Wood composite material surface defect judgment method

Abstract

The invention belongs to the technical field of visual inspection, and particularly relates to a wood composite material surface defect judgment method which can efficiently and accurately detect surface defects of a wood composite material by combining an AI computer visual processing technology and a generative adversarial network, remarkably improve production efficiency and product quality, reduce labor cost and improve product quality. The product market competitiveness is improved, and wide application prospects are achieved in the fields of furniture manufacturing, floor laying and the like.

Description

Technical Field

[0001] The invention belongs to the technical field of visual inspection, and in particular relates to a method for judging surface defects of a wood composite material. Background Art

[0002] Wood composite materials are widely used in furniture manufacturing, floor laying and other fields due to their excellent performance and low cost. However, during the production process, cracks, scratches, knots, bubbles and other defects are prone to appear on the surface of wood composite materials. These defects will affect the appearance and service life of the product. Therefore, corresponding testing of such products is required in the production process.

[0003] Traditional detection methods mainly rely on manual and simple machine vision technology, and have problems such as low detection efficiency and low accuracy. With the development of artificial intelligence technology, computer vision processing technology has made major breakthroughs in image recognition and analysis, and is expected to solve many deficiencies in existing technologies. However, the training of defect detection models usually requires a large number of defect image samples, which are often difficult to obtain, resulting in detection errors in actual applications of the model and requiring a lot of time and effort to supplement samples. Summary of the invention

[0004] The purpose of the present invention is to provide a method for judging surface defects of wood composite materials, so as to solve the problems mentioned in the background technology.

[0005] In order to achieve the above technical objectives, the technical solution adopted by the present invention is as follows:

[0006] A method for judging surface defects of wood composite materials includes an image acquisition module, a GAN data enhancement module, a result output module and a result output module, and the specific steps of the judgment are as follows:

[0007] Step 1: Use the image acquisition module to collect the material surface image;

[0008] Step 2: Use the GAN data enhancement module to generate defect image samples and expand the training data set;

[0009] Step 3: The result output module performs noise filtering and image enhancement on the collected images, uses a deep learning algorithm to extract image features, and classifies material surface defects based on the extracted features;

[0010] Step 4: The result output module displays the test results and generates a test report.

[0011] In the step 1, the image acquisition module uses a high-resolution industrial camera and a multi-spectral light source to perform comprehensive image acquisition on the material surface, and adjusts the camera's focal length, aperture and exposure time according to the material's characteristics and detection requirements to ensure image clarity and detail integrity.

[0012] In the step 2, the generative adversarial network (GAN) model is used to perform data enhancement on the existing defect images to generate a large number of high-quality defect image samples. At the same time, the conditional GAN ​​model uses the defect type as a conditional label to generate various types of defect images and expand the training data set to improve the generalization ability of the defect detection model.

[0013] In the step three, the collected image is preprocessed, and the image noise is filtered by using median filtering. The neighborhood area of ​​each pixel is sorted, and the median of the area is used as the new pixel value to effectively remove the salt and pepper noise.

[0014] In the step three, Laplace sharpening technology is used to perform image enhancement processing, and the edge of the image is enhanced by calculating the second-order derivative of the image to improve the clarity of the image.

[0015] In the step three, a pre-trained convolutional neural network is used to extract features from the pre-processed image, and a high-dimensional feature vector is extracted through a convolutional layer, a pooling layer, and a fully connected layer.

[0016] In the step three, a support vector machine classification algorithm is used to classify the extracted high-dimensional feature vectors, and an accurate classification of material surface defects is achieved by training a classification model.

[0017] In step 4, the inspection structure displayed in real time includes but is not limited to the type, location and severity information of the defect. The inspection results are marked on the original image, a detailed inspection report is generated, and the results are stored in a database for reference in quality analysis and process improvement. At the same time, an alarm signal is issued according to the severity of the defect to prompt the operator to handle it in time.

