Artificial board surface defect intelligent detection method and system based on machine vision

Abstract

The invention discloses an artificial board surface defect intelligent detection method and system based on machine vision, and particularly relates to the technical field of artificial board surface defect detection. An artificial board surface image is collected, image preprocessing, feature extraction, defect segmentation and classification recognition are carried out through a deep learning algorithm, a multi-scale convolutional neural network is adopted to carry out feature extraction on the image, a shallow convolutional layer captures small defect details, a deep convolutional layer recognizes global features of large defects, and a multi-scale convolutional neural network is adopted to carry out feature extraction on the image. Accurate defect segmentation is carried out through a Mask R-CNN model, a redundant frame is removed in combination with a non-maximum suppression algorithm, the classification problem of adjacent defects is corrected by using an error correction algorithm in combination with the spatial relationship and morphological characteristics of the defects, and defect information is fed back to a production line control system in real time; and defective products are automatically removed or production process parameters are automatically adjusted, so that the automation level of a production line is effectively improved, the product quality is optimized, and human intervention and production cost are reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of surface defect detection of wood-based panels, and particularly relates to an intelligent detection method and system for surface defects of wood-based panels based on machine vision. Background Art

[0002] Intelligent detection of surface defects of wood-based panels refers to the process of using artificial intelligence technologies, especially computer vision and deep learning algorithms, to automatically detect defects on the surface of wood-based panels (such as medium-density fiberboard, particleboard, etc.). This detection method collects images of the surface of wood-based panels through a camera device, and then analyzes the images through intelligent algorithms to identify various possible defects on the surface, such as cracks, bubbles, scratches, color differences, etc. However, when there are multiple defects on the surface of wood-based panels, traditional machine vision systems may be interfered by the defects. For example, when multiple cracks or bubbles are close to each other, the system may identify them as a larger defect, or misjudge similar defects as different types, resulting in misclassification. Summary of the Invention

[0003] The purpose of the present invention is to provide an intelligent detection method and system for surface defects of wood-based panels based on machine vision to solve the deficiencies in the background art.

[0004] To achieve the above purpose, the present invention provides the following technical solutions: An intelligent detection method for surface defects of wood-based panels based on machine vision, including: Collecting an image of the surface of a wood-based panel using a high-resolution industrial camera and performing preprocessing; Extracting features from the preprocessed image using a multi-scale convolutional neural network, where the shallow convolutional layer is used to extract the detailed features of small defects, and the deep convolutional layer is used to extract the global features of large defects, and generating a feature fusion map; Based on the extracted feature fusion map, using the Mask R-CNN model to segment the defect regions and removing redundant detection frames through the non-maximum suppression algorithm; Classifying each segmented defect region and identifying the defect type using a deep learning model; for the interference of adjacent defects, combining the spatial relationship and morphological features between the defects, and performing classification correction through an error correction algorithm; Real-time feedback of defect information to the production line control system, and the control system automatically rejects defective products or adjusts production process parameters according to the feedback information.

[0005] Preferably, an image of the surface of a wood-based panel is collected using a high-resolution industrial camera, and denoising, grayscale conversion, and contrast enhancement processing are performed.

[0006] Preferably, after extracting shallow and deep features, the feature maps of different layers are fused: the feature maps output by the shallow convolutional layer and the deep convolutional layer are concatenated in the channel dimension to obtain a multi-scale feature map. By assigning different weights to the shallow and deep feature maps and performing weighted summation on them, the fused feature map will be passed as input to the convolutional layer to further extract features, and finally, classification or regression tasks are performed through the fully connected layer to identify specific defect types and locations.

[0007] Preferably, the Mask R-CNN model first uses a convolutional neural network as a feature extractor to extract feature maps from the input feature fusion map. Based on the feature maps, the Region Proposal Network (RPN) generates candidate regions. The RPN generates multiple candidate boxes in a sliding window manner to mark the regions with defects. Each candidate box is assigned a confidence score to indicate whether there are defects in the region. Through the RoIAlign layer, the candidate regions generated by the RPN are mapped back to the feature maps and aligned. For each candidate region, the Mask R-CNN model uses convolutional operations to process it twice to generate a binary mask for each region. Through the generated mask, the edges of the defect regions can be segmented, and each region is classified according to the features of the candidate region to identify different types of defects.

[0008] Preferably, the non-maximum suppression algorithm is used to remove redundant boxes: First, the candidate boxes are sorted in descending order according to the confidence scores of each candidate box. The confidence score represents the credibility of the candidate region being a defect region. From the sorted candidate boxes, the box with the highest confidence is selected as the reference box, and the reference box is considered as the final detection box. For the remaining candidate boxes, calculate their overlap with the reference box, usually measured by IoU. The IoU value represents the ratio of the overlapping area of the two boxes to their union area. If the IoU value of two boxes exceeds the set threshold, it is considered that they are redundant boxes, and the boxes with high overlap are removed. Otherwise, the candidate boxes are retained.

[0009] Preferably, the method for obtaining the bounding box overlap rate is as follows: Define two bounding boxes: the ground truth box represents the actual position of the object, and the predicted box represents the position of the object predicted by the model. Calculate the intersection region. The intersection region is the rectangular region of the overlapping part of the two bounding boxes. Calculate the area of the intersection region: the width and height of the intersection are respectively: ; ; The intersection area Aint is: ; Calculate the area of the union region: The areas of the ground truth box and the predicted box are respectively: Agt = (x2 - x1) × (y2 - y1), Apred = (p3 - p1) × (p4 - p2); x1, y1 are the coordinates of the upper left corner of the ground truth box, x2, y2 are the coordinates of the lower right corner of the ground truth box, p1, p2 are the coordinates of the upper left corner of the predicted box, and p3, p4 are the coordinates of the lower right corner of the predicted box; The union area Aunion is: ; The union area is equal to the total area of the two boxes minus the intersection area; The bounding box overlap ratio is the ratio of the intersection area to the union area: ; where Aint is the area of the intersection region, Agt is the area of the ground truth box, Apred is the area of the predicted box, Aunion is the area of the union region, representing the total area of the two boxes, minus the overlapping part.

