Ultra-wide printing paper defect online detection system and method

Abstract

The invention provides an ultra-wide printing paper defect online detection system, the system comprises a high-speed image acquisition unit, an edge AI calculation unit and an MSAF-ACPM model deployed on the edge AI calculation unit, the model captures image features in different receptive fields in parallel through a multi-scale feature extractor, and the image features are dynamically weighted and fused by a scale attention fusion module to obtain an image feature fusion model; and then local defect judgment is corrected by utilizing global printing state information through a self-adaptive context sensing module, so that real paper defects and environment interference artifacts are distinguished, and accurate and real-time online detection of multi-scale and complex-form paper defects in the ultra-wide printing process is realized; the method has the remarkable advantages in the aspects of detection precision, real-time processing capacity and system robustness. Meanwhile, the invention further provides a model training method and a detection method applying the system.

Description

Technical Field

[0001] This invention relates to an online defect detection system for ultra-wide-format printed paper, suitable for quality control in high-speed printing production lines with a width greater than 2 meters and a line speed of up to several hundred meters per minute. Background Technology

[0002] Currently, the global printing industry is undergoing an unprecedented wave of digital transformation and industrial upgrading. Particularly in packaging printing, newsprint printing, and high-end commercial printing, the demand for wider printing widths continues to grow, with ultra-wide-format printing equipment exceeding 2 meters becoming the industry mainstream. These machines, with their extremely high production efficiency, can print large sizes of paper or film in a single pass, significantly reducing production costs and shortening delivery cycles. However, with the rapid expansion of printing widths and the continuous increase in linear speeds (many ultra-wide-format printing presses have linear speeds exceeding 300 meters per minute), online real-time inspection of print quality faces unprecedented challenges. Paper defects, namely various defects in printed materials such as ink spots, scratches, mold spots, creases, stains, color differences, missing or excessively large dots, can cause enormous losses if they are not detected and identified in time during high-speed production.

[0003] To address this challenge, the industry has developed various online paper defect detection technologies. Among them, traditional machine vision-based inspection systems are currently the most widely used solution. These systems typically employ high-resolution linear image sensors, coupled with LED light sources, to scan the surface of printed materials. Pre-set image processing algorithms, such as threshold segmentation, edge detection, and template matching, are used to identify areas that differ from normal printed images. However, for printing formats larger than 2 meters wide, traditional machine vision systems have significant limitations.

[0004] First, the bandwidth bottleneck in data acquisition and processing is the primary pain point. A printing press with a width of 2 meters, operating at a linear speed of 300 meters per minute, generates an extremely large amount of image data per second. Traditional data acquisition cards and image processing units struggle to complete image acquisition, transmission, preprocessing, and defect detection in real time under such high bandwidth. This results in the detection system being unable to fully keep up with the printing press's linear speed, or having to sacrifice some image resolution and detection accuracy for real-time performance, making it easy to miss minute or low-contrast paper defects.

[0005] Secondly, insufficient ability to identify complex and varied paper defects is another key pain point. Traditional image processing algorithms have poor robustness when faced with complex scenarios such as uneven lighting, variations in paper texture, and fluctuations in printing processes. They rely on fixed thresholds or templates, resulting in low recognition rates for unstructured and diverse paper defects, and are prone to false positives or false negatives.

[0006] Third, the application of deep learning in ultra-wide format print inspection still faces "computing power walls" and "data walls." Although deep learning, especially convolutional neural networks (CNNs), has achieved great success in image recognition, effectively deploying it in ultra-wide format, high-speed online detection systems still faces many challenges. Training an accurate deep learning model requires massive, precisely labeled defect datasets, which are costly to collect and label. Meanwhile, deploying complex deep learning models to real-time detection systems requires powerful computing capabilities. In ultra-wide format, high-speed scenarios, traditional GPU servers may not meet real-time requirements, or the model structure may need to be significantly simplified, sacrificing performance. Furthermore, the generalization ability of existing models across different printing batches and paper types has not yet been fundamentally resolved.

