Solid wood texture identification and classification method and system based on global features

Abstract

The invention discloses a global feature-based solid wood texture identification and classification method and system, and the system comprises a picture collection unit, a preprocessing unit, a filtering unit and a classification model unit, and is characterized in that a plurality of solid wood composite floor pictures are collected and preprocessed to obtain preprocessed floor pictures; the collected images are classified according to straight lines and patterns, a global feature method is used for conducting floor surface texture feature vector extraction on the preprocessed floor images, feature dimension reduction operation is conducted according to an overall data set in combination with a nonlinear method, and filtering images in different scale directions are obtained; the extracted feature vectors are substituted into a support vector machine model for classification training, meanwhile, optimal hyper-parameters of the support vector machine are obtained through a grid search algorithm in the training process, and the trained model is stored; and the to-be-classified solid wood composite floor pictures are classified through the trained model. According to the method, the influence caused by color difference possibly existing on the wood surface can be overcome, and the operation cost is reduced.

Description

Technical Field

[0001] The invention relates to a solid wood texture recognition and classification method and system based on global features, belonging to the field of machine vision target detection and recognition. Background Art

[0002] In the production process of solid wood composite flooring, classifying surface texture is a complex and critical task. Traditional texture classification methods mainly rely on manual visual inspection, which has the problems of strong subjectivity, low efficiency, high false detection rate and easy fatigue of inspectors. With the rapid development of machine vision technology, visual inspection methods have partially replaced manual automatic inspection. However, due to the color difference and texture diversity on the surface of solid wood, there are many interference factors in the accuracy of classification. In addition, the traditional texture feature extraction method has high computational cost and slow processing speed, which is difficult to meet the production scenarios with high real-time requirements. Therefore, the present invention proposes a solid wood texture recognition and classification method and system based on global features (Gist), which can overcome the influence of surface color difference and reduce computational cost, and provide a new solution for the automatic classification of solid wood composite flooring texture. Summary of the invention

[0003] Purpose of the invention: In order to solve the problem of unsatisfactory texture classification of solid wood composite flooring due to surface color difference, the present invention provides a solid wood texture recognition and classification method and system based on global features. By extracting global features and combining support vector machine classification, the goal of overcoming the influence of surface color difference and reducing computing costs is achieved, thereby reducing the labor cost and false detection rate of enterprises in production and life.

[0004] Technical solution: To achieve the above purpose, the technical solution adopted by the present invention is:

[0005] A solid wood texture recognition and classification method based on global features comprises the following steps:

[0006] Step S1: Collect multiple solid wood composite floor pictures and perform preprocessing to obtain preprocessed floor pictures.

[0007] Step S2: Classify the collected images according to straight lines and patterns, use the global feature method to extract the texture feature vector of the floor surface from the preprocessed floor image, and perform feature dimension reduction operations based on the overall data set combined with a nonlinear method to obtain filtered images in different scale directions.

[0008] Step S3: The extracted feature vector is brought into the Adaboost strategy integration model for classification training. At the same time, the grid search algorithm is used during the training process to obtain the optimal hyperparameters of the support vector machine and save the trained model.

[0009] Step S4: Classify the solid wood composite flooring images to be classified using the trained model.

[0010] Preferably, the method for obtaining filtered images in different scale directions in step S2 is as follows:

[0011] Step S21: First, the input image is grayscaled, resized to 256*256, and normalized.

[0012] Step S22: Establish 8 scales, each scale is a Gabor kernel in 8 directions, that is, select eight different variances of 7, 5, 3.5, 2.5, 1.75, 1.25, 0.875, and 0.625, and each variance calculates the Gabor kernel at eight angles of 0°, 30°, 45°, 60°, 90°, 120°, 135°, and 150°.

[0013] Step S23: Divide the obtained filtered images in different scale directions into multiple 7*7 area blocks and perform average pooling operations.

[0014] Step S24: Combine the means of all small blocks into a feature vector and normalize it by dividing it by the L2 norm of the vector, and finally obtain a 2304-dimensional feature vector.

[0015] Step S25: traverse the data set of 2304-dimensional feature vectors of all images.