[0010] Preferably, the morphological feature analysis includes calculating the morphological feature similarity difference value: Extract morphological features for each defect region: Area A: representing the size of the defect region, Aspect Ratio: Circularity, Perimeter P; Calculate the difference in morphological features: For two defect regions, their morphological features are represented as vectors: V1 = (A1, Aspect Ratio1, Circularity1, P1) and V2 = (A2, Aspect Ratio2, Circularity2, P2), where each vector element represents the corresponding morphological feature, calculate the morphological feature similarity difference value, and calculate the difference between the two defect morphological features through the Euclidean distance as the morphological feature similarity difference value.

[0011] Preferably, normalize the bounding box overlap ratio and the morphological feature similarity difference value so that they are both in the range of [0, 1], and perform weighted summation on the normalized bounding box overlap ratio and the morphological feature similarity difference value to calculate the defect evaluation factor; Compare the obtained defect evaluation factor with a preset threshold. If the defect evaluation factor is greater than or equal to the preset threshold, divide the detection region into a defect region; if the defect evaluation factor is less than the preset threshold, divide the detection region into a normal detection region.

[0012] Preferably, by analyzing the spatial relationship and morphological features between defects, design an error correction algorithm for classification correction: Use a deep learning model to independently classify each defect region to obtain the preliminary type of each defect, and cluster the defect features through the adjacency matrix and spatial relationship rules; For each clustering region, determine whether the defects within the region belong to the same defect. If the morphological features of adjacent defects are similar and their spatial position overlap is high, they are merged into one defect; The merged defect region will be reclassified as a whole; For misclassification cases, correct the classification results of adjacent defects through an algorithm.

[0013] The present invention also provides an intelligent detection system for surface defects of wood-based panels based on machine vision, including an image acquisition and processing module, a feature extraction module, a defect segmentation module, an error correction module, and a feedback control module; Image acquisition and processing module: Use a high-resolution industrial camera to acquire the surface image of the wood-based panel and perform preprocessing; Feature extraction module: Use a multi-scale convolutional neural network to extract features from the preprocessed image. The shallow convolutional layer is used to extract the detailed features of small defects, and the deep convolutional layer is used to extract the global features of large defects, and generate a feature fusion map; Defect segmentation module: Based on the extracted feature fusion map, use the Mask R-CNN model to segment the defect region and remove redundant detection frames through the non-maximum suppression algorithm; Error correction module: Classify each segmented defect region and use a deep learning model to identify the defect type; for the interference of adjacent defects, combine the spatial relationship and morphological features between defects, and perform classification correction through an error correction algorithm; Feedback control module: Real-time feedback the defect information to the production line control system, and the control system automatically rejects defective products or adjusts production process parameters according to the feedback information.

[0014] In the above technical solution, the technical effects and advantages provided by the present invention: 1. The present invention combines deep learning and computer vision technologies to automatically detect surface defects of wood-based panels through a high-resolution industrial camera. This method can accurately identify various defect types, such as cracks, bubbles, scratches, etc., and perform efficient feature extraction and defect segmentation through a multi-scale convolutional neural network and the Mask R-CNN model, solving the interference problem generated by traditional vision systems when defects are adjacent or overlapping. By real-time feedback of defect information to the production line control system, this system can automatically reject defective products or adjust production process parameters, effectively reducing manual intervention and improving the automation and intelligence level of the production process.

[0015] 2. The error correction algorithm of the present invention can further optimize defect classification by combining the spatial relationship and morphological features between defects, improving the accuracy and robustness of the system. Through the application of this system, not only the detection efficiency of the production line is improved, but also the quality consistency of the products is greatly enhanced, the losses caused by human judgment errors are reduced, and the production cost is lowered. With the adaptive optimization of the system to the production environment, the detection accuracy and production efficiency can also be gradually improved, promoting the intelligent upgrade of the wood-based panel production process. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments recorded in the present invention, and those of ordinary skill in the art can also obtain other drawings based on these drawings.

[0017] Figure 1 It is the method mind map of the present invention.

[0018] Figure 2 It is the system module mind map of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0020] Example 1. Please refer to Figure 1 As shown, a method for intelligent detection of wood-based panel surface defects based on machine vision in this embodiment includes: Collect the surface image of the wood-based panel using a high-resolution industrial camera and perform preprocessing; Use a multi-scale convolutional neural network to extract features from the preprocessed image. The shallow convolutional layer is used to extract the detailed features of small defects, and the deep convolutional layer is used to extract the global features of large defects, and generate a feature fusion map; Based on the extracted feature fusion map, use the Mask R-CNN model to segment the defect area and remove redundant detection frames through the non-maximum suppression algorithm; Classify each segmented defect area and use a deep learning model to identify the defect type; for the interference of adjacent defects, combine the spatial relationship and morphological features between defects, and perform classification correction through an error correction algorithm; The defect information is fed back to the production line control system in real time, and the control system automatically removes defective products or adjusts production process parameters according to the feedback information.

[0021] Install high-resolution industrial cameras at key positions on the wood-based panel production line to ensure that every detail on the surface of the wood-based panel can be captured, and the resolution of image acquisition can meet the detection accuracy requirements, usually at least 300 DPI (dots per inch) or higher. The industrial camera should have high-speed acquisition capabilities to adapt to the high-frequency operation of the production line.

[0022] Use the camera to take images of the surface of the wood-based panel, covering as much of the surface area as possible to ensure that the captured images contain all possible defect locations. The acquisition frequency of the camera is usually several to dozens of images per second, adjusted according to the speed of the production line.

[0023] The position of the camera should be optimized according to the layout of the production line to minimize the impact of factors such as reflection and shadow on the image quality. Usually, the camera is installed above or on the side of the production line to ensure that surfaces at different angles can be captured.

[0024] After image acquisition, perform preprocessing steps to enhance the image quality, making defect features more prominent for subsequent defect recognition. The specific steps are as follows: Denoising processing: The acquired images are often interfered by noise, which may come from factors such as the camera's sensor, ambient light, or surface reflection. To remove these unnecessary interference signals, appropriate denoising algorithms are used. For example: Gaussian filtering: Apply the Gaussian filtering algorithm to smooth the image to remove high-frequency noise, especially the noise at the edges of the image, and keep the key features of the defects from being blurred.