[0007] In summary, existing technical solutions suffer from varying degrees of technical bottlenecks in terms of detection accuracy, real-time performance, robustness, and deployment cost. Therefore, there is an urgent need for an ultra-wide-format, high-speed online defect detection solution for printed paper that can balance high accuracy, high real-time performance, strong robustness, and efficient deployment. Summary of the Invention

[0008] To address these issues, the present invention provides an online defect detection system for ultra-wide-format printed paper, thereby solving the aforementioned technical problems.

[0009] An online defect detection system for ultra-wide-format printing is applied to the online detection of paper defects in high-speed printing processes with a width greater than 2 meters and a linear speed greater than 300 meters per minute. The system includes: A high-speed image acquisition unit is configured to simultaneously acquire high-resolution image data covering the entire printing width during the printing process; An edge AI computing unit is communicatively connected to the high-speed image acquisition unit and is configured to preprocess and infer the acquired image data. as well as, The MSAF-ACPM model is deployed in the edge AI computing unit; The MSAF-ACPM model includes: The multi-scale feature extractor (MSFE) is used to receive preprocessed image data and generate multiple feature maps at different receptive fields or resolutions in parallel. The Scale Attention Fusion (SAFM) module is connected to the output of the multi-scale feature extractor. It is used to evaluate the intra-scale importance of feature maps at each scale and assign weights between scales to generate a unified feature representation after fusion. An adaptive context-aware module (ACPM) is connected to the output of the scale attention fusion module and is configured to receive the unified feature representation and global context information, and dynamically adjust the response intensity of local defect features through a context gating mechanism. The Defect Classification and Location Head (DCLH), connected to the output of the Adaptive Context Aware Module, is used to predict the type, location bounding box, and confidence level of paper defects.

[0010] The multi-scale feature extractor (MSFE) includes multiple parallel branches, wherein at least one branch employs a shallow convolutional network to capture high-frequency detail features, at least one branch employs a deep network with dilated convolutions to capture low-frequency global features, and / or at least one branch integrates a lightweight Transformer encoder to model long-range dependencies.

[0011] In the Scale Attention Fusion (SAFM) module, the intra-scale importance evaluation is achieved by generating an attention mask by applying convolution and sigmoid activation on each scale feature map; the inter-scale weight allocation is obtained by globally pooling each scale feature map and then calculating it through a fully connected layer and a Softmax function.

[0012] The global context information in the Adaptive Context Aware Module (ACPM) is provided by a separately deployed lightweight global analysis module. The global analysis module is a convolutional neural network or statistical feature extractor with a computational complexity lower than that of the MSAF-ACPM model backbone network. The global context information extracted by the module includes the average brightness of the entire image, texture variance statistical features, and / or real-time process parameters from the printing press control system, including linear speed and tension.

[0013] The edge AI computing unit includes an FPGA acceleration card, which is configured to perform image bad pixel correction, gain correction, and shallow convolution operations in the multi-scale feature extractor (MSFE).

[0014] The edge AI computing unit is further configured to divide the preprocessed ultra-wide image data into multiple data blocks with overlapping regions along the width direction. The overlapping regions are used to ensure the integrity detection of cross-block defects and eliminate the boundary effect of model inference, and to allocate the data blocks to multiple GPUs for parallel inference.

[0015] This invention also provides a training method for an MSAF-ACPM model, used to train the model described above. The method includes the following steps: Dataset construction steps: Collect ultra-wide format printed images covering various paper types, inks, and printing conditions, and annotate paper defects at the pixel level; Data augmentation steps: Randomly rotate, flip, scale, and adjust the brightness / contrast of the dataset, and use defect synthesis technology to overlay the labeled paper defect samples onto the normal background image to expand the training samples; Model training steps: An end-to-end supervised learning approach is adopted to optimize model parameters under the guidance of a combined loss function, which is a weighted sum of Focal Loss and Dice Loss; Model quantization steps: After training, the model is quantized using INT8 and a quantization-aware training strategy is adopted to minimize accuracy loss.