[0016] Step S26: Calculate the local reconstruction weight coefficient vector according to the Euclidean distance of each neighboring data point.

[0017] Step S27: Finally, the data points are mapped to a 6-dimensional data space using the calculation matrix derived from the local reconstruction weight coefficient vector to obtain filtered images in different scale directions.

[0018] Preferably: the feature extraction and dimensionality reduction operation in step S2 is as follows:

[0019] Calculate the reconstruction error of the extracted feature vector:

[0020]

[0021] The target mapping space evaluates the quantitative relationship:

[0022]

[0023] Where ∈(W) is the reconstruction error parameter of the target low-dimensional space dimensionality reduction, N is the total number of data, kH represents the height of the pooling window, kW represents the width of the pooling window, input represents the input, m represents the starting position of the height of the pooling window, n represents the starting position of the width of the pooling window, and N iis the batch size, C i Represents the number of channels, Q i is the set of neighboring points of the data point, ω ij For x j x i The reconstruction weight of i The neighboring point set, x represents the horizontal coordinate position of the pixel, y represents the vertical coordinate position of the pixel, θ represents the tilt angle of the Gabor kernel function image, γ represents the aspect ratio, σ represents the standard deviation of the Gaussian function, λ represents the wavelength of the filter, ψ represents the phase offset, min Y trace represents the minimization of the trace operation, Y is the defined embedding mapping matrix, and W is the reconstruction weight ω ij The weight matrix composed of , I is the unit matrix.

[0024] Preferably, the preprocessing method is: firstly grayscale and Gaussian filter the image, then binarize it using the threshold segmentation method, then draw a mask based on the binary image, and overlay it on the original color image to obtain a floor image with an isolated background.

[0025] A solid wood texture recognition and classification system based on global features, used to implement the solid wood texture recognition and classification method based on global features, comprising an image acquisition unit, a preprocessing unit, a filtering unit, and a classification model unit, wherein:

[0026] The picture acquisition unit is used to acquire multiple pictures of solid wood composite flooring.

[0027] The preprocessing unit is used to preprocess the collected multiple solid wood composite floor pictures to obtain preprocessed floor pictures.

[0028] The filtering unit is used to classify the collected images according to straight lines and patterns, use the global feature method to extract the texture feature vector of the floor surface from the preprocessed floor image, and perform feature dimension reduction operations based on the overall data set combined with a nonlinear method to obtain filtered images in different scale directions.

[0029] The classification model unit is used to extract feature vectors and bring them into the multi-support vector machine model based on Adaboost strategy integration for classification training. At the same time, the grid search algorithm is used to obtain the kernel function and penalty factor of the support vector machine with the highest accuracy when completing the classification task during the training process, and then the trained model is saved. The trained model is used to classify the solid wood composite flooring pictures to be classified.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] The present invention converts the original 4-direction and 4-scale Gabor filter extraction features into 8-scale and 8-direction feature extraction algorithms based on improved global features, which can ensure the extraction of more feature information in dealing with complex solid wood textures, and can effectively reduce the model overfitting risk caused by the increasing feature dimension by combining the feature dimension reduction method, and at the same time, the generalization ability of the classifier is further improved by integrating the support vector machine using the Adaboost strategy, so compared with the traditional machine vision classification model, the method proposed by the present invention can effectively overcome the influence of the color difference that may exist on the wood surface, and compared with the traditional color, the texture feature extraction process reduces the computing cost of the computing device hardware. Overall, it has the advantages of good robustness, high recognition accuracy, fast processing speed, and convenient use and maintenance, and better solves the problems of strong subjectivity, low efficiency, high false detection rate, and easy fatigue in the current solid wood composite floor texture classification relying on manual visual defect detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a flowchart of the method of the present invention.

[0033] Figure 2 It is a basic structural diagram of the image acquisition system of the present invention.