[0025] Mean filtering: In some simple scenarios, mean filtering technology can also be used to denoise the image.

[0026] Grayscale processing: Convert the color image to a grayscale image to reduce the computational load and simplify subsequent processing. In grayscale images, defects are usually more obvious through brightness changes and can be more easily distinguished from the background. This step is achieved by converting the RGB channel information in the image into brightness values.

[0027] Contrast enhancement: To make surface defects more prominent, especially those tiny cracks, bubbles, etc., the image needs to be contrast-enhanced. Common contrast enhancement methods include: Histogram equalization: By adjusting the pixel distribution of the image, make the brightness value distribution of the image more uniform, thereby increasing the contrast of the image and highlighting the defect area.

[0028] Local contrast enhancement: For local regions of an image, fine-tune the contrast to make local defects more easily distinguishable.

[0029] Edge enhancement: Use edge detection algorithms (such as Sobel operator, Canny edge detection, etc.) to enhance the edges in the image, making the edges of surface defects more obvious. Edge enhancement can effectively highlight the contours of defects such as cracks and scratches, facilitating subsequent defect localization.

[0030] If the image captured by the camera has geometric distortion (such as perspective effects caused by an incorrect camera angle or a tilted surface), then the image needs to be perspective-corrected. Usually, a perspective transformation matrix is used to restore the tilted image to a front view, ensuring the accurate proportion and position of objects in the image.

[0031] The surface of the wood-based panel may affect the image quality due to changes in environmental lighting. Especially, the reflected light will affect the display of defects. At this time, the lighting of the image can be adjusted through a lighting equalization algorithm to make the contrast between the defect area and the background more obvious.

[0032] According to the gray value of the image, by setting an appropriate threshold, binarize the image (that is, divide the pixel points in the image into two categories according to the threshold, usually the defect area and the non-defect area). Binarization helps to quickly segment the main defect areas in the image, providing a clear image basis for subsequent feature extraction and classification.

[0033] During the actual operation of the production line, each image will be processed through the above preprocessing steps to ensure that the image quality remains in the best state. The real-time and high efficiency of image preprocessing are the keys to ensuring the normal operation of the entire detection system. Especially in a high-speed production environment, it is necessary to ensure that each frame of the image can be processed timely and accurately.

[0034] The preprocessed image will be passed to the subsequent deep learning network for tasks such as feature extraction, defect detection, and classification to ensure that defects can be accurately identified and located.

[0035] Through the above detailed image acquisition and preprocessing steps, this embodiment ensures that the machine vision system can stably and accurately capture and process the defect images on the surface of the wood-based panel in a complex production environment, providing high-quality image data support for subsequent defect detection, classification, and production line control.

[0036] In this embodiment, a multi-scale convolutional neural network is used to extract features from the preprocessed image. This network contains multiple convolutional layers, which are divided into shallow and deep structures and are used to extract features of different scales respectively. The shallow convolutional layers are used to capture the detailed features of small defects in the image, while the deep convolutional layers are mainly responsible for extracting the global features of large defects in the image. Through the multi-scale structure, the network can identify defects of different sizes and shapes simultaneously, improving the detection accuracy.

[0037] Shallow convolutional layers: Feature extraction: The shallow convolutional layers mainly focus on the local detailed features in the image, such as tiny cracks, fine scratches or small bubbles, etc. Through multiple convolutional operations, these layers can extract the texture information, edge features and pattern changes in a small range of the image.

[0038] Convolution kernel size: The shallow convolutional layers use small-sized convolution kernels (such as 3x3 or 5x5) to finely capture the details of small defects in the image. At this stage, the stride of the convolutional layer is usually small to retain higher-resolution features.

[0039] Activation function: The output of the shallow convolutional layers undergoes a non-linear transformation through an activation function (such as ReLU), increasing the representational ability of the model, so as to accurately extract the features of tiny defects.

[0040] Deep convolutional layers: Feature extraction: The deep convolutional layers are used to extract the global features of large defects in the image, such as larger cracks, obvious bubbles or surface depressions, etc. At this time, the deep convolutional layers begin to focus on a larger range of image content and identify more complex defect morphologies.

[0041] Convolution kernel size: The deep convolutional layers use larger-sized convolution kernels (such as 7x7 or 9x9) to capture global information and identify large-sized defects. The stride at this time is larger to reduce the computational complexity and enhance the expression of global features.

[0042] Activation function and pooling: The output of the deep convolutional layers is activated through an activation function (such as ReLU), and a pooling layer (such as max pooling or average pooling) is used for dimensionality reduction processing to further extract the high-level information of the features, while reducing the spatial resolution of the features to ensure the expression of the features at a larger scale.

[0043] Feature fusion map generation: Feature map fusion: After extracting the shallow and deep features, the system fuses the feature maps of different layers. Through feature fusion, the detailed features of small defects are combined with the global features of large defects. Common fusion methods include: Feature concatenation: Concatenate the feature maps output by the shallow convolutional layer and the deep convolutional layer in the channel dimension to obtain a multi-scale feature map. This method helps to retain the different-scale information extracted by each convolutional layer.

[0044] Weighted summation: Assign different weights to the shallow and deep feature maps and perform weighted summation on them. This method fuses according to the importance of each layer of feature maps, ensuring that important features can have a greater impact on the final result.

[0045] The fused feature map: The fused feature map contains the defect information of all scales in the image, which can help the detection system better identify defects of different sizes and shapes. This feature map is passed to the subsequent layers of the network for further processing, such as defect classification and localization.

[0046] The fused feature map will be passed as input to the subsequent convolutional layers to further extract high-level features and perform the final classification or regression tasks through the fully connected layer to identify specific defect types and locations.

[0047] During the network training process, a large number of labeled defect images are used for supervised learning, and the weights of the network are optimized through the backpropagation algorithm, so that the accuracy of feature extraction is continuously improved.

[0048] Through the multi-scale convolutional neural network, feature extraction is performed on the preprocessed image. The shallow convolutional layer can extract the detailed features of small defects, and the deep convolutional layer can capture the global features of large defects. The generation of the feature fusion map effectively combines the feature information of different scales, enhancing the network's ability to identify defects of different sizes and types. Finally, the fused feature map provides a solid foundation for subsequent defect classification and localization, improving the detection accuracy and robustness.