[0016] This invention also provides an online detection method for printing paper defects based on the MSAF-ACPM model. Using the aforementioned system, the method includes the following steps: Image acquisition step: The high-speed image acquisition unit acquires image data covering the printing width at a frame rate matching the printing line speed; Hardware-accelerated preprocessing steps: On the FPGA of the edge AI computing unit, bad pixel correction, gain correction and shallow feature extraction are performed on the raw image data; Data segmentation and parallel inference steps: The preprocessed image is divided into data blocks with overlapping regions and distributed to multiple GPUs. The deployed MSAF-ACPM model performs multi-scale feature extraction, scale attention fusion, context awareness and defect localization in parallel. Results fusion and response steps: The detection results of each data block are fused on the CPU to eliminate boundary effects, and the final detection results are sent to the printing press control system to trigger alarm, marking, or rejection instructions.

[0017] In the result fusion and response step, the printing press control system performs a graded response based on the severity of the received defect information: for minor defects, only logs are recorded; for serious defects, an audible and visual alarm is triggered, inkjet marking is activated, or the paper cutter is driven to discard the defective product.

[0018] The method further includes a trend analysis and early warning step: continuously collecting defect data, and when the density or frequency of a specific type of defect exceeds a threshold within a preset time, sending an early warning signal of process parameter drift or equipment failure to the printing press control system.

[0019] Beneficial Effects: This invention provides an online detection system for ultra-wide-format printing paper defects based on multi-scale attention fusion and adaptive context awareness. The system includes a high-speed image acquisition unit, an edge AI computing unit, and an MSAF-ACPM model deployed on top of it. The model captures image features in parallel under different receptive fields through a multi-scale feature extractor, and then dynamically weights and fuses these features using a scale attention fusion module. Finally, an adaptive context awareness module uses global printing state information to correct local defect judgments, distinguishing between genuine paper defects and environmental interference artifacts. The system employs a data block and parallel inference architecture, combined with FPGA hardware-accelerated preprocessing, effectively overcoming the real-time bottleneck in ultra-wide-format high-speed scenarios. This invention achieves accurate, real-time online detection of multi-scale, complex-shaped paper defects during ultra-wide-format printing, achieving significant advantages in detection accuracy, real-time processing capability, and system robustness. Furthermore, this invention provides a model training method and a detection method applying the system. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to specific embodiments.

[0021] This invention provides an online defect detection system for ultra-wide format printing paper (MSAF-ACPM) based on multi-scale attention fusion and adaptive context awareness, aiming to overcome the shortcomings of existing ultra-wide format high-speed printing paper defect detection solutions in terms of accuracy, real-time performance, robustness, and deployment cost. This system achieves accurate, real-time, and online detection of complex, multi-scale paper defects by organically integrating multi-scale feature learning, adaptive attention mechanisms, context awareness technology, and an efficient hardware and software co-architecture.

[0022] The system includes a high-speed image acquisition unit, an edge AI computing unit, and an MSAF-ACPM model deployed on the edge AI computing unit.

[0023] The high-speed image acquisition unit is configured on the printing production line to acquire raw image data covering the entire printing width. Specifically, the high-speed image acquisition unit includes a high-resolution line scan camera array and a synchronous illumination system. In this embodiment, the high-resolution line scan camera array includes eight CXP-12 interface line scan cameras densely arranged along the printing width direction, with a total pixel width covering a 2.5-meter width and sampling 10 pixels per millimeter. The camera frame rate is matched to the printing line speed, reaching approximately 6.67 kHz at a line speed of 400 meters per minute. The synchronous illumination system is a high-brightness, uniform LED light source array, ensuring flicker-free illumination across the entire printing width.

