ​​of the background region to zero, and retaining only the coating foreground thickness data to generate a preliminary thickness map. Then, connected component analysis is performed on the preliminary thickness map to identify all continuous coating pixel regions and calculate the pixel area of ​​each region. Small interfering connected components with areas smaller than a preset minimum region threshold are all set to zero, eliminating sporadic false thickness responses caused by dust and optical path noise, and outputting an interference-free corrected thickness image. The image segmentation rectangular statistical window unit uniformly divides the entire corrected thickness image into rectangular statistical windows of uniform size, supporting window sliding and overlapping segmentation methods. The window step size is smaller than the window side length to increase local statistical density, cover typical spatial scales of coating defects, and balance the statistical accuracy of random noise and local thickness fluctuations. The single-window thickness mean and standard deviation calculation unit traverses all effective coating pixels within a single window, calculating the arithmetic mean and standard deviation of the thickness within the window. The standard deviation value directly reflects the coating thickness fluctuation range within the window. The window validation node checks the total number of valid coating pixels and their standard deviation within the window. Windows with fewer than a preset threshold of pixels or a standard deviation of zero are considered invalid and the process jumps directly to the window sliding loop traversal unit, discarding the statistical data for that set. Valid windows that meet the statistical criteria are stored in the valid window standard deviation storage unit, where the current window's standard deviation value is retained for subsequent mean calculation. The window sliding loop traverses the entire image region unit, moving the statistical window horizontally and vertically according to a preset step size. After completing a single-window calculation, it automatically generates the next set of window coordinates, continuously performing single-window statistical operations in a loop. The window validation node determines whether all image regions have been traversed. If not, the process jumps to the image division rectangular statistical window unit to start a new round of window calculation, forming a complete sliding window loop. After all windows have been traversed, the process enters the valid standard deviation averaging output coating uniformity result unit, which summarizes all retained valid window standard deviation values ​​and calculates the arithmetic mean. This average value is the quantitative result representing the overall coating uniformity. The entire process integrates branch judgment and cyclic sliding window statistical logic, and there is no single linear calculation path. Each level of operation unit eliminates image interference and quantifies coating thickness fluctuations, thus fully realizing the entire numerical calculation process from the predicted image to the final uniformity index.

[0034] The binary mask signal and the coating relative thickness map signal are multiplied pixel-by-pixel. That is, for each spatial location in the image, the value of the binary mask is multiplied by the corresponding thickness value. At background pixel locations, since the mask value is zero, the multiplication result is zero; at coating pixel locations, the mask value is the logical truth value, and the thickness value is preserved. This operation achieves accurate separation of the coating region from the background region, forming a preliminary thickness map containing only the coating region. To further eliminate interference from small, sporadic responses, connected component analysis is performed on the preliminary thickness map, marking each unconnected white region and calculating its pixel area. The thickness values ​​of all pixels within connected regions whose pixel area is less than a preset minimum coating region area threshold are forcibly set to zero. This area threshold is set according to the minimum spatial scale of coating defects that need to be identified in actual production, filtering out tiny spots that obviously cannot constitute an effective coating, ultimately obtaining a spatially pure corrected thickness map signal.

[0035] When quantifying coating uniformity based on the corrected thickness map signal, the effective area of ​​the entire image is uniformly divided into several equal-sized, non-overlapping rectangular windows along the row and column directions. The window size is reasonably selected based on the typical spatial scale of coating uniformity defects. Typically, the window side length needs to cover several times the width of the expected local coating defect, thus capturing spatial variation information reflecting regional coating thickness fluctuations while effectively suppressing random noise in individual pixels. For each rectangular window, the thickness values ​​of all pixels within it are statistically analyzed, and the arithmetic mean and standard deviation of the window's thickness are calculated. To exclude windows that cannot provide reliable statistics due to being located at the edge of the cardboard or having too few effective coating pixels, a preset effective pixel count threshold is set, filtering out windows with a total pixel count below this threshold; simultaneously, windows with a standard deviation of zero are also filtered out, as these windows are usually located in completely uncoated pure background areas and do not reflect coating uniformity. The arithmetic mean of the standard deviations of all valid windows that pass the filtering is taken; this mean is the coating uniformity quantification result signal. The smaller the value of this signal, the smaller the coating thickness fluctuation amplitude in each local area, and the better the overall coating uniformity. When comparing products with different nominal coating thicknesses, the uniformity index can be calculated by using the global thickness variation coefficient, which is the ratio of the standard deviation to the arithmetic mean of all pixel thickness values. In order to prevent division by zero anomalies, a preset maximum deviation value can be directly assigned when the arithmetic mean is too low.