[0018] Compared with the prior art, the present invention has at least the following advantages:

[0019] This application combines AI computer vision processing technology and generative adversarial networks to efficiently and accurately detect defects on the surface of wood composite materials, significantly improve production efficiency and product quality, reduce labor costs, and improve product market competitiveness. It has broad application prospects in furniture manufacturing and floor laying. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The present invention is further illustrated by means of the non-limiting examples given in the accompanying drawings.

[0021] Figure 1 It is a structural schematic diagram of a wood composite material surface defect detection method that combines AI computer vision processing technology and a generative adversarial network in the present invention.

[0022] Figure 2 It is a block diagram of an image acquisition module in a wood composite material surface defect detection method that combines AI computer vision processing technology and a generative adversarial network in the present invention.

[0023] Figure 3 It is a block diagram of a GAN data enhancement module in a wood composite material surface defect detection method combining AI computer vision processing technology and a generative adversarial network in the present invention.

[0024] Figure 4 It is a block diagram of a defect detection module in a wood composite material surface defect detection method that combines AI computer vision processing technology and a generative adversarial network in the present invention.

[0025] Figure 5 It is a block diagram of a result output module in a wood composite material surface defect detection method combining AI computer vision processing technology and a generative adversarial network according to the present invention. DETAILED DESCRIPTION

[0026] In order to enable those skilled in the art to better understand the present invention, the technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0027] like Figure 1-5 As shown, a method for judging surface defects of wood composite materials includes an image acquisition module, a GAN data enhancement module, a defect detection module and a result output module, and the specific steps of the judgment are as follows:

[0028] Step 1: Use the image acquisition module to collect the material surface image;

[0029] Step 2: Use the GAN data enhancement module to generate defect image samples and expand the training data set;

[0030] Step 3: The result output module performs noise filtering and image enhancement on the collected images, uses a deep learning algorithm to extract image features, and classifies material surface defects based on the extracted features;

[0031] Step 4: The result output module displays the test results and generates a test report.

[0032] In the step 1, the image acquisition module acquires the image of the material surface, selects a high-resolution industrial camera and a multi-spectral light source (including visible light, infrared light, ultraviolet light, etc.), and ensures that the multi-spectral image of the surface of the wood composite material can be acquired. The image acquisition steps are as follows:

[0033] Equipment calibration: Use the calibration board to calibrate the camera, adjust the camera's focal length, aperture and exposure time to ensure accurate camera focus and clear images, adjust the position and brightness of the multi-spectral light source to ensure uniform illumination of the wood composite surface and reduce the impact of shadows and reflections on image quality.

[0034] Continuous shooting: Set the image acquisition frequency according to the speed of the production line and the inspection requirements. You can use the trigger mode to detect the material position through the sensor and trigger the camera to shoot, ensuring that every material surface is captured.

[0035] Data transmission and storage: The collected image data is transmitted to the central processing unit in real time through a high-speed interface (GigE), and stored and managed according to timestamps or other identifiers to ensure data integrity and traceability.

[0036] In step 2, the GAN data enhancement module includes data preparation, GAN model training and data enhancement;

[0037] The specific method of data preparation is: annotate the original image obtained by the image acquisition module, use the tool LabelImg to manually mark the defect area, divide the image into normal image and defect image, and generate the corresponding annotation file.

[0038] The specific steps of GAN model training are:

[0039] S1: The specific method of data preprocessing is: resizing the original image (unified to 256x 256 pixels), graying it, and dividing it into a training set and a test set.

[0040] S2: The specific method of the model architecture is: using a generative adversarial network (GAN), including a generator and a discriminator, and the training process steps are as follows:

[0041] ① Generator: Input a noise vector (random number) and generate fake images similar to real defect images through multi-layer convolution and deconvolution networks. The loss function uses generative adversarial loss to encourage the generator to generate images that cannot be distinguished by the discriminator.

