A method for sorting eucalyptus veneers using an image classification model

Abstract

A method for classifying eucalyptus veneer by using an image classification model, comprising the following steps: (1) collecting eucalyptus veneer original image data set and preprocessing, dividing the data set in the proportion of 7-8:2-3 as the training set and test set of the model; (2) model selection analysis is carried out on the image classification algorithm, and the model suitable for eucalyptus veneer classification task is selected; (3) the selected model is optimized in the aspects of efficiency, accuracy and stability; (4) the training set is used to train the optimized image classification model, and the test set is used to test the model performance, and the image classification algorithm is used to classify the eucalyptus veneer; (5) the image classification algorithm is packaged, and is cooperated with the production line motion control system to form a complete set of eucalyptus veneer intelligent classification system. The present application can realize accurate and rapid classification of eucalyptus veneer, and the provided method has high intelligent level and can be applied to the automatic production line of the wood-based panel enterprise.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision and wood-based panel production, and particularly relates to a method for classifying eucalyptus veneer by using an image classification model. BACKGROUND

[0002] Veneer is a thin wood sheet material made from logs by rotary cutting or sawing, which is mainly used for producing plywood and other laminated materials. High-quality veneer is generally used as the face plate of plywood, fine woodworking board and other wood-based panels, and poor-quality veneer is used as the back plate and core plate. Therefore, accurate and rapid classification of wood veneer is an effective means to improve the quality and production efficiency of wood-based panels.

[0003] At present, the classification method of veneer adopted by wood-based panel manufacturers is mainly manual classification. However, human beings are prone to fatigue, resulting in unstable efficiency and accuracy of manual classification. At present, there are few reports on the method of classifying wood veneer by using image classification algorithm technology, but there are many methods of using computer vision to detect wood veneer defects in the field. Scholars first studied the wood defect classification algorithm based on machine learning. However, in actual production, not only should the wood defects be accurately classified, but also the wood defects should be accurately positioned, so scholars have carried out research on the application of target detection algorithm to wood defect identification. According to the above research, deep learning has enough ability to be applied to plywood industrial production. However, previous scholars aimed at the classification and identification of veneer defects, and did not accurately classify and identify the veneer. According to the actual needs of the production line, the present application establishes an intelligent classification model of eucalyptus veneer based on image classification algorithm. SUMMARY

[0004] In view of the problems of low classification efficiency, poor accuracy and low utilization rate of wood veneer in the industry, the present application aims to provide an intelligent classification model of eucalyptus veneer based on image classification algorithm technology, which can realize accurate and rapid classification of eucalyptus veneer, and the method provided has high intelligent level and can be applied to the automatic production line of wood-based panel enterprises.

[0005] To solve the above technical problems, the technical scheme of the present application is as follows: a method for classifying eucalyptus veneer by using an image classification model, comprising the following steps:

[0006] (1) collecting eucalyptus veneer original image data set and performing pretreatment, dividing the treated data set in a ratio of 7-8:2-3 as a training set and a test set of the model;

[0007] (2) performing model selection analysis on the image classification algorithm, and selecting a model suitable for the classification task of eucalyptus veneer;

[0008] (3) optimizing the selected model in terms of efficiency, accuracy and stability.

[0009] (4) Using the training set to train the optimized image classification model, and testing the model performance using the test set, using the image classification algorithm to grade the eucalyptus veneer; the optimized image classification model is to add a spatial pyramid pooling module in the model,

[0010] (5) Encapsulating the image classification algorithm, deploying the eucalyptus veneer sorting system, and working with the production line motion control system to form a complete set of eucalyptus veneer intelligent sorting system.

[0011] The eucalyptus veneer data set preprocessing of step (1) is to use gamma transformation, Laplace transformation, random cropping, random horizontal flip, Gaussian blur, Gaussian noise and salt and pepper noise processing for data enhancement.

[0012] Step (2) the model selection analysis is to compare the training of the front-line algorithm based on visual self-attention model architecture and convolutional neural network architecture, and select the algorithm with high precision and efficiency for the sorting system, which has a continuous improvement channel.

[0013] Step (3) the optimization of the selected model refers to first introducing a spatial attention mechanism module and a channel attention mechanism module in the mobile visual self-attention model to improve the network's ability to focus on key information, make the network adaptively focus on the area of interest, suppress unnecessary features, reduce the interference of texture on classification accuracy, add a spatial pyramid pooling module in the model, which can ignore the input size and produce fixed length output to increase the scale invariance of the image, reduce overfitting, secondly, train the optimized mobile visual self-attention model, during the model training process, use the back propagation algorithm to calculate the gradient of the model and optimize the model parameters to enhance the model's classification ability for different veneer types, at the same time, use data enhancement and noise processing technology to preprocess the original data to improve the model's robustness and generalization ability.