[0036] like Figure 5As shown, this invention also provides a cardboard surface coating uniformity detection system based on multi-sensor fusion, comprising: a synchronous triggering and image acquisition module, including a rotary encoder, a synchronous triggering controller, a linear array visible light camera, and a linear array infrared thermal imager; the rotary encoder is mechanically coupled to the conveyor line to output displacement pulse signals; the synchronous triggering controller generates row trigger signals based on the displacement pulse signals and simultaneously sends them to the two cameras, enabling the two cameras to acquire visible light image row signals and infrared image row signals of the same physical row on the cardboard surface row by row, accumulating to form spatially corresponding visible light image signals and infrared image signals; a multispectral image preprocessing module, used to perform geometric registration of the visible light image signals and infrared image signals to eliminate parallax and scale differences, and stack the registered images in the channel dimension to generate multi-channel registered image signals; and a multi-task neural network inference module, which... The system deploys a pre-trained multi-task convolutional neural network, which includes a shared encoder, a mask decoder, and a thickness decoder. The shared encoder extracts multi-scale features from the multi-channel registered image signal. The mask decoder generates a coating presence confidence map signal based on the multi-scale features, and the thickness decoder generates a coating relative thickness map signal based on the multi-scale features. A mask generation and thickness correction module performs adaptive threshold binarization and morphological filtering on the coating presence confidence map signal to obtain a binary mask signal, and multiplies the binary mask signal with the coating relative thickness map signal pixel by pixel to obtain a corrected thickness map signal. A windowing statistics and uniformity output module divides the corrected thickness map signal into multiple non-overlapping rectangular windows, calculates the standard deviation of the coating thickness value in each window, and takes the arithmetic mean of the standard deviations of all windows as the coating uniformity measurement result signal output.

[0037] The present invention provides a method and system for detecting the uniformity of coatings on cardboard surfaces based on multi-sensor fusion. This method utilizes a rotary encoder to simultaneously trigger a linear array visible light camera and a linear array infrared thermal imager, acquiring spatially corresponding images of the cardboard surface line by line. After geometric registration and channel stitching, these images are input into a pre-constructed multi-task convolutional neural network. This network extracts features using a shared encoder, and a mask decoder generates a coating presence confidence map, while a thickness decoder generates a coating relative thickness map. The confidence scores are then analyzed. Figure 2 After value-based scaling, the thickness map is multiplied pixel by pixel to obtain a corrected thickness map containing only the coating area. Then, the mean of the standard deviation of the thickness of each window is statistically analyzed by dividing the map into windows to output the coating uniformity quantification result.

[0038] Therefore, the method and system for detecting the uniformity of coatings on cardboard surfaces based on multi-sensor fusion of the present invention can solve the problem of the inability to reliably detect the uniformity of transparent adhesive coatings online and in a non-contact manner.

[0039] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A method for detecting the uniformity of coating on cardboard surfaces based on multi-sensor fusion, characterized in that, include: S1: The displacement pulse signal generated by the rotary encoder triggers the synchronous trigger controller, which generates a line trigger signal and sends it to the linear array visible light camera and the linear array infrared thermal imager. It collects the visible light image line signal and the infrared image line signal corresponding to the space, and accumulates them to form the visible light image signal and the infrared image signal. S2: Perform geometric registration and stitching on the visible light image signal and infrared image signal to obtain a multi-channel registered image signal; S3: Input the multi-channel registered image signal into a multi-task convolutional neural network, extract features by a shared encoder, generate a coating presence confidence map signal by a mask decoder, and generate a coating relative thickness map signal by a thickness decoder. S4: The confidence map signal of the coating is binarized to obtain a binary mask signal, which is then multiplied pixel by pixel with the relative thickness map signal of the coating to obtain a corrected thickness map signal; S5: Calculate the thickness standard deviation by dividing the window according to the modified thickness map signal, and take the average standard deviation of each window as the coating uniformity quantification result signal.

2. The method according to claim 1, characterized in that, The geometric registration described in S2 includes: correcting the distortion of the visible light image signal using pre-calibrated intrinsic parameters of a linear array visible light camera, and correcting the distortion of the infrared image signal using pre-calibrated intrinsic parameters of a linear array infrared thermal imager; scaling and translation transformations of the distortion-corrected infrared image signal based on pre-calibrated relative spatial pose parameters between the two sensors, so that each pixel corresponds to the same position on the cardboard surface as the visible light image signal; the stitching involves stacking the registered visible light image signal and the registered infrared image signal in the channel dimension, so that each pixel position of the multi-channel registered image signal contains a visible light intensity channel value and at least one infrared radiation channel value. The calculation formula for the geometric transformation mapping of the infrared image signal is as follows: ， in, To match the infrared pixel coordinate vectors of the visible light pixel positions after registration, k is the global scale scaling factor of the infrared image, and H is the relative spatial pose homography matrix between the linear array visible light camera and the infrared thermal imager. This is a mapping function for infrared image distortion correction. This represents the original infrared pixel coordinate vector before distortion correction. To calibrate the intrinsic parameter matrix of the infrared thermal imager, This is the infrared distortion coefficient vector. For a two-dimensional vector, the 2-norm operator, The d-th order infrared distortion coefficient, Operators for solving matrix determinants This is the x-axis translation offset. This is the y-axis translation offset. Construct an operator for a two-dimensional column vector, where d is the distortion order traversal variable.