[0042] ② Discriminator: Input real images and generated images, perform feature extraction and classification through a multi-layer convolutional network, and output the probability of true or false. The loss function uses cross entropy loss. The discriminator is trained to correctly distinguish between real images and generated images.

[0043] ③ Training iteration: Alternately train the generator and the discriminator until the generator can generate high-quality defect images and the discriminator cannot distinguish between real and fake images. Use the optimization algorithm Adam to update the parameters, and adjust the learning rate and batch size to speed up training.

[0044] The specific method of data enhancement is: using the trained GAN model to generate a large number of defect images, further expanding the data set through image transformations such as rotation, flipping, and scaling, improving the generalization ability of the model, and improving the generative adversarial network model;

[0045] First, the conditional label y is introduced into the generator, so that the input of the generator consists of random noise z and the conditional label y. Secondly, the same conditional label y is introduced into the discriminator, that is, the input of the discriminator consists of two parts: one is the real data x or the data generated by the generator, and the other is the corresponding conditional variable y;

[0046] During the training process, the generator and the discriminator are trained alternately. The discriminator continuously improves its discrimination ability in each iteration to distinguish between real data and generated data. The generator tries to deceive the discriminator and generate more realistic data. Through this adversarial training, the generator can generate more realistic samples, thereby enriching the diversity of the training data set.

[0047] After introducing conditional variables, the optimization function of GAN is as follows:

[0048]

[0049] For binary or multi-classification data, setting conditional information (category labels) as constraints can guide data generation, thereby producing outputs with specific characteristics and further alleviating the overfitting problem.

[0050] In step 3, the defect detection module specifically includes three parts: image preprocessing, feature extraction and defect classification. The defect detection steps are specifically as follows:

[0051] S1: The image preprocessing steps specifically include:

[0052] ① Normalize the image, resize the collected image (unified to 224x 224 pixels), normalize (scale the pixel value to the range of 0-1) and adjust the color to ensure the consistency of the input image.

[0053] ② Remove noise. Use image processing techniques such as median filtering to remove noise in the image, improve image quality, and retain key features. Specifically, introduce edge detection algorithm and Sobel operator to obtain the gradient values ​​in the horizontal and vertical directions of the image, so as to calculate the gradient amplitude and gradient direction of the pixel point. Based on the median filtering algorithm of image gradient, protect the edge by judging whether the pixel point is in the edge area, and perform median filtering on the noise area to remove noise. The specific implementation steps are as follows:

[0054] (1) Input an original grayscale image.

[0055] (2) Determine whether the central pixel is noise. If the difference between the maximum value and the central pixel, and the difference between the central pixel and the minimum value in the filter are both less than 0, the pixel is determined to be noise, and median filtering is performed to output the result; otherwise, it is marked as a pixel for further inspection.

[0056] (3) Use the Sobel operator to calculate the gradient of the original image.

[0057] (4) Calculate the difference value of the pixels in the filter window and determine whether the central pixel is an edge. If the gradient value is greater than the preset threshold and the difference value does not exceed the preset threshold, it is determined to be an edge and its value is retained; otherwise, perform median filtering.

[0058] (5) Traverse all pixels of the image to obtain a filtered image, in which the pixel values ​​at the edge remain unchanged and the pixel values ​​at non-edges are processed by median filtering.

[0059] (6) Output the filtered image.

[0060] The mathematical expression of the improved median filtering algorithm can be written as:

[0061]

[0062] Among them, Y i is the new pixel value in the filter, Δ is the sum of the differences, and kernel is the filter neighborhood size.

[0063] ③ Perform image enhancement. In order to further enhance the image, the Laplace operator is introduced to sharpen the image after median filtering to improve the clarity of the image edges and details.