[0034] Figure 3 It is a flow chart of the implementation of the overall method of the present invention. DETAILED DESCRIPTION

[0035] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these examples are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0036] In order to achieve the goal of completing texture classification under the premise that the data set contains the influence of color defects, this embodiment provides a solid wood texture recognition and classification method based on global features. First, the surface image of the floor workpiece is obtained by an industrial image acquisition device and the background is isolated by an image processing method. Then, the features in the image are extracted by the Gist method. The dimension of the extracted feature vector is reduced by the local linear embedding method and the corresponding classification label is produced. Finally, the feature vector results are classified by a support vector machine, and the optimal hyperparameters are found by an optimization algorithm, such as Figure 1-3 As shown, the specific steps include:

[0037] Step S1: Using Figure 2The industrial camera shown collects multiple solid wood composite floor pictures. The industrial camera includes a transmission device, a light source, a light source condition device, an industrial linear array CCD camera and a control device image storage, and performs preprocessing to isolate the background and remove noise to obtain preprocessed floor pictures.

[0038] An industrial camera is used to acquire images, and appropriate parameters are adjusted during the process to accurately reflect the texture information of the floor surface.

[0039] Since background information and noise clutter are also stored in the image during the image acquisition process, which has a certain impact on the feature extraction of the image, a series of preprocessing operations are required. In another embodiment of the present invention, the preprocessing method is: first, the image is grayed and Gaussian filtered, where the filter kernel is 3*3; then, the threshold segmentation method is used for binarization; then, a mask is drawn based on the binary image, and it is covered on the original color image to obtain a floor image with an isolated background, such as Figure 3 shown.

[0040] Step S2: Classify the collected images according to straight lines and patterns, use the global feature method to extract the texture feature vector of the floor surface from the preprocessed floor image, and perform feature dimensionality reduction operations based on the overall data set combined with a nonlinear method to obtain filtered images of different scales and directions, and create labels in them.

[0041] Calculate the dimension reduction reconstruction error of the extracted feature vector:

[0042]

[0043]

[0044] Where ∈(W) is the reconstruction error parameter of the target low-dimensional space dimensionality reduction, N is the total number of data, Q i is the set of neighboring points of the data point, ω ij For x j x i The reconstruction weight of i The set of neighboring points of .

[0045]

[0046] H 1 is the average pooling operation, where N i is the batch size, C i Represents the number of channels, the size of the output feature map is 256×256, m represents the starting position of the height of the pooling window, n represents the starting position of the width of the pooling window, kH represents the height of the pooling window, kW represents the width of the pooling window, i is an imaginary unit, out(N i ,Ci ,256,256) represents the output variable of the average pooling operation.

[0047]

[0048] H 2 represents the input variable,

[0049] is the imaginary component of the Gabor filter, is the real component of the Gabor filter, where x, y represent the pixel coordinates, λ represents the wavelength of the filter, θ represents the tilt angle of the Gabor kernel function image, ψ represents the phase offset, σ represents the standard deviation of the Gaussian function, σ=7, and γ represents the aspect ratio.

[0050] Then, the solved multiple groups of reconstruction error ∈(W) functions are decomposed into weight matrices, and the mapping space Y is obtained by solving the generalized eigenvalue method. The calculation formula is as follows:

[0051] min Y trace(Y T (IW) T (IW)Y)stY T Y=I

[0052] Where Y is the defined embedding mapping matrix and W is the reconstruction weight ω ij The weight matrix composed of , I is the unit matrix.

[0053] The global feature method (Gist method) extracts features mainly by generating Gabor filters at different directional scales and performing average pooling calculations on the whole to obtain a complete feature vector. In another embodiment of the present invention, the method for obtaining filtered images of different scale directions in step S2 is as follows:

[0054] Step S21: First, the input image is grayscaled, resized to 256*256, and normalized.

[0055] Step S22: Establish 8 scales, each scale is a Gabor kernel of 8 directions, that is, select eight different variances of 7, 5, 3.5, 2.5, 1.75, 1.25, 0.875, and 0.625, and calculate the Gabor kernel at eight angles of 0°, 30°, 45°, 60°, 90°, 120°, 135°, and 150° for each variance;

[0056] In another embodiment of the present invention, taking Gabor filtering with a scale of 7 and an angle of 45° as an example, the calculation formula is as follows:

[0057]

[0058] The real component is:

[0059]

[0060] The imaginary component is:

[0061]

[0062] in:

[0063] x'=xcos45°+ysin45°;

[0064] y'=-xsin45°+ycos45°;

[0065] Where x, y represent the pixel coordinate position, λ represents the wavelength of the filter, θ represents the tilt angle of the Gabor kernel function image, ψ represents the phase offset, σ represents the standard deviation of the Gaussian function, σ=7, γ represents the aspect ratio, which determines the ellipticity of the Gabor kernel function image, i is the imaginary unit, and x', y' represent the pixel coordinates after transformation.