[0024] The edge AI computing unit, deployed at the edge of the production line, provides real-time data processing and analysis. In one specific embodiment, this edge AI computing unit is a high-performance embedded computing platform with the following hardware configuration: FPGA accelerator card: using the Xilinx Virtex UltraScale+ series, used to offload preprocessing tasks such as image bad line / bad pixel correction, gain correction, and shallow convolution feature extraction; multiple high-performance GPUs: such as four NVIDIA H100s, providing parallel computing capabilities with a total computing power of several thousand TFLOPS (INT8); high-performance CPU: used for task scheduling, result fusion, and system control; large-capacity high-speed memory and storage: used to cache real-time data streams and models.

[0025] The MSAF-ACPM model is a hybrid architecture model that integrates multi-scale feature extraction, attention fusion, context awareness, and efficient reasoning capabilities. The MSAF-ACPM model comprises four collaboratively operating sub-modules: Multi-Scale Feature Extractor (MSFE), Scale Attention Fusion Module (SAFM), Adaptive Context Awareness Module (ACPM), and Defect Classification and Localization Head (DCLH).

[0026] The Multi-Scale Feature Extractor (MSFE) is configured with a parallel processing strategy, specifically including: a shallow branch based on early layers of ResNet-18 (conv1 to layer2) using 3x3 convolutional kernels to capture high-frequency details such as tiny ink spots and minor scratches; a deep branch based on later layers of ResNet-18 (layer3 to layer4) and introducing dilated convolutions with dilation rates of 2, 4, and 8 to capture low-frequency information such as large-area stains and creases; and a global branch integrating a two-layer lightweight Transformer encoder to model long-distance dependencies between the two ends of the image. Each branch outputs feature maps at different scales.

[0027] The Scale Attention Fusion (SAFM) module is configured with an adaptive fusion strategy, specifically including: Intra-scale attention: For each scale's feature map, an attention mask is generated using a 1x1 convolution + sigmoid function to highlight defect-sensitive regions within that scale; Inter-scale attention: Global average pooling is performed on each scale's feature map to obtain a global descriptor, which is then processed by a fully connected layer and a Softmax function to calculate the fusion weights for each scale. Finally, a weighted sum is obtained to obtain a unified fused feature map.

[0028] The Adaptive Context Aware Module (ACPM) is configured as a context gating mechanism, specifically including: an independent global analysis module (running on the CPU) that calculates statistical features such as average brightness and texture variance of the entire image in real time; the ACPM receives the fused feature map and global context features, and generates modulation factors through a gating unit (fully connected layer + sigmoid) to dynamically adjust the response intensity at each position in the fused feature map in order to suppress pseudo-defect responses caused by global illumination changes.

[0029] The Defect Classification and Localization Head (DCLH) adopts an anchor-based design similar to YOLOv7, predicting the offset, class probability, and confidence score of multi-scale anchor boxes on the enhanced feature map output by ACPM.

[0030] Furthermore, the MSAF-ACPM model is trained as follows: S11: Dataset Construction and Augmentation. 100,000 ultra-wide format printed images were collected, covering various materials such as coated paper and offset paper, and pixel-level annotations were performed by experts. Random rotation, flipping, brightness / contrast perturbation, and defect synthesis technology were used to overlay 20,000 real defect samples onto 50,000 normal backgrounds, ultimately forming 500,000 training samples. S12: Model Training and Optimization. ImageNet pre-trained weights were used for initialization. The optimizer was AdamW, with an initial learning rate of 1e. -4 Cosine annealing scheduling was employed. The loss function was a weighted sum of FocalLoss (α=0.75, γ=2) and Dice Loss (weight ratio 1:1). Training was conducted for 200 epochs using an early stopping strategy. S13: Model Quantization and Deployment. After training, Quantization-Aware Training (QAT) was used to convert the model to INT8 precision, resulting in a 4x reduction in model size, a 2.8x increase in inference speed, and a precision loss of <0.5%.