[0049] In the multi-scale convolutional neural network (CNN), through the aforementioned steps, a feature fusion map that combines shallow and deep features has been obtained. This fusion map contains defect information of different scales, including the local features of small defects and the global features of large defects. This feature fusion map will be used as the input of the Mask R-CNN model.

[0050] Mask R-CNN model architecture: Backbone Network: The Mask R-CNN model first uses a convolutional neural network (such as ResNet, VGG, etc.) as a feature extractor to extract high-level feature maps from the input feature fusion map. These feature maps capture the spatial information of the defects and provide the possible defect locations for each region.

[0051] Region Proposal Network (RPN): Based on the feature map, the Region Proposal Network (RPN) generates candidate regions (region proposals). The RPN generates multiple candidate bounding boxes in a sliding window manner to mark the regions where defects may exist. Each candidate bounding box is assigned a confidence score to indicate whether there are defects in that region.

[0052] RoIAlign layer: Through the RoIAlign (Region of Interest Align) layer, the candidate regions generated by the RPN are mapped back to the feature map and aligned to ensure that the size of each candidate bounding box is consistent. This step can refine the candidate bounding boxes more precisely, providing more accurate inputs for subsequent classification and segmentation.

[0053] Defect region segmentation: Secondary convolution: For each candidate region, the Mask R-CNN model performs secondary processing on it using convolution operations to generate a binary mask for each region. This mask represents the exact pixel positions of the defects in the candidate bounding box. Through the generated mask, the edges of the defect regions can be segmented at a fine-grained level, ensuring that the contours of the defects are clearly visible.

[0054] Class classification: Mask R-CNN not only generates masks but also classifies the types of defects, such as cracks, bubbles, scratches, etc. The model classifies each region based on the features of the candidate region to identify different types of defects.

[0055] Non-Maximum Suppression (NMS): Redundant bounding box removal: After the Mask R-CNN generates multiple candidate bounding boxes and assigns confidence scores to them, there may be cases where multiple detection bounding boxes cover the same defect region. To avoid repeated detection of defects by multiple overlapping bounding boxes, which may affect subsequent defect localization and classification, the Non-Maximum Suppression (NMS) algorithm is used to remove redundant bounding boxes.

[0056] Steps of the NMS algorithm: Candidate bounding box sorting: First, the candidate bounding boxes are sorted in descending order according to the confidence scores of each candidate bounding box. The confidence score represents the credibility of the candidate region being a defect region.

[0057] Select the high-confidence bounding box: From the sorted candidate bounding boxes, select the bounding box with the highest confidence as the reference bounding box, considering this box as the final detection bounding box.

[0058] Calculate the overlap degree: For the remaining candidate bounding boxes, calculate their overlap degree with the reference bounding box, usually measured by IoU. The IoU value represents the ratio of the overlapping area of the two bounding boxes to their union area.

[0059] Threshold judgment: If the IoU value of two bounding boxes exceeds the set threshold (usually 0.5), they are considered redundant bounding boxes, and the bounding box with a higher overlap degree is removed. Otherwise, the candidate bounding box is retained.

[0060] Repeat operation: Repeat the above steps until all candidate bounding boxes are processed.

[0061] Segmentation result output: Defect area calibration: Through the segmentation mask of Mask R-CNN, the pixel information of each defect area is accurately extracted, and its defect type is determined according to the classification result. Finally, after non-maximum suppression, the bounding box of each defect area and its corresponding defect category and accurate mask are output.

[0062] Defect location and type output: Each detected defect area will be accompanied by its location (represented by a bounding box), type (such as crack, bubble, etc.) and the mask information of the defect, as the input for subsequent defect evaluation and production line control.

[0063] Transmit the defect area, defect type and mask information optimized by NMS to the production line control system to automatically remove defective products or adjust the production process (such as adjusting temperature, pressure, etc.).

[0064] Segmenting the defect area of the feature fusion map through the Mask R-CNN model and removing redundant detection bounding boxes through the non-maximum suppression algorithm can significantly improve the accuracy and robustness of defect detection. This method can accurately segment the edge of each defect, and at the same time effectively eliminate the redundancy problem caused by the overlap of multiple candidate bounding boxes, providing high-quality data support for subsequent defect classification and production line control.

[0065] In the foregoing steps, the Mask R-CNN model was used to segment the defect area of the input feature fusion map, and the redundant detection bounding boxes were removed through the non-maximum suppression (NMS) algorithm to ensure the accurate positioning of each independent defect area. After these steps, the system obtains the bounding box of each defect and the corresponding mask, where each mask represents the accurate pixel position of the defect.

[0066] Defect type classification: Deep learning model classification: After obtaining the mask of each defect area, use the deep learning model to classify the type of each defect. The purpose of the classification task is to identify different types of defects (such as cracks, bubbles, scratches, etc.). The specific steps are as follows: Feature extraction: Using each segmented defect area mask as input, feature extraction is performed through a deep convolutional neural network (CNN). The model extracts features from aspects such as local details, texture, and shape to distinguish different types of defects.

[0067] Fully-connected layer classification: After feature extraction, the extracted high-dimensional features are input into the fully-connected layer for processing, and finally the classification results of each defect area are output. The output of the classification task is usually a probability distribution, indicating the probabilities that the defect belongs to different types (such as cracks, bubbles, scratches, etc.).

[0068] Softmax activation function: The classification layer usually uses the Softmax activation function to handle multi-class tasks. By calculating the probabilities of each class, the type of the defect area is finally determined. Each defect area will be assigned to a class according to the highest probability value.

[0069] Defect type recognition process: Training dataset: Training of the deep learning model requires a large number of labeled datasets. The datasets contain defect images of known types, and the defect types of each image have been labeled. Through supervised learning, the model can learn the features of different defect types.

[0070] For example, cracks usually present linear or curved shapes, bubbles may appear as local bulges in circular or elliptical shapes, and scratches are usually in the form of long and thin bands. By learning these features, the model can identify and classify different types of defects.