[0066] In this embodiment, Gabor filters of four scales are applied in four directions respectively, for a total of 16 groups.

[0067] Step S23: Divide the obtained filtered images in different scale directions into multiple 7*7 area blocks and perform average pooling operations.

[0068] In another embodiment of the present invention, the calculation method of the average pooling operation in step S23 is as follows:

[0069]

[0070] Among them, out(N i ,C i ,256,256) represents the average pooling operation output variable, input(N i ,C i ,7×256,7×256) represents the input variable, input variable input(N i ,C i , stride[0]×h+m, stride[1]×ω+n), N i is the batch size, C i Represents the number of channels. The size of the output feature map is 256×256. The pooling window slides the coordinate information of the selected area in steps of 7 to determine the sampling interval of the pooling operation. m represents the starting position of the height of the pooling window, n represents the starting position of the width of the pooling window, kH represents the height of the pooling window, and kW represents the width of the pooling window.

[0071] Step S24: Combine the means of all small blocks into a feature vector and normalize it by dividing it by the L2 norm of the vector, finally obtaining a 576-dimensional feature vector.

[0072] Since the original feature vector dimension is too high, the complexity of the trained support vector machine model will be too high, increasing the risk of overfitting. Therefore, the local linear embedding method is preferably used for feature dimensionality reduction. This method mainly completes the low-dimensional space mapping of the data by finding a fixed number of neighboring data for each data point in each data set and minimizing the reconstruction error.

[0073] Step S25: traverse the data set of 576-dimensional feature vectors of all images. In this embodiment, search for the five data points with the closest Euclidean distance to the target data point.

[0074] Step S26: Calculate the local reconstruction weight coefficient vector according to the Euclidean distance of each neighboring data point.

[0075] Step S27: Finally, the data points are mapped to a 6-dimensional data space using the calculation matrix derived from the local reconstruction weight coefficient vector to obtain filtered images in different scale directions.

[0076] In another embodiment of the present invention, the objective function of the reconstruction error in step S26 is as follows:

[0077]

[0078] Among them, ∈(W) is the reconstruction error, which is used to measure the accuracy of linear reconstruction of data points by their neighboring points, N is the total number of data, x i is the data point, Q i is the set of neighboring points of the data point, ω ij For x j x i The reconstruction weight of i The set of neighboring points.

[0079] Then the method of solving the generalized eigenvalue is used to obtain the mapping space Y, which is calculated as follows:

[0080] M=(IW) T (IW);

[0081] min Y trace(Y T Y T Y=I

[0082] Where M is the intermediate variable, Y is the defined embedding mapping matrix, and W is the number of ω ij The weight matrix composed of , I is the unit matrix.

[0083] Step S3: The extracted feature vector is brought into the support vector machine model for classification training. At the same time, the grid search algorithm is used during the training process to obtain the optimal hyperparameters of the support vector machine and the trained model is saved.

[0084] The grid search algorithm divides the kernel function type and penalty factor of SVM when completing the classification task into a grid with a fixed range and step size and brings it into the model for training and recognition, and saves the parameters used by the model with the best accuracy. The kernel function type ranges from linear kernel function (Linear), radial basis kernel function (RBF), and sigmoid kernel function, the penalty factor ranges from 0 to 100, and the step size is 0.5.

[0085] In another embodiment of the present invention, the images that are misdetected during classification are labeled and added to the training data set, and feature extraction and model training are performed again to optimize the classification accuracy of the model.

[0086] Since the support vector machine as a classifier contains multiple hyperparameters and all of them affect the accuracy of classification, it is necessary to use an optimization algorithm to find the hyperparameters that can maximize the classification accuracy when corresponding to the input feature vector set. Preferably, the present invention selects a grid search algorithm as the optimization algorithm, which mainly traverses and searches for the optimal hyperparameters of the model by generating a grid with a fixed step size.