[0064] The Laplacian operator of a two-dimensional image function is defined as:

[0065]

[0066] Among them, the two second-order partial differentials can be expressed as:

[0067]

[0068]

[0069] Since the second-order differentials in the x and y directions are linear operations, the Laplace transform is a linear transform. Based on the above formula, the discrete Laplace operator of two variables can be expressed as:

[0070]

[0071] By constructing a Laplace template and applying the operator to the original image, the image can be sharpened. The final sharpened image function is expressed as:

[0072]

[0073] S2: Use the feature extraction unit to process the image, use the convolutional neural network (VGG) to extract the deep features of the image, and design a reasonable network structure, including convolutional layers, pooling layers, fully connected layers, etc., to ensure the effectiveness and efficiency of feature extraction. Specifically:

[0074] ① Input the sharpened image into the pre-trained convolutional neural network, perform convolution operation through the two-dimensional convolution layer, automatically extract the defect features of the image, and generate a feature map with the help of the ReLU activation function;

[0075] The convolution operation formula is as follows:

[0076]

[0077] Among them, Y is the output feature map, X is the input image, W is the convolution kernel, b is the bias term, and M and N are the convolution kernel sizes.

[0078] Using the ReLU activation function, the formula is as follows:

[0079] f(x)=max(0,x)

[0080] ② Perform downsampling operation through the maximum pooling layer, compress the feature map and pass it to the next layer;

[0081] The pooling operation formula is as follows:

[0082]

[0083] Among them, Y is the output feature map after pooling, X is the input feature map, and m and n are the pooling window sizes.

[0084] ③Extract image features layer by layer through multi-layer two-dimensional convolution to obtain high-dimensional feature vectors. Each layer of convolution operation is followed by ReLU activation function and maximum pooling operation to form deep features;

[0085] ④Fuse the feature maps output by different convolution blocks:

[0086] Using interpolation method, the feature map with resolution lower than the preset value is scaled to be equal to the feature map with the highest resolution;

[0087] The interpolation formula is as follows:

[0088]

[0089] The feature maps output by multiple convolution blocks are concatenated in the channel dimension to obtain the final feature map;

[0090] The splicing operation formula is as follows:

[0091] F concat =concat(F 1 ,F 2 ,…,F n )

[0092] Among them, F i is the output feature map of the i-th convolutional block, F concat It is the concatenated feature map.

[0093] ⑤ Reduce the dimension of the extracted feature map through the fully connected layer or the global average pooling layer to generate a fixed-length feature vector for subsequent classification tasks;

[0094] The operation formula of the fully connected layer is as follows:

[0095] y=σ(Wx+b)

[0096] Among them, x is the flattened feature map, W is the weight matrix, b is the bias term, and σ is the activation function.

[0097] For the global average pooling layer, a fixed-length feature vector is generated by averaging all elements of each feature map;

[0098] The global average pooling formula is as follows:

[0099]

[0100] Among them, y k is the output value of the kth channel, H and W are the height and width of the feature map, and X is the input feature map.

[0101] S3: Use the defect classification unit to input the extracted image features into the classifier for classification, determine whether the image has defects based on the classification results, and further refine the defect type, specifically:

[0102] ① The classification algorithm is selected using support vector machine (SVM):

[0103] (1) Using the linear kernel function The feature vector extracted by VGG is mapped to a high-dimensional space, and the optimal hyperplane is found through global search. The optimal classification decision function is obtained as follows:

[0104]

[0105] in, and b *is the solution in the formula, n is the number of training samples, x j is the input vector, x j is the corresponding expected value.

[0106] (2) Introduce the ten-fold cross-validation method to train SVM. Divide the data set into ten subsets, use nine subsets for training each time, and use one subset for validation, and repeat this cycle ten times to train the optimal decision function and validate it in the test set to obtain the optimal SVM hyperparameters and improve the accuracy and generalization ability of the model.

[0107] (3) The SVM model optimized by ten-fold cross-validation is used to classify and identify the image features automatically extracted by the deep learning network, thereby achieving accurate classification of material surface defects.