[0014] In step (4), the spatial pyramid pooling module uses two layers of 3x3 convolution and one layer of 1x1 convolution to extract features and compress the number of channels. At the same time, a lightweight design is used to reduce the number of channels of the convolution layer by half to reduce the complexity of the model. In the pooling layer, a combination of maximum pooling and average pooling is used to extract features of different scales. ReLU is used as the activation function to improve the model's nonlinear fitting ability.

[0015] In step (4), the spatial pyramid pooling module is added to the third layer of the convolution layer of the mobile visual self-attention mechanism model, because the front half of the mobile visual self-attention mechanism model contains multiple lightweight convolution layers and attention mechanisms, which can well extract image features. The spatial pyramid pooling is a more comprehensive processing of these features to extract different scale information of the features. Therefore, adding the spatial pyramid pooling module to the third layer of the convolution layer of the mobile visual self-attention mechanism model can better extract features, thereby improving the performance of the model.

[0016] The step (5) of encapsulating the image classification algorithm and cooperating with the production line motion control system refers to first selecting a high-performance GPU server, deploying the trained algorithm to the selected GPU server, using the TensorFlow Serving model service to deploy the algorithm to the GPU server, controlling the execution of the algorithm through the control system of the pipeline equipment, calling the classification model to classify the wood veneer, and outputting the output result of the algorithm to the control system of the pipeline equipment through the device interface. The control system of the pipeline equipment will distribute the wood veneer to the corresponding processing process for processing according to the output result of the algorithm, and finally realize the intelligent detection of eucalyptus veneer.

[0017] The image classification algorithm proposed in the application classifies the eucalyptus veneer, solves the problems of low efficiency and poor accuracy of veneer detection in the plywood factory, and improves the production efficiency and material utilization. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 The technical roadmap of the method for detecting eucalyptus veneer by using the image classification model.

[0019] Figure 2 The intelligent detection system effect diagram of the method for detecting eucalyptus veneer by using the image classification model.

[0020] Figure 3 The data set exploratory analysis diagram of the method for detecting eucalyptus veneer by using the image classification model. The first row from left to right is an image example of the eucalyptus veneer after gray processing, and the second row from left to right is an image example of the corresponding first row A, B and C grade veneer after gamma transformation.

[0021] Figure 4The schematic diagram of data enhancement of the method for detecting eucalyptus wood veneer by using an image classification model according to the present application. From top to bottom, from left to right are respectively a single board original drawing detail example, a single board Gaussian blur processing drawing detail example, a single board Gaussian noise processing drawing detail example, a single board salt and pepper noise processing drawing detail example, a single board random clipping processing drawing example, and a single board random horizontal rotation drawing detail example.

[0022] Figure 5 The training precision comparison chart of two image classification models and their optimized models used in the method for detecting eucalyptus wood veneer by using an image classification model according to the present application. The left chart is a training precision comparison chart of a mobile visual self-attention model before optimization, a mobile visual self-attention model optimized by using a spatial pyramid pooling module, and a mobile visual self-attention model optimized by using an attention mechanism and a spatial pyramid pooling module. Figure Three The bar line from low to high is a training precision comparison chart of a mobile visual self-attention model before optimization, a mobile visual self-attention model optimized by using a spatial pyramid pooling module, and a mobile visual self-attention model optimized by using an attention mechanism and a spatial pyramid pooling module. The right chart is a training precision comparison chart of a residual neural network 50 before and after optimization.

[0023] Figure 6 The confusion matrix of the image classification model. The left chart represents the detection precision confusion matrix based on the image feature engineering method, which represents the precision of the image feature engineering method for classifying eucalyptus wood veneer. The right chart is a detection precision confusion matrix of eucalyptus wood veneer based on a mobile visual self-attention model, which represents the test precision of the mobile visual self-attention model for detecting eucalyptus wood veneer.

[0024] Figure 7 The detection result visualization of the eucalyptus wood veneer grading model based on the mobile visual self-attention model. The left chart is a correct example of detecting A-grade board, which represents that the model evaluates the board as A-grade board with a confidence of 0.98. The middle chart is a correct example of detecting B-grade board, which represents that the model evaluates the board as B-grade board with a confidence of 0.95. The right chart is a correct example of detecting C-grade board, which represents that the model evaluates the board as B-grade board with a confidence of 0.93. DETAILED DESCRIPTION

[0025] The technical solutions of the present application are further described by the following examples, but the protection scope of the present application is not limited thereby. Obviously, the described examples are part of the examples of the present application, but not all the examples.

[0026] Example 1

[0027] The method for detecting eucalyptus wood veneer by using an image classification model according to the present application comprises the following steps:

[0028] (1) Data set sampling

[0029] In the black background, a 2000 million pixel CMOS industrial camera is used with a 1 inch 8mm lens, and the shooting distance is set to 0.8m. The eucalyptus veneer with a size of 1.27m*0.63m is photographed.