[0087] In this embodiment, feature vector sets of different categories are randomly scattered and mixed into an overall data set; and the labeled feature vector data set is divided into a training set and a test set according to a certain ratio and then brought into the support vector machine model training. At the same time, a grid search algorithm is used during the training process to obtain the optimal hyperparameters of the support vector machine, and finally the overall integrated model hyperparameters for classification are obtained.

[0088] Step S4: Classify the solid wood composite flooring images to be classified using the trained model.

[0089] After the training is completed, the model is exported and combined with the feature extraction process to achieve the classification of the solid wood composite floor texture, and the classification results are printed on the original image and transmitted to the terminal classification equipment or staff. Combined with the previous feature extraction process, the classification task can be completed using the classifier model (trained model).

[0090] A solid wood texture recognition and classification system based on global features, used to implement the solid wood texture recognition and classification method based on global features, comprising an image acquisition unit, a preprocessing unit, a filtering unit, and a classification model unit, wherein:

[0091] The picture acquisition unit is used to acquire multiple pictures of solid wood composite flooring.

[0092] The preprocessing unit is used to preprocess the collected multiple solid wood composite floor pictures to obtain preprocessed floor pictures.

[0093] The filtering unit is used to classify the collected images according to straight lines and patterns, use the global feature method to extract the texture feature vector of the floor surface from the preprocessed floor image, and perform feature dimension reduction operations based on the overall data set combined with a nonlinear method to obtain filtered images in different scale directions.

[0094] The classification model unit is used to extract feature vectors and bring them into the Adaboost strategy integration model for classification training. At the same time, the grid search algorithm is used to obtain the optimal hyperparameters of the support vector machine during the training process, and the trained model is saved. The trained model is used to classify the solid wood composite flooring images to be classified.

Claims

1. A solid wood texture recognition and classification method based on global features, characterized in that: The following steps are involved: Step S1: collecting multiple solid wood composite floor pictures and preprocessing them to obtain preprocessed floor pictures; Step S2: Classify the collected images according to straight lines and patterns, use the global feature method to extract the texture feature vector of the floor surface from the preprocessed floor image, and perform feature dimension reduction operation based on the overall data set combined with a nonlinear method to obtain filtered images in different scale directions; Step S3: The extracted feature vector is brought into the Adaboost strategy integration model based on the support vector machine model for classification training. At the same time, the grid search algorithm is used during the training process to obtain the optimal hyperparameters of each support vector machine base model, and the trained model is saved; Step S4: Classify the solid wood composite flooring images to be classified using the trained model.

2. The solid wood texture recognition and classification method based on global features according to claim 1 is characterized in that: The feature extraction and dimensionality reduction operation in step S2 is as follows: Calculate the reconstruction error of the extracted feature vector: The target mapping space evaluates the quantitative relationship: min Y trace(Y T (I-W) T (I-W)Y); s.t.Y T Y=I Where ∈(W) is the reconstruction error parameter of the target low-dimensional space dimensionality reduction, N is the total number of data, kH represents the height of the pooling window, kW represents the width of the pooling window, input represents the input, m represents the starting position of the height of the pooling window, n represents the starting position of the width of the pooling window, and N i is the batch size, C i Represents the number of channels, Q i is the set of neighboring points of the data point, ω ij For x j The reconstruction weight of xi, N(i) represents the set of neighboring points of xi, x represents the horizontal coordinate position of the pixel, y represents the vertical coordinate position of the pixel, θ represents the tilt angle of the Gabor kernel function image, γ represents the aspect ratio, σ represents the standard deviation of the Gaussian function, λ represents the wavelength of the filter, ψ represents the phase offset, min Y trace represents the minimization of the trace operation, Y is the defined embedding mapping matrix, and W is the reconstruction weight ω ij The weight matrix composed of , I is the unit matrix.