[0071] Multi-class classification: This embodiment supports the classification of multiple defect types and can identify multiple defect categories, including but not limited to cracks, bubbles, scratches, depressions, color differences, etc. Each defect area is assigned a class label, such as "crack", "bubble" or "scratch", and a confidence score for this class is given.

[0072] Defect type post-processing: Confidence threshold: For the classification results of each defect type, a confidence threshold is set. Only when the confidence of the classification result exceeds the set threshold, the area is considered to be confirmed as a specific defect type. For example, when the confidence is lower than 0.8, the area may be marked as "to be confirmed" or "other" category to avoid misclassification.

[0073] Output of defect type: After the classification and post-processing steps of the deep learning model, each defect area will be assigned a clear defect type. For example, a certain area may be identified as "crack" and assigned a high confidence value, such as 0.95; while another area is identified as "bubble" with a confidence of 0.87.

[0074] Output of defect type recognition result: Defect type and location: Each segmented defect area will output the following information: Defect Location: The location of the defect is calibrated by the coordinates of the bounding box, usually a rectangular box in the image, indicating the position of the defect in the image.

[0075] Defect Type: The type label (such as crack, bubble, scratch, etc.) of each defect and its corresponding confidence score, representing the probability that the defect belongs to a certain type.

[0076] Mask Information: The mask of the defect area can be used to accurately describe the pixel range of the defect, helping with subsequent defect processing.

[0077] Transmit information such as defect type, location, and confidence to the production line control system. The control system takes corresponding actions according to the defect type. For example, if a "bubble" defect is detected, the production parameters are automatically adjusted; if a "crack" or "scratch" defect is found, the unqualified products may need to be removed.

[0078] The classification results can also be used for the recording and analysis of production data. By statistically analyzing the defect types, quality problems in the production process can be identified and optimized and improved.

[0079] Through the classification and recognition of defect types by the deep learning model, this embodiment can accurately identify the type of each defect and provide detailed defect location and classification information. This can not only provide a basis for quality control on the production line but also improve the automation and intelligence level of the detection system, ensuring the consistency and reliability of product quality.

[0080] In practical applications, there may be multiple adjacent defects on the surface of the wood-based panel, such as multiple cracks, bubbles, or scratches, and these defects may touch or overlap each other in the image. Due to the adjacent positions and morphological similarities of these defects, traditional defect classification methods may lead to misjudgment or missed judgment, especially between defect types with similar morphologies. For example, two adjacent small cracks may be misjudged as a large crack, or two defects with similar morphologies may be misclassified as different types. To solve this problem, combining the spatial relationship and morphological characteristics between defects, a classification correction is carried out through an error correction algorithm.

[0081] Spatial Relationship Modeling: Adjacency Matrix Construction: First, use the position and size information of the defect areas to establish an adjacency matrix between the defects. Each defect area is calibrated in the image by its bounding box, and defects that are close in space or have similar morphologies will form associations in the adjacency matrix. For example, when the distance between the bounding boxes of two defects is less than a set threshold, they are regarded as adjacent defects and marked in the adjacency matrix.

[0082] Spatial relationship rule: By setting spatial relationship rules, the system can identify the relationships between adjacent defects, such as whether they belong to the same physical defect. For example, if two defects have similar types and a large spatial overlap, they may be different parts of the same defect. This rule can be set based on the Intersection over Union (IoU value). If the IoU value of two defects exceeds the set threshold, they are considered to belong to the same physical defect.

[0083] The method for obtaining the Intersection over Union is as follows: Define two bounding boxes: Ground truth box: Represents the actual position of the object. It is usually represented by the coordinates of its upper left corner (x1, y1) and lower right corner (x2, y2), representing the boundaries of the rectangle.

[0084] Predicted box: Represents the position of the object predicted by the model. It is also represented by the coordinates of its upper left corner (p1, p2) and lower right corner (p3, p4).

[0085] Calculate the intersection area: The intersection area is the rectangular area where the two bounding boxes overlap. The coordinates of its upper left and lower right corners can be calculated as follows: Upper left corner of the intersection area: xint = max(x1, p1), yint = max(y1, p2); Lower right corner of the intersection area: xint2 = min(x2, p3), yint2 = min(y2, p4); If the width or height of the intersection area is negative, it means the two boxes do not overlap and the intersection area is 0.

[0086] Calculate the area of the intersection area: The width and height of the intersection are respectively: ; ; The intersection area Aint is: ; If the width or height of the intersection is negative, then Aint = 0, indicating no intersection.

[0087] Calculate the area of the union area: The areas of the ground truth box and the predicted box are respectively: Agt = (x2 - x1) × (y2 - y1), Apred = (p3 - p1) × (p4 - p2); The union area Aunion is: ; The union area is equal to the total area of the two boxes minus the intersection area.

[0088] Intersection over Union Is the ratio of the intersection area to the union area: ; If the IoU is close to 1, it means that the overlapping degree of the two bounding boxes is very high; if it is close to 0, it means that the two boxes hardly overlap. Among them: x1, y1: the coordinates of the upper left corner of the ground truth box. x2, y2: the coordinates of the lower right corner of the ground truth box. p1, p2: the coordinates of the upper left corner of the predicted box. p3, p4: the coordinates of the lower right corner of the predicted box. Aint: the area of the intersection region, representing the overlapping part of the two boxes. Agt: the area of the ground truth box. Apred: the area of the predicted box. Aunion: the area of the union region, representing the total area of the two boxes minus the overlapping part.

[0089] Morphological feature analysis: Morphological feature extraction: Analyze the geometric morphology of the defects based on the morphological features (such as area, aspect ratio, circularity, edge smoothness, etc.) of each defect region extracted by the deep learning model. Through the morphological features, it can be further determined whether two adjacent defects have the same or similar morphological features. For example, if two adjacent defects have similar areas and shapes, the system may determine that they are different parts of the same defect.

[0090] Similarity judgment: For defects with similar morphological features, use similarity measures (such as cosine similarity, Euclidean distance, etc.) for judgment. If the morphological feature differences between two adjacent defects are small and their positions are close, they can be considered to belong to the same defect type.