[0031] The model is deployed on an edge AI unit and performs detection as follows: S21: High-speed image acquisition. Eight linear scan cameras simultaneously acquire a line of 2.5-meter wide data (25,000 pixels) at a frame rate of 6.67kHz. S22: FPGA hardware-accelerated preprocessing. The FPGA performs bad pixel correction, gain correction, and completes convolution calculations for shallow branches of the MSFE. S23: Data partitioning and parallel inference. The preprocessed data is divided into 16 tiles with 50mm overlap and distributed to four H100 GPUs for parallel inference. The single-tile inference latency is ≤5ms. S24: Result fusion and production response. The CPU fuses the results of each tile to eliminate duplicate detections. The detection results (type, location, confidence level) are packaged and sent to the printing press main control system. The system performs a graded response according to preset rules: minor defects are logged; serious defects trigger alarms, inkjet printing, or scrapping. S25: Trend analysis and intelligent early warning. The system continuously analyzes the defect data stream. If the density of a certain type of defect (such as scratches) exceeds 5 per square meter within one hour, it sends an "equipment wear" warning to the main control system.

[0032] Furthermore, the detection system also includes a central control and user interface unit, integrated into the host computer in the main control room of the printing press. This unit provides real-time visualization, alarm linkage, historical data query, and system configuration functions.

[0033] The technical effects of the present invention will be described below with reference to specific embodiments.

[0034] Example 1: Configuration: For a printing line with a 2.5-meter width and a linear velocity of 400 meters per minute. High-speed image acquisition unit: 8 CXP-12 line scan cameras. Edge AI computing unit: 1 Xilinx FPGA + 4 NVIDIA H100 GPUs. MSAF-ACPM model: MSFE (ResNet-18 + dilated convolution + lightweight Transformer), SAFM, ACPM, DCLH (YOLOv7 header). Workflow: The system processes approximately 66,700 lines of data per second. Test results: 99.2% recall rate and 0.3% false positive rate for detecting 0.1mm diameter ink dots.

[0035] Comparative Example 1 (without SAFM module): The difference from Example 1 is that the MSAF-ACPM model removes SAFM and directly splices the features of each branch of MSFE. Test results: The recall rate for 0.1mm ink dots decreased to 92.5%, and the recall rate for 5mm stains decreased to 88.7%. The study shows that SAFM is crucial for adaptively handling multi-scale defects.

[0036] Example 2: Based on Example 1, a trend analysis and early warning function is added. When the system detects that scratch defects occur continuously within one hour and the density exceeds a threshold, it automatically sends an early warning to the main control system. Field verification showed that this function successfully provided an early warning of a guide roller bearing wear failure 3 hours in advance, avoiding batch quality problems.

[0037] The key parameters and performance comparisons of each embodiment and comparative example are shown in the table below:

[0038] The MSAF-ACPM system of this invention achieves a balance between detection accuracy, real-time performance, robustness, and engineering feasibility through innovative multi-scale attention fusion and context-aware architecture, combined with efficient hardware and software co-design, providing an advanced online quality inspection solution for ultra-wide format high-speed printing.

[0039] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made using the content of the present invention specification, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. An on-line inspection system for super-wide format printing paper, characterized by, The paper disease online detection system is applied to a high-speed printing process with a width greater than 2 meters and a line speed greater than 300 meters / minute, and comprises: a high-speed image acquisition unit configured to synchronously acquire high-resolution image data covering the entire printing width during the printing process; an edge AI computing unit in communication connection with the high-speed image acquisition unit and configured to pre-process and infer the acquired image data; and an MSAF-ACPM model deployed in the edge AI computing unit; the MSAF-ACPM model comprises: a multi-scale feature extractor (MSFE) for receiving pre-processed image data and generating feature maps at multiple different receptive fields or resolutions in parallel; a scale attention fusion module (SAFM) connected to the output end of the multi-scale feature extractor and configured to perform intra-scale importance evaluation and inter-scale weight distribution on the scale feature maps to generate a unified feature representation after fusion; an adaptive context perception module (ACPM) connected to the output end of the scale attention fusion module and configured to receive the unified feature representation and global context information and dynamically adjust the response strength of local defect features through a context gating mechanism; a defect classification and localization head (DCLH) connected to the output end of the adaptive context perception module and configured to predict the category, location bounding box and confidence of the paper disease.