[0030] (2) Data set preprocessing

[0031] According to the color and defect characteristics of the wood veneer, data augmentation is performed by using gamma transformation, Laplace transformation, random cropping, random horizontal flipping, Gaussian blur, Gaussian noise and salt and pepper noise processing to increase the generalization ability and robustness of the model, thereby improving the training efficiency and accelerating the training speed. According to the national standard and industry specification, the data set is labeled, and the eucalyptus veneer is tested by grading using the image gray threshold processing method;

[0032] (3) Model selection analysis

[0033] The same data set is used to compare and train the residual neural network algorithm based on the convolutional neural network architecture and the mobile visual self-attention model based on the self-attention model architecture, and the running efficiency and accuracy of the model are tested and compared to select the excellent model (mobile visual self-attention model) as the basic application model.

[0034] (4) Model optimization

[0035] According to the characteristics of irregular size of wood defects in eucalyptus veneer data set and similarity of part of defect information and wood texture, channel attention mechanism module, spatial attention mechanism module and spatial pyramid pooling are added to the mobile visual self-attention model algorithm to optimize modeling, and the test results show that the effect is improved;

[0036] (5) Establishment of eucalyptus veneer intelligent sorting system

[0037] The eucalyptus veneer quality sorting algorithm is coordinated with the production line motion control system to form a complete set of eucalyptus veneer intelligent sorting system under the conditions of fixed shooting speed, conveyor belt speed and light source. The collaborative sorting system completes image acquisition, image processing and data transmission from the host computer (high-performance computer), data reception, data processing, logic control and finally executes the mechanism of the lower computer (embedded system).

[0038] In step (1), the final data set image has a bit depth of 24, a resolution of 5472*3648 and a color space of RGB. A total of 8000 images are collected, including 1500 A-grade boards, 4000 B-grade boards and 2500 C-grade boards.

[0039] In step (2), the eucalyptus veneer is preliminarily sorted by thresholding the gray value of the processed eucalyptus veneer image, and the accuracy is about 0.76, wherein the classification accuracy of A-grade board is 0.93, the classification accuracy of B-grade board is 0.84, and the classification accuracy of C-grade board is 0.51.

[0040] In step (3), the classification accuracy of the residual neural network 50 algorithm based on the convolutional neural network architecture for eucalyptus veneer is 0.85; the classification accuracy of the improved residual neural network 50 algorithm for eucalyptus veneer is 0.92; and the accuracy of the mobile visual self-attention model is 0.99. The model test efficiency is shown in Table 1.

[0041] Table 1 Efficiency of eucalyptus veneer detection model

[0042]

[0043] In step (3), the mobile visual self-attention mechanism model is developed by Apple Inc. The model is open source software using Apache 2.0 license. Apache 2.0 license emphasizes the terms of copyright and patent, and the restrictions on derivative works. According to the license, software using Apache 2.0 license can be used, modified and distributed for free, including private and commercial use. At the same time, the license also allows the inclusion of patent claims in derivative works.

[0044] In step (4), the spatial pyramid pooling module adopts two layers of 3x3 convolution and one layer of 1x1 convolution to extract features and compress the number of channels. At the same time, the invention uses lightweight design to reduce the number of channels of convolution layer by half to reduce the complexity of the model. In terms of pooling layer, the invention adopts the combination of maximum pooling and average pooling to extract features of different scales. Finally, the invention uses ReLU as the activation function to improve the nonlinear fitting ability of the model.

[0045] In step (4), the spatial pyramid pooling module is added to the third layer of the convolution layer of the mobile visual self-attention mechanism model, because the first half of the mobile visual self-attention mechanism model contains multiple lightweight convolution layers and attention mechanisms, which can well extract image features. The spatial pyramid pooling is a more comprehensive processing of these features to extract different scale information of the features. Therefore, adding the spatial pyramid pooling module to the third layer of the convolution layer of the mobile visual self-attention mechanism model can better extract features and improve the performance of the model.

[0046] In step (5), the hardware device is selected to configure a high-performance GPU server to improve the calculation speed and concurrent processing capability of the algorithm.

[0047] In step (5), the algorithm is deployed to the GPU server using the TensorFlow Serving model service.

[0048] In step (5), the data acquisition device selects high-quality, high-resolution cameras or sensors and processes the collected data, such as image enhancement and noise removal.

[0049] In step (5), the output results of the algorithm can be output to the control system of the pipeline equipment through the device interface, so as to proceed to the next intelligent filling processing.