3. The method according to claim 1, characterized in that, The multi-task convolutional neural network consists of a shared encoder, a mask decoder, and a thickness decoder. The shared encoder contains multiple serially connected convolutional layers, which extract features from the multi-channel registered image signal step by step, outputting multi-scale feature maps with progressively decreasing spatial resolution and progressively increasing channel number. Both the mask decoder and the thickness decoder adopt a top-down progressive upsampling structure. At each upsampling stage, the feature map from the previous stage is upsampled and then concatenated with the feature map of the corresponding scale in the shared encoder through skip connections to achieve channel dimension concatenation, followed by convolutional operations to achieve feature fusion. The output layer of the mask decoder generates a coating presence confidence map signal through a single-channel convolutional layer and an sigmoid nonlinear activation function that maps the output value to the zero-to-one interval. The output layer of the thickness decoder generates a coating relative thickness map signal through a single-channel convolutional layer and an identity mapping activation function. The mask decoder and the thickness decoder share multi-scale features from the shared encoder during each upsampling stage, but each has independent convolution parameters.

4. The method according to claim 3, characterized in that, The multi-task convolutional neural network employs a joint loss function for parameter optimization during training. This joint loss function is defined as a weighted sum of the mask branch loss and the thickness branch loss. The mask branch loss is calculated using a binary cross-entropy loss based on the coating existence confidence map signal and pre-acquired true binary mask annotations. The thickness branch loss is calculated using a smoothed absolute deviation loss based on the coating relative thickness map signal and pre-acquired true relative thickness map annotations. This smoothed absolute deviation loss is squared when the absolute value of the difference between the predicted and labeled values ​​of a single pixel is less than a preset switching threshold, and linear when it is greater than or equal to the switching threshold. The weighting coefficients of the joint loss function are set according to the order-of-magnitude ratio of the mask branch loss and the thickness branch loss in the early stages of training, ensuring that the gradient magnitudes of the two losses match during backpropagation. During training, all learnable parameters in the shared encoder, mask decoder, and thickness decoder are synchronously updated using the backpropagation algorithm. The calculation formula for the joint loss function is as follows: ， ， ， in, Here, λ represents the overall joint loss value of the multi-task convolutional neural network, λ is the weighting coefficient of the mask branch loss, and N is the total number of pixels in the multi-channel registered image. The single-pixel binary cross-entropy loss function is... The actual coating binary mask annotation value for the i-th pixel. Let α be the confidence prediction value for the existence of the coating at pixel i, and let α be the weighting coefficient for the thickness branch loss. The single-pixel smoothing absolute deviation loss function is... This is the predicted relative thickness of the coating for the i-th pixel. This represents the actual relative thickness value of the i-th pixel. This is a conditional indicator function.

5. The method according to claim 3, characterized in that, The value of each pixel in the coating relative thickness map signal output by the thickness decoder has a positive linear mapping relationship with the physical thickness of the adhesive coating on the cardboard surface at that pixel location. The positive linear mapping relationship is determined through a pre-calibration process, which includes: acquiring a calibration cardboard board with a known area of ​​continuous coating thickness variation; acquiring its multi-channel registration image signal on the detection system and inputting it into the multi-task convolutional neural network to obtain a coating relative thickness map signal; simultaneously acquiring the actual coating thickness spatial distribution map of the calibration cardboard board using a high-precision thickness measurement device; spatially aligning the actual coating thickness spatial distribution map with the coating relative thickness map signal; taking several sampling points with the coating relative thickness value as the independent variable and the actual coating thickness value as the dependent variable to perform linear regression, obtaining the slope and intercept parameters of the linear transformation and storing them. During online detection, the pixel values ​​in the coating relative thickness map signal are multiplied by the slope and the intercept is added to obtain a coating thickness map signal that characterizes the absolute coating thickness. The corrected thickness map signal is calculated based on this coating thickness map signal.

6. The method according to claim 1, characterized in that, The binarization process of the coating confidence map signal in S4 is specifically as follows: statistically analyze the gray-level histogram of the coating confidence map signal, use the maximum inter-class variance criterion to traverse possible gray-level values ​​as candidate thresholds, calculate the inter-class variance between the foreground pixel class and the background pixel class under each candidate threshold, and select the candidate threshold that maximizes the inter-class variance as the adaptive segmentation threshold. Pixels whose values ​​in the confidence map signal of the coating are greater than or equal to the adaptive segmentation threshold are identified as foreground pixels of the coating and assigned a first logical value in the binary mask. Pixels whose values ​​are less than the adaptive segmentation threshold are identified as background pixels and assigned a second logical value in the binary mask to form an initial binary mask. Then, morphological closing and morphological opening operations are sequentially performed on the initial binary mask to fill the holes inside the coating area caused by confidence fluctuations and remove isolated false detection points in the background area, thereby obtaining the binary mask signal.