2. The system of claim 1, wherein, The multi-scale feature extractor (MSFE) comprises multiple parallel branches, at least one of which adopts a shallow convolutional network to capture high-frequency detail features, at least one of which adopts a deep network with a hollow convolution to capture low-frequency global features, and / or at least one of which integrates a lightweight Transformer encoder to model long-distance dependencies.

3. The system of claim 1, wherein, In the scale attention fusion module (SAFM), the intra-scale importance evaluation is achieved by applying convolution and Sigmoid activation on each scale feature map to generate an attention mask; and the inter-scale weight distribution is calculated by globally pooling each scale feature map and passing it through a fully connected layer and a Softmax function.

4. The system of claim 1, wherein, The global context information in the adaptive context perception module (ACPM) is provided by a separately deployed lightweight global analysis module, which is a convolutional neural network or a statistical feature extractor with a lower computational complexity than the main network of the MSAF-ACPM model, and the extracted global context information includes average brightness, texture variance statistical features of the entire image, and / or real-time process parameters from the printing machine control system, including line speed and tension.

5. The system of claim 1, wherein, The edge AI computing unit comprises an FPGA acceleration card configured to perform image bad pixel correction, gain correction, and shallow convolution operations in the multi-scale feature extractor (MSFE).

6. The system of claim 1, wherein, The edge AI computing unit is also configured to divide the pre-processed super-wide image data into multiple data blocks with overlapping regions along the width direction, the overlapping regions being used to ensure the integrity detection of cross-block defects and eliminate the boundary effects of model inference, and to distribute the data blocks to multiple GPUs for parallel inference.

7. A method for training a MSAF-ACPM model, characterized in that, The method for training the MSAF-ACPM model according to any one of claims 1-6, comprising the following steps: A data set construction step: collecting super-wide format printing images under various paper types, inks and printing conditions, and pixel-level labeling of paper defects therein; A data augmentation step: randomly rotating, flipping, scaling, brightness / contrast adjusting the data set, and using defect synthesis technology to superimpose labeled paper defect samples on normal background images to expand training samples; A model training step: using end-to-end supervised learning to optimize model parameters under the guidance of a combined loss function, which is the weighted sum of Focal Loss and Dice Loss; A model quantization step: after training, the model is quantized to INT8, and a quantization-aware training strategy is used to minimize precision loss.

8. A method for online detection of print paper disease based on MSAF-ACPM model, characterized in that, The method for applying the system according to any one of claims 1-6, comprising the following steps: An image acquisition step: using the high-speed image acquisition unit to acquire image data covering the printing width at a frame rate matching the printing line speed; A hardware acceleration preprocessing step: performing bad pixel correction, gain correction and shallow feature extraction on the original image data on the FPGA of the edge AI computing unit; A data blocking and parallel inference step: dividing the preprocessed image into data blocks with overlapping areas, assigning them to multiple GPUs, and performing multi-scale feature extraction, scale attention fusion, context awareness and defect positioning by the deployed MSAF-ACPM model in parallel; A result fusion and response step: fusing the detection results of each data block on the CPU to eliminate boundary effects, and sending the final detection results to the printing machine control system to trigger alarm, marking or rejection instructions.

9. The method of claim 8, wherein, In the result fusion and response step, the printing machine control system performs hierarchical response according to the severity of the received defect information: only logs for minor defects; for serious defects, trigger audible and light alarms, start inkjet marking or drive the paper cutter to discard unqualified products.

10. The method of claim 8, wherein, The method further comprises a trend analysis and early warning step: continuously collecting defect data, and when the density or frequency of a specific type of defect exceeds the threshold within a preset time, sending a process parameter drift or equipment failure warning signal to the printing machine control system.

Images