[0091] Calculate the similarity difference value of morphological features: Extract the following morphological features for each defect region: Area A: Represents the size of the defect region. A = width × height; Aspect Ratio: The aspect ratio represents the aspect ratio of the defect. Aspect Ratio = length / width. Circularity: Represents the degree of circularity of the defect, usually defined as: ; where A is the area and P is the perimeter. The circularity of a circular object is close to 1, and the closer it is to 1, the rounder it is. Perimeter: Represents the edge length of the defect.

[0092] Calculate the differences in morphological features: For two defect regions (assumed to be Region 1 and Region 2), their morphological features can be represented as vectors: V1 = (A1, Aspect Ratio1, Circularity1, P1) and V2 = (A2, Aspect Ratio2, Circularity2, P2), where each vector element represents the corresponding morphological feature. Calculate the similarity difference value of morphological features, and calculate the difference between the morphological features of the two defects through the Euclidean distance. The smaller the similarity difference value of morphological features, the more similar the two defect morphologies are. According to the calculated similarity difference value of morphological features, a threshold can be set to determine whether the two defect regions belong to the same type. For example, when the similarity difference value of morphological features is less than a certain threshold (such as 0.5), it is considered that the two defect morphologies are similar and may belong to the same physical defect; when the similarity difference value of morphological features is greater than or equal to a certain threshold (such as 0.8), it is considered that they are morphologically similar.

[0093] Normalize the bounding box overlap rate and the similarity difference value of morphological features so that they are both in the range of [0, 1], and perform a weighted sum of the normalized bounding box overlap rate and the similarity difference value of morphological features to calculate the defect evaluation factor.

[0094] Compare the obtained defect evaluation factor with the preset threshold. If the defect evaluation factor is greater than or equal to the preset threshold, divide the detection region into a defect region; if the defect evaluation factor is less than the preset threshold, divide the detection region into a normal detection region.

[0095] By analyzing the spatial relationship and morphological features between defects, design an error correction algorithm for classification correction. The specific steps are as follows: Use a deep learning model to independently classify each defect region to obtain the preliminary type of each defect (such as crack, bubble, scratch, etc.). At the same time, record information such as the spatial position and morphological features of each defect.

[0096] Cluster the defect features through the adjacency matrix and spatial relationship rules. For each clustering region, judge whether the defects in the region belong to the same defect. If the morphological features of adjacent defects are similar and their spatial position overlap is high, they are merged into one defect.

[0097] The merged defect region will be reclassified as a whole. The merged region will be re-input into the classification model for accurate defect type recognition. If defects that were originally misjudged as different types are merged, they will be more accurately recognized as a unified defect type.

[0098] For misclassified cases, the classification results of adjacent defects are corrected by an algorithm. For example, two adjacent regions originally classified as "crack" and "scratch" are corrected to the same type of defect (such as crack) after morphological feature analysis. At the same time, if multiple defects belong to the same physical defect, the system will group them into a large defect to avoid over-segmentation.

[0099] After error correction, the final classification results will be obtained for all defect regions. These classification results not only include the type of defect (such as crack, bubble, scratch, etc.), but also update the specific location and size of the defect. For the interference problem of multiple adjacent defects, the corrected results can accurately reflect the true type and distribution of the defects.

[0100] The final classification results will be output to the production line control system for subsequent quality control and production adjustment. For example, if a "crack" defect is detected, the control system may automatically adjust the processing technology; if multiple adjacent "bubble" defects are found, the automatic rejection device may be activated.

[0101] The error correction algorithm can be continuously optimized according to the feedback results in actual production. By learning new data, the model can continuously adjust the spatial relationship rules and the determination criteria of morphological features, thereby improving the accuracy of classification correction. The optimization process of the algorithm usually adopts incremental learning and online learning methods to ensure that the system can adaptively improve its performance in different production environments.

[0102] By combining the spatial relationship and morphological features between defects and implementing the error correction algorithm, the interference problem of adjacent defects can be effectively solved. This algorithm not only improves the accuracy of classification, but also avoids misclassifying defects with similar morphology or close positions as different types, improving the robustness and precision of the intelligent detection system.

[0103] In this embodiment, an intelligent detection system based on machine vision is used to detect defects on the surface of wood-based panels. By real-time feedback of the detected defect information to the production line control system and automatically rejecting defective products or adjusting production process parameters according to the feedback information, the stability and consistency of product quality are ensured. The following is a detailed description of this process: The machine vision system on the production line uses a high-definition industrial camera to collect the surface image of the wood-based panel in real time, and uses a deep learning model (such as Mask R-CNN) for defect recognition and classification. After the image processing process is completed, the system can mark the location, type (such as crack, bubble, scratch, etc.) and size of each defect, and calculate its specific distribution on the surface.

[0104] Once a defect is detected, the system will encapsulate the detailed information of the defect (including defect type, location, size, confidence, etc.) into a data packet. The data packet usually contains the following fields: Defect ID: The number that uniquely identifies the defect.

[0105] Type: The classification result of the defect (such as crack, bubble, etc.).

[0106] Location coordinates: The coordinate information of the defect on the plate (such as the upper left and lower right corners of the bounding box).

[0107] Dimension information: Features such as the area, aspect ratio, perimeter of the defect.

[0108] Confidence level: The confidence level of the model's classification of the defect, indicating the credibility of the classification result.

[0109] Data transmission to the production line control system: After the defect information is extracted, the data packet is transmitted in real time to the central control system of the production line through a high-speed communication protocol (such as Ethernet / IP, Modbus TCP, Profinet, etc.). At this time, after receiving the defect data packet, the production line control system will further analyze and take appropriate processing measures.

[0110] The control system determines which products have defects based on the location and type information in the received defect data packet. For example, if a "bubble" defect is detected, the control system will, according to the location and dimension information of the defect, instruct the automatic rejection system to remove the defective product from the production line. The automatic rejection system can use the following several devices: Robotic arm rejection: The robotic arm precisely locates the defect position and removes the defective product from the production line.

[0111] Air jet device: According to the defect location information, the air jet device will be automatically triggered to "blow away" the defective product from the production line, preventing it from entering the next process.

[0112] Slide rail system: For small-sized defects, the slide rails of the production line may be adjusted to guide the defective plate through the sorting system for automatic rejection.