3. The solid wood texture recognition and classification method based on global features according to claim 2 is characterized in that: The method of obtaining the filtered images in different scale directions in step S2 is as follows: Step S21: firstly, the input image is grayscaled, resized to 256*256, and normalized; Step S22: Generate a Gabor filter and establish 8 scales, each scale is a Gabor kernel in 8 directions, that is, select eight different variances of 7, 5, 3.5, 2.5, 1.75, 1.25, 0.875, and 0.625, and each variance calculates the Gabor kernel at eight angles of 0°, 30°, 45°, 60°, 90°, 120°, 135°, and 150°; Step S23: Divide the obtained filtered images in different scale directions into multiple 7*7 area blocks and perform average pooling operation; Step S24: combine the means of all small blocks into a feature vector, and normalize it by dividing it by the L2 norm of the vector, and finally obtain a 2304-dimensional feature vector; Step S25: traverse the data set of 2304-dimensional feature vectors of all pictures; Step S26: Calculate the local reconstruction weight coefficient vector according to the Euclidean distance of each neighboring data point; Step S27: Finally, the data points are mapped to a 6-dimensional data space using the calculation matrix derived from the local reconstruction weight coefficient vector to obtain filtered images in different scale directions.

4. The solid wood texture recognition and classification method based on global features according to claim 3 is characterized in that: The calculation formula of the Gabor filter with an angle of 45° in step S22 is as follows: The real component is: The imaginary component is: in: x'=xcos45°+ysin45°; y'=-xsin45°+ycos45°; Where x, y represent the pixel coordinate position, λ represents the wavelength of the filter, θ represents the tilt angle of the Gabor kernel function image, ψ represents the phase offset, σ represents the standard deviation of the Gaussian function, σ=7, γ represents the aspect ratio, i is the imaginary unit, and x', y' represent the pixel coordinates after transformation.

5. The solid wood texture recognition and classification method based on global features according to claim 4 is characterized in that: The calculation method of the average pooling operation in step S23 is as follows: Among them, out(N i ,C i ,256,256) represents the average pooling operation output variable, input(N i ,C i ,7×256,7×256) represents the input variable, N i is the batch size, C i Represents the number of channels, the size of the output feature map is 256×256, m represents the starting position of the height of the pooling window, n represents the starting position of the width of the pooling window, kH represents the height of the pooling window, and kW represents the width of the pooling window.

6. The solid wood texture recognition and classification method based on global features according to claim 5 is characterized in that: The objective function of the reconstruction error in step S26 is as follows: Among them, ∈(W) is the reconstruction error, N is the total number of data, xi is the data point, Qi is the set of neighboring points of the data point, ω ij For x j The reconstruction weight of xi, N(i) represents the set of neighboring points of xi; Then the method of solving the generalized eigenvalue is used to obtain the mapping space Y, which is calculated as follows: M=(I-W) T (I-W); min Y trace(Y T MY); s.t.Y T Y=I Where M is the intermediate variable, Y is the defined embedding mapping matrix, and W is the number of ω ij The weight matrix composed of , I is the unit matrix.

7. The method for identifying and classifying solid wood texture based on global features according to claim 6, characterized in that: The preprocessing method is as follows: first, grayscale and Gaussian filter the image; then use the threshold segmentation method to binarize it; then draw a mask based on the binary image and overlay it on the original color image to obtain a floor image with an isolated background.

8. A solid wood texture recognition and classification system based on global features, characterized in that: The method for realizing the solid wood texture recognition and classification method based on global features as described in claim 1 comprises a picture acquisition unit, a preprocessing unit, a filtering unit, and a classification model unit, wherein: The picture acquisition unit is used to acquire multiple pictures of the solid wood composite floor; The preprocessing unit is used to preprocess the collected multiple solid wood composite floor pictures to obtain preprocessed floor pictures; The filtering unit is used to classify the collected images according to straight lines and patterns, extract the texture feature vector of the floor surface from the pre-processed floor image using a global feature method, and perform feature dimension reduction operations based on the overall data set combined with a nonlinear method to obtain filtered images in different scale directions; The classification model unit is used to bring the extracted feature vector into the Adaboost strategy integration model with the support vector machine as the base model for classification training. At the same time, during the training process, the grid search algorithm is used to obtain the optimal hyperparameters of each support vector machine base model, and the trained model is saved; the solid wood composite flooring images to be classified are classified through the trained model.

Images