[0108] ② Perform classification model training and use feature vectors and corresponding labels (defect types) to train the classification model. Improve classification accuracy by optimizing classification boundaries or decision trees. Use cross-validation, grid search and other methods to adjust model parameters to prevent overfitting.

[0109] ③ Perform defect determination, input the feature vector of the test image into the classification model, and output the classification result. Determine whether the image has defects based on the classification result, and further refine the defect type. Perform manual re-inspection on samples that are difficult to classify to improve the robustness of the system.

[0110] In step 4, the result output module generates a test report based on the test results and displays the classification and positioning information. Manual re-inspection is performed on samples that are difficult to classify to ensure the robustness of the system, specifically:

[0111] ①Result display: The inspection results are displayed in real time on the display interface of the central control system, including defect type, location, severity and other information. The following steps and technologies are used:

[0112] (1) Real-time display:

[0113] Image overlay technology: Display the inspection results on the original image in the form of image overlay. Use different colors or symbols to mark defects of different types and severity, so that operators can intuitively see the location and nature of the defects.

[0114] Interface layout: Design a concise and clear user interface to display real-time inspection images, defect statistics, and historical inspection data. The interface should have zoom, pan, and detailed information viewing functions to facilitate operator inspection and analysis.

[0115] (2) Detailed information display:

[0116] Defect Information Panel: Displays detailed defect information of the current inspection image on one side of the interface or in a pop-up window, including defect type, location (coordinates), area, and severity.

[0117] Historical data backtracking: Provides a timeline or log function, allowing operators to view the test results within a certain period of time in the past for comparison and analysis.

[0118] ②Data storage and feedback: Store the test results in the database and generate statistical reports to facilitate quality analysis and process improvement. The specific steps and techniques are as follows:

[0119] (1) Data storage:

[0120] Database design: Establish a dedicated database to store inspection results, including image data, defect information, timestamps, and production batches, etc. The database must have efficient data access and management capabilities.

[0121] Data compression and backup: For large amounts of image data, compression technology is used to reduce storage space usage, and data is backed up regularly to ensure data security.

[0122] (2) Statistical report generation:

[0123] Report content: Generate reports including defect frequency, distribution, severity statistics, etc. Reports can be classified by time period, production batch or defect type.

[0124] Automatic analysis: Use data mining and statistical analysis techniques to automatically analyze test results, identify common defects and high-incidence areas, and provide a basis for quality control and process improvement.

[0125] (3) Data feedback, quality analysis and process improvement:

[0126] Quality analysis: Use stored test data to conduct long-term quality analysis to identify process bottlenecks and improvement directions. Generate trend charts and comparison charts to help managers make decisions.

[0127] Process improvement: According to the analysis results, adjust the production process and parameters to reduce the occurrence of defects and improve product quality and the stability of the production process.

[0128] ③Alarm mechanism: For serious defects, the system automatically sends out an alarm signal to prompt the operator to deal with it in time. A multi-level alarm mechanism is adopted to take different treatment measures according to the severity and location of the defects. The specific steps and techniques are as follows:

[0129] (1) Automatic alarm:

[0130] Alarm type: The system sets a variety of alarm types, including visual, sound and light alarms, and SMS / email notifications. Choose different alarm methods according to the severity of the defect.

[0131] Alarm priority: Set up a multi-level alarm mechanism, graded according to the severity of the defect. For example, a minor defect (such as a surface scratch) will issue a low-level alarm, and a serious defect (such as a crack or fall-off) will issue a high-level alarm.

[0132] (2) Alarm processing:

[0133] Suspend production: For serious defects, the system can automatically suspend the production line to prevent defective products from entering the next process.

[0134] Automatic rejection: An automatic rejection device is installed on the production line. For defective products detected, the system can control the device to reject them to prevent unqualified products from entering the market.

[0135] Manual re-inspection: For defects that are difficult to classify or high-risk, the system prompts the operator to conduct manual re-inspection to ensure the accuracy and reliability of the test results.