[0113] Automatic adjustment of production process parameters: In addition to rejecting defective products, the production line control system will also automatically adjust the production process parameters according to the detected defect type to optimize the production process and reduce the occurrence of defects. Common adjustment methods include: Adjusting temperature: If multiple "bubble" defects are detected, it indicates that there may be a problem with the temperature control during the production process. The control system will automatically adjust the temperature of the heating device to prevent the material from generating bubbles due to excessive temperature.

[0114] Adjusting pressure: When more "cracks" or "uneven surface" defects are detected, it may be caused by insufficient or uneven pressure. The control system will adjust the pressure of the pressing equipment to ensure uniform pressure during the production process.

[0115] If the defect is caused by the production line speed being too fast or too slow, the system will automatically adjust the production line speed to ensure that each sheet of plate can get the appropriate processing time during the processing process to avoid surface defects caused by operating too fast or too slow.

[0116] Based on real-time feedback, the control system will also monitor the real-time data of the production process and verify the adjustment effect based on the defect information and real-time data of the production process. If the adjusted process parameters still fail to effectively reduce defects, the control system will continue to adjust the process settings until the defect rate is reduced to an acceptable range.

[0117] Automatic optimization of process parameters: In order to improve production efficiency, the control system can also use historical defect data to perform trend analysis in a data-driven manner to predict and optimize process parameters for future production. For example, based on past defect data, the system can predict the changing trend of certain production process parameters and automatically adjust production parameters to further reduce the defect rate.

[0118] The control system displays the defect detection results and production process adjustments to the operator in real time, and displays detailed information about each defect (type, location, confidence level, etc.) and automatically adjusted production parameters through the interface. The operator can monitor the production process in real time, view the effect of defect handling, and make further manual intervention decisions.

[0119] If the number of detected defects exceeds a predetermined threshold (for example, the number of defects or the type of defects exceeds the allowable range), the control system will notify the operator or manager through the alarm system. The system will trigger emergency response procedures, such as suspending the production line, initiating additional inspections, etc., to ensure that production quality is not affected.

[0120] The control system also generates regular production reports, including defect detection results, the number of defective products rejected, adjusted process parameters, and the overall quality status of the production line. These reports not only help the production line to control quality, but also provide a basis for future production improvements.

[0121] The control system will conduct intelligent learning based on historical defect data and production feedback, and automatically optimize the defect detection model and production process parameters through machine learning algorithms. This adaptive optimization process enables the system to gradually improve its detection and adjustment accuracy as the production environment and raw materials change.

[0122] The system continuously collects and analyzes new data, and adjusts the algorithm and production process parameters in real time. For example, in different production batches, the system can automatically adjust the defect detection model according to the differences in materials to ensure that the production line of each batch can operate under the best conditions.

[0123] By real-time feedback of defect information to the production line control system and automatically removing defective products or adjusting production process parameters according to the feedback information, this embodiment realizes a highly automated quality control process. This system not only improves production efficiency, reduces manual intervention, but also ensures high product quality and Example 2, please refer to Figure 2 As shown, the intelligent detection system for surface defects of wood-based panels based on machine vision in this embodiment includes an image acquisition and processing module, a feature extraction module, a defect segmentation module, an error correction module, and a feedback control module; Image acquisition and processing module: Use a high-resolution industrial camera to collect images of the wood-based panel surface and perform preprocessing; Feature extraction module: Use a multi-scale convolutional neural network to extract features from the preprocessed image. The shallow convolutional layer is used to extract the detailed features of small defects, and the deep convolutional layer is used to extract the global features of large defects, and generate a feature fusion map; Defect segmentation module: Based on the extracted feature fusion map, use the Mask R-CNN model to segment the defect area and remove redundant detection frames through the non-maximum suppression algorithm; Error correction module: Classify each segmented defect area and use a deep learning model to identify the defect type; for the interference of adjacent defects, combine the spatial relationship and morphological features between defects and perform classification correction through an error correction algorithm; Feedback control module: Real-time feedback the defect information to the production line control system, and the control system automatically removes defective products or adjusts production process parameters according to the feedback information.

[0124] The above formulas are all dimensionless and take their numerical calculations. The formulas are obtained by collecting a large amount of data for software simulation to get a formula closest to the real situation. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0125] It should be understood that the term "and / or" in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist simultaneously, and B exists alone. These three situations, where A and B can be singular or plural. In addition, the character " / " in this article generally represents an "or" relationship between the front and back associated objects, but it may also represent an "and / or" relationship, which can be specifically understood by referring to the context before and after.

[0126] Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be implemented by electronic hardware, or by a combination of computer software and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods for each specific application to implement the described functions, but such implementation should not be considered to exceed the scope of this application.

[0127] As described above, it is only the specific implementation manner of this application, but the protection scope of this application is not limited thereto. Any person skilled in the art within the technical scope disclosed in this application can easily think of changes or substitutions, which should all be covered within the protection scope of this application.

Claims

1. An intelligent detection method for surface defects of wood-based panels based on machine vision, characterized in that: Including: Collect the surface image of the wood-based panel using a high-resolution industrial camera and perform preprocessing; Use a multi-scale convolutional neural network to extract features from the preprocessed image. The shallow convolutional layer is used to extract the detailed features of small defects, and the deep convolutional layer is used to extract the global features of large defects, and generate a feature fusion map; Based on the extracted feature fusion map, use the Mask R-CNN model to segment the defect area and remove redundant detection boxes through the non-maximum suppression algorithm; Classify each segmented defect area and use a deep learning model to identify the defect type; for the interference of adjacent defects, combine the spatial relationship and morphological features between defects, and perform classification correction through an error correction algorithm; Real-time feedback the defect information to the production line control system, and the control system automatically removes defective products or adjusts production process parameters according to the feedback information.

2. The intelligent detection method for surface defects of wood-based panels based on machine vision according to claim 1, wherein: Collect the surface image of the wood-based panel using a high-resolution industrial camera and perform denoising, grayscale conversion, and contrast enhancement processing.