7. The method according to claim 6, characterized in that, The process of obtaining the corrected thickness map signal in S4 includes: performing pixel-by-pixel multiplication between the binary mask signal and the coating relative thickness map signal, so that the thickness value at the position corresponding to the background pixel in the binary mask signal is set to zero, while the thickness value at the position corresponding to the foreground pixel of the coating is retained, generating a preliminary thickness map containing only the coating region; performing connected component analysis on the preliminary thickness map, marking each connected region and calculating the pixel area of ​​each connected region, setting the thickness value of all pixels in the connected regions whose pixel area is less than the preset minimum coating region area to zero, the preset minimum coating region area being determined according to the lower limit of the spatial scale required for coating defect identification, thereby eliminating sporadic interference responses and obtaining the corrected thickness map signal.

8. The method according to claim 1, characterized in that, In S5, the process of dividing the window, calculating the thickness standard deviation, and taking the average is as follows: The pixel area covered by the corrected thickness map signal is uniformly divided into several equal-sized, non-overlapping rectangular windows along the row and column directions; for each rectangular window, the thickness values ​​of all pixels falling within the window are counted, and the arithmetic mean and standard deviation of the thickness values ​​within the window are calculated; after filtering out invalid windows whose total number of pixels is less than a preset effective pixel count threshold or whose standard deviation is zero, the arithmetic mean of the standard deviations of all valid windows is calculated, and this arithmetic mean is output as the coating uniformity measurement result signal. The formula for calculating the arithmetic mean is as follows: ， ， ， ， Where U represents the final uniformity measurement result of the coating, and M represents the total number of valid rectangular windows after filtering. W represents the set of all valid windows, where W is the set of all rectangular windows used to partition the image. Let w be the standard deviation of the coating thickness within the w-th window. The number of effective pixels of the coating within the w-th window. Correct the absolute coating thickness value for the j-th pixel in the w-th window. The arithmetic mean of the coating thickness within the w-th window. This is a function to indicate the validity of a window. The preset effective pixel count threshold is defined, w is the window traversal variable, and j is the pixel traversal variable within the window.

9. The method according to claim 8, characterized in that, The size of the rectangular window is set according to the typical spatial scale range of the uniformity defect of the cardboard adhesive coating, so that the side length of a single rectangular window covers several times the width of the expected local coating defect, in order to retain the standard deviation information reflecting the regional coating thickness fluctuation while suppressing random noise; when dividing the window, the adjacent windows can also be partially overlapped, and the adjacent windows slide in both the row and column directions with a preset step size, which is smaller than the side length of the rectangular window, in order to increase the local statistical density.

10. A system for detecting the uniformity of coating on cardboard surfaces using the multi-sensor fusion method according to any one of claims 1-9, characterized in that, include: The synchronous triggering and image acquisition module includes a rotary encoder, a synchronous triggering controller, a linear array visible light camera, and a linear array infrared thermal imager. The rotary encoder is mechanically coupled to the conveyor line and outputs a displacement pulse signal. The synchronous triggering controller generates a row trigger signal based on the displacement pulse signal and sends it to the two cameras simultaneously, so that the two cameras acquire the visible light image row signal and infrared image row signal of the same physical row on the surface of the carton line one by one, and accumulate them to form the spatially corresponding visible light image signal and infrared image signal. The multispectral image preprocessing module is used to perform geometric registration on the visible light image signal and the infrared image signal to eliminate parallax and scale differences, and to stack the registered images in the channel dimension to generate a multi-channel registered image signal. The multi-task neural network inference module is equipped with a pre-trained multi-task convolutional neural network. The network includes a shared encoder, a mask decoder, and a thickness decoder. The shared encoder extracts multi-scale features from the multi-channel registered image signal. The mask decoder generates a coating presence confidence map signal based on the multi-scale features. The thickness decoder generates a coating relative thickness map signal based on the multi-scale features. The mask generation and thickness correction module is used to perform adaptive threshold binarization and morphological filtering on the coating existence confidence map signal to obtain a binary mask signal, and multiply the binary mask signal with the coating relative thickness map signal pixel by pixel to obtain a corrected thickness map signal; The windowed statistics and uniformity output module is used to divide the corrected thickness map signal into multiple non-overlapping rectangular windows, calculate the standard deviation of the coating thickness value in each window, and take the arithmetic mean of the standard deviations of all windows as the coating uniformity quantification result signal output.

Images