[0050] In step (5), the host computer and the lower computer contain the following components:

[0051] The host computer contains a high-performance computer responsible for collecting and processing eucalyptus veneer images, which requires high-speed, large-capacity memory, multi-core processors, and high-performance graphics cards. Industrial cameras are responsible for collecting eucalyptus veneer images and need to have high resolution, high sensitivity, and high speed. Image processing software is responsible for feature extraction, classification analysis, and other processing operations on collected images, usually requiring efficient, accurate algorithms and computing resources. Communication interfaces are responsible for data transmission between the host computer and the lower computer, usually supporting high-speed, reliable data transmission protocols such as Ethernet, serial ports, USB, etc. The lower computer contains an embedded system responsible for receiving data sent by the host computer and performing data processing and logical control, usually requiring high reliability, low power consumption, and high integration. For example, industrial controllers, single-board computers, etc. Sensors are responsible for detecting and measuring eucalyptus veneer to obtain data and perform analysis. For example, laser sensors, photoelectric sensors, etc. Actuators are responsible for implementing classification, sorting, and processing of eucalyptus veneer. Including conveyors, cutting machines, sanding machines, painting machines, etc. Communication interfaces are responsible for data transmission between the host computer and the lower computer, usually supporting high-speed, reliable data transmission protocols such as Ethernet, serial ports, USB, etc.

Claims

1. A method for sorting eucalyptus veneer using an image classification model, characterized in that, Includes the following steps: (1) Collect the original image dataset of eucalyptus veneer and preprocess it. Divide the dataset into 7-8:2-3 ratios as the training set and test set of the model. (2) Conduct model selection analysis on image classification algorithms and select models suitable for eucalyptus veneer sorting tasks; (3) Optimize the selected model to improve efficiency, accuracy, and stability; The optimization of the selected model refers to first introducing spatial attention and channel attention modules into the mobile vision self-attention model to improve the network's ability to focus on key information, enabling the network to adaptively focus on regions of interest, suppress unnecessary features, and reduce the interference of texture on classification accuracy. A spatial pyramid pooling module is added to the model, which can ignore the input size and produce a fixed-length output to increase the scale invariance of the image and reduce overfitting. Secondly, the mobile vision self-attention model is trained. During the model training process, the backpropagation algorithm is used to calculate the model's gradient and optimize the model parameters to enhance the model's classification ability for different thin plate types. At the same time, data augmentation and noise reduction techniques are used to preprocess the original data to improve the model's robustness and generalization ability. (4) The optimized image classification model is trained using the training set and the model performance is tested using the test set. The eucalyptus veneer is graded using the image classification algorithm. The optimized image classification model is a model with the addition of a spatial pyramid pooling module. The spatial pyramid pooling module employs two 3×3 convolutional layers and one 1×1 convolutional layer to extract features and compress the number of channels. It also utilizes a lightweight design, halving the number of channels in the convolutional layers to reduce model complexity. For the pooling layers, a combination of max pooling and average pooling is used to extract features at different scales. ReLU is used as the activation function to improve the model's non-linear fitting ability. The spatial pyramid pooling module is added to the third convolutional layer of the mobile visual self-attention mechanism model. This is because the first half of the model contains multiple lightweight convolutional layers and an attention mechanism, which effectively extract image features. Spatial pyramid pooling, on the other hand, processes these features more comprehensively, extracting information at different scales. Therefore, adding the spatial pyramid pooling module to the third convolutional layer of the mobile visual self-attention mechanism model allows for better feature extraction, thereby improving model performance. (5) The image classification algorithm is encapsulated and worked in conjunction with the production line motion control system to form a complete intelligent sorting system for eucalyptus veneer.

2. The method for sorting eucalyptus veneer using an image classification model according to claim 1, characterized in that, The eucalyptus veneer dataset preprocessing in step (1) involves data augmentation using gamma transform, Laplace transform, random cropping, random horizontal flipping, Gaussian blur, Gaussian noise, and salt-and-pepper noise.

3. The method for sorting eucalyptus veneer using an image classification model according to claim 1, characterized in that, The model selection analysis described in step (2) involves comparing and training cutting-edge algorithms based on visual self-attention model architecture and convolutional neural network architecture, selecting algorithms with superior accuracy and efficiency to apply to the sorting system, which has a continuous improvement channel.

4. The method for sorting eucalyptus veneer using an image classification model according to claim 1, characterized in that, Step (5) involves encapsulating the image classification algorithm and cooperating with the production line motion control system. First, a high-performance GPU server is selected and configured. The trained algorithm is then deployed to the selected GPU server using TensorFlow Serving. The algorithm is executed by the control system of the production line equipment. The classification model is called to classify the wood veneer, the target detection algorithm is called to detect defects, and the defect contour segmentation algorithm is called to segment the defect. The output of the algorithm is output to the control system of the production line equipment through the equipment interface for the next step of intelligent patching.

Images