3. The intelligent detection method for surface defects of wood-based panels based on machine vision according to claim 1, wherein: After extracting the shallow and deep features, fuse the feature maps of different layers: splice the feature maps output by the shallow convolutional layer and the deep convolutional layer in the channel dimension to obtain a multi-scale feature map. By assigning different weights to the shallow and deep feature maps, perform weighted summation on them. The fused feature map will be used as the input to the convolutional layer to further extract features, and perform the final classification or regression task through the fully connected layer to identify the specific defect type and location.

4. An intelligent detection method for surface defects of wood-based panels based on machine vision according to claim 3, characterized in that: The Mask R-CNN model first uses a convolutional neural network as a feature extractor to extract a feature map from the input feature fusion map. Based on the feature map, the Region Proposal Network (RPN) generates candidate regions. The RPN generates multiple candidate boxes in a sliding window manner to mark the regions with defects. Each candidate box will be assigned a confidence score to indicate whether there are defects in the region; through the RoIAlign layer, the candidate regions generated by the RPN are mapped back to the feature map and aligned; For each candidate region, the Mask R-CNN model uses convolutional operations to process it twice to generate a binary mask for each region. Through the generated mask, the edge of the defect area can be segmented, and each region is classified according to the characteristics of the candidate region to identify different types of defects.

5. The intelligent detection method for surface defects of wood-based panels based on machine vision according to claim 4, characterized in that: Adopt the non-maximum suppression algorithm to remove redundant boxes: First, sort the candidate boxes in descending order according to the confidence score of each candidate box. The confidence score represents the credibility of the candidate region as a defect region; Select the box with the highest confidence score from the sorted candidate boxes as the reference box and consider the reference box as the final detection box; For the remaining candidate boxes, calculate their overlap degree with the reference box, usually measured by IoU. The IoU value represents the ratio of the overlapping area of the two boxes to their union area; If the IoU value of two boxes exceeds the set threshold, it is considered that they are redundant boxes, and the boxes with high overlap degree are removed; otherwise, the candidate boxes are retained.

6. The intelligent detection method for surface defects of wood-based panels based on machine vision according to claim 5, characterized in that: The method for obtaining the bounding box overlap rate is as follows: Define two bounding boxes: the ground truth box represents the actual position of the object, and the predicted box represents the position of the object predicted by the model; Calculate the intersection area. The intersection area is the rectangular area of the overlapping part of the two bounding boxes. Calculate the area of the intersection area: The width and height of the intersection are respectively: ; ; The intersection area Aint is: ; Calculate the area of the union area: The areas of the ground truth box and the predicted box are respectively: Agt = (x2 - x1) × (y2 - y1), Apred = (p3 - p1) × (p4 - p2); x1, y1 are the coordinates of the upper left corner of the ground truth box, x2, y2 are the coordinates of the lower right corner of the ground truth box, p1, p2 are the coordinates of the upper left corner of the predicted box, and p3, p4 are the coordinates of the lower right corner of the predicted box; The union area Aunion is as follows: ; the union area is equal to the total area of the two bounding boxes minus the intersection area; the bounding box overlap ratio is the ratio of the intersection area to the union area: ; where Aint is the area of the intersection region, Agt is the area of the ground truth bounding box, Apred is the area of the predicted bounding box, and Aunion is the area of the union region, representing the total area of the two bounding boxes minus the overlapping part.

7. The intelligent detection method for surface defects of wood-based panels based on machine vision according to claim 6, characterized in that: Morphological feature analysis includes calculating the difference value of morphological feature similarity: Extract morphological features for each defect area: Area A: representing the size of the defect area, Aspect Ratio: Circularity, Perimeter P; Calculate the difference of morphological features: For two defect areas, their morphological features are represented as vectors: V1=(A1, AspectRatio1, Circularity1, P1) and V2=(A2, Aspect Ratio2, Circularity2, P2), where each vector element represents the corresponding morphological feature. Calculate the difference value of morphological feature similarity, and calculate the difference between the morphological features of the two defects through the Euclidean distance as the difference value of morphological feature similarity.

8. An intelligent detection method for surface defects of wood-based panels based on machine vision according to claim 7, characterized in that: Normalize the bounding box overlap rate and the difference value of morphological feature similarity so that they are both within [0,1], and perform weighted summation on the normalized bounding box overlap rate and the difference value of morphological feature similarity to calculate the defect evaluation factor; Compare the obtained defect evaluation factor with the preset threshold. If the defect evaluation factor is greater than or equal to the preset threshold, divide the detection area into a defect area; If the defect evaluation factor is less than the preset threshold, divide the detection area into a normal detection area.

9. The intelligent detection method for surface defects of wood-based panels based on machine vision according to claim 8, characterized in that: By analyzing the spatial relationship and morphological features between defects, design an error correction algorithm for classification correction: Use a deep learning model to independently classify each defect area to obtain the preliminary type of each defect, and cluster the defect features through the adjacency matrix and spatial relationship rules; For each clustering area, judge whether the defects in the area belong to the same defect. If the morphological features of adjacent defects are similar and their spatial position overlap degree is high, merge them into one defect; The merged defect area will be reclassified as a whole; For misclassification cases, correct the classification results of adjacent defects through the algorithm.

10. An intelligent detection system for surface defects of wood-based panels based on machine vision, which is used to implement an intelligent detection method for surface defects of wood-based panels based on machine vision according to any one of claims 1-9, characterized in that: Including an image acquisition and processing module, a feature extraction module, a defect segmentation module, an error correction module, and a feedback control module; Image acquisition and processing module: Use a high-resolution industrial camera to collect the surface image of the wood-based panel and perform preprocessing; Feature extraction module: Use a multi-scale convolutional neural network to extract features from the preprocessed image. The shallow convolutional layer is used to extract the detailed features of small defects, and the deep convolutional layer is used to extract the global features of large defects and generate a feature fusion map; Defect segmentation module: Based on the extracted feature fusion map, use the Mask R-CNN model to segment the defect area and remove redundant detection boxes through the non-maximum suppression algorithm; Error correction module: Classify each segmented defect area and use a deep learning model to identify the defect type; for the interference of adjacent defects, combine the spatial relationship and morphological features between defects and perform classification correction through the error correction algorithm; Feedback control module: Real-time feedback the defect information to the production line control system, and the control system automatically rejects the defective products or adjusts the production process parameters according to the feedback information.

Images