[0136] (3) Alarm recording and analysis:

[0137] Record storage: All alarm information will be recorded and stored in the database, including alarm time, defect type, treatment measures, etc.

[0138] Analysis and Improvement: Regularly analyze alarm records, identify the causes of high-frequency alarms, formulate improvement measures, optimize the alarm mechanism, reduce false alarms and missed alarms, and improve the robustness of the system.

[0139] In summary, an embodiment of a method for judging surface defects of wood composite materials based on the present application is explained. By combining AI computer vision processing technology and generative adversarial networks, it can efficiently and accurately detect defects on the surface of wood composite materials, significantly improve production efficiency and product quality, reduce labor costs, and improve product market competitiveness. It has broad application prospects in furniture manufacturing and floor laying and other fields.

[0140] The above embodiments are merely illustrative of the principles and effects of the present invention, and are not intended to limit the present invention. Anyone familiar with the technology may modify or change the above embodiments without violating the spirit and scope of the present invention. Therefore, all equivalent modifications or changes made by those with ordinary knowledge in the technical field without departing from the spirit and technical ideas disclosed by the present invention should still be covered by the claims of the present invention.

Claims

1. A method for judging surface defects of wood composite materials, characterized by: It includes image acquisition module, GAN data enhancement module, result output module and result output module. The specific steps of judgment are as follows: Step 1: Use the image acquisition module to collect the material surface image; Step 2: Use the GAN data enhancement module to generate defect image samples and expand the training dataset; Step 3: The result output module performs noise filtering and image enhancement on the collected images, uses a deep learning algorithm to extract image features, and classifies material surface defects based on the extracted features; Step 4: The result output module displays the test results and generates a test report.

2. The method for judging surface defects of wood composite materials according to claim 1, characterized in that: In step 1, the image acquisition module uses a high-resolution industrial camera and a multi-spectral light source to comprehensively capture images of the material surface. At the same time, the camera's focal length, aperture, and exposure time are adjusted according to the material's characteristics and detection requirements to ensure image clarity and detail integrity.

3. The method for judging surface defects of wood composite materials according to claim 2, characterized in that: In the second step, the generative adversarial network (GAN) model is used to perform data enhancement on the existing defect images to generate a large number of high-quality defect image samples. At the same time, the conditional GAN ​​model uses the defect type as the conditional label to generate various types of defect images and expand the training data set to improve the generalization ability of the defect detection model.

4. The method for judging surface defects of wood composite materials according to claim 3, characterized in that: In the step three, the collected image is preprocessed and image noise is filtered using median filtering. The neighborhood area of ​​each pixel is sorted and the median of the area is used as the new pixel value to effectively remove salt and pepper noise.

5. The method for judging surface defects of wood composite materials according to claim 4, characterized in that: In the step three, Laplace sharpening technology is used to perform image enhancement processing, and the edges of the image are enhanced by calculating the second-order derivative of the image to improve the clarity of the image.

6. The method for judging surface defects of wood composite materials according to claim 5, characterized in that: In the step three, a pre-trained convolutional neural network is used to extract features from the pre-processed image, and a high-dimensional feature vector is extracted through a convolutional layer, a pooling layer, and a fully connected layer.

7. The method for judging surface defects of wood composite materials according to claim 6, characterized in that: In the step three, a support vector machine classification algorithm is used to classify the extracted high-dimensional feature vectors, and an accurate classification of material surface defects is achieved by training a classification model.

8. The method for judging surface defects of wood composite materials according to claim 7, characterized in that: In step 4, the inspection structure displayed in real time includes but is not limited to the type, location and severity of the defect. The inspection results are marked on the original image, a detailed inspection report is generated, and the results are stored in a database for reference in quality analysis and process improvement. At the same time, an alarm signal is issued according to the severity of the defect to prompt the operator to take timely action.

Images