Micro image classification of 19 high-value hardwood species based on texture feature fusion

Abstract

For classification of wood species with similar microstructure, 19 high-value hardwood species of Papilionaceae, Ebenaceae, and Caesalpiniaceae were used as experimental objects. Images of transverse sections, radial sections, and tangential sections were collected by Micro CT. Local binary patterns (LBP) are often used for feature extraction. LBP deformed forms such as uniform LBP, rotation-invariant LBP, and rotation-invariant uniform LBP were fused with Gray-Level Co-Occurrence Matrix (GLCM) to form three fusion features. The fusion features were combined with support vector machine (SVM) or BP neural network to realize wood classification. The texture feature fusion method was found to be better than the single feature classification. Among them, the effect of uniform LBP and GLCM fusion was the best, and the classification accuracy combined with SVM was the highest. The evaluation of the classification of 19 kinds of hardwood mainly depended on transverse sections, and its accuracy was higher than that of the radial and tangential sections. Therefore, the classification results of transverse sections should be taken as the main evaluation basis for the classification of the 19 high-value hardwood species.

Full Article

Micro Image Classification of 19 High-value Hardwood Species Based on Texture Feature Fusion

Xiaoxia Yang,^a Hailan Jiang,^a Lixin Ma,^a Wenhu Yang,^a,* Xiaohan Zhao,^a Cuiping Hu,^b and Zhedong Ge ^c

For classification of wood species with similar microstructure, 19 high-value hardwood species of Papilionaceae, Ebenaceae, and Caesalpiniaceae were used as experimental objects. Images of transverse sections, radial sections, and tangential sections were collected by Micro CT. Local binary patterns (LBP) are often used for feature extraction. LBP deformed forms such as uniform LBP, rotation-invariant LBP, and rotation-invariant uniform LBP were fused with Gray-Level Co-Occurrence Matrix (GLCM) to form three fusion features. The fusion features were combined with support vector machine (SVM) or BP neural network to realize wood classification. The texture feature fusion method was found to be better than the single feature classification. Among them, the effect of uniform LBP and GLCM fusion was the best, and the classification accuracy combined with SVM was the highest. The evaluation of the classification of 19 kinds of hardwood mainly depended on transverse sections, and its accuracy was higher than that of the radial and tangential sections. Therefore, the classification results of transverse sections should be taken as the main evaluation basis for the classification of the 19 high-value hardwood species.

DOI: 10.15376/biores.18.2.3373-3386

Keywords: High-value hardwood; Feature fusion; Texture feature; Image classification

Contact information: a: School of New Generation Information Technology Industry, Shandong Polytechnic, Jinan, Shandong 250101 China; b: Linshu County Garden Sanitation Guarantee Service Center, Linyi, Shandong276700 China; c: School of Information and Electrical Engineering, Shandong Jianzhu University, Jinan, Shandong 250101, China; *Corresponding author: Yang_mess@163.com

INTRODUCTION

Wood has renewable and environmentally friendly advantages. It can be used as the raw material for flooring, furniture, buildings, and interior decoration. Its use is related to the species and features. Many high-value species with better quality are imported from abroad, so it is of great significance for wood to be classified correctly. Traditionally, experts slice the wood and use optical light microscopy to observe its transverse sections, radial sections, and tangential sections (Barmpoutis et al. 2018; Jahanbanifard et al. 2020). However, this method is cumbersome, time-consuming, and requires highly trained professionals. In this paper, micro computed tomography (micro CT) was used to collect high-resolution microscopic images. This is a more advanced method that allows faster collection of data. Large numbers of images on the three sectional planes are collected in a very short time. These abundant images are the premise of wood classification and identification by computer vision and machine learning. At present, the majority of species identification is completed by computer vision with digital images (Hu et al. 2019; Cao et al. 2021; Fabijańska et al. 2021), which is more efficient and accurate than the traditional method (Kamal et al. 2017).

Machine learning and pattern recognition have been applied to macroscopic and microscopic images recognition of wood, and the method of texture feature extraction is widely used in wood classification (Zhao et al. 2014). Local phase quantization and linear discriminant method were used to classify 77 commercial timber species, and the average success rate was approximately 88% (Rosa et al. 2017). Local binary patterns (LBP) combined with support vector machine (SVM) were used to classify 2240 microscopic images of 112 wood species, and the accuracy reached 98.6% (Martins et al. 2013). Rotation-invariant LBP and SVM were used to classify 46 Brazilian species, and the accuracy reached 97.67%. By using the fusion of classifiers, the accuracy rate was improved by 0.33% (Souza et al. 2020). Six texture feature extraction techniques and three classifiers were used to classify hardwood species images. The results showed that the reduced dimension texture feature uniform completed local binary pattern (DWTCLBPu2) combined with LDA classifier had the best classification accuracy (Yadava et al. 2015). A multi-resolution texture classification method based on anisotropic diffusion and local directional binary pattern (LDBP) was used to evaluate and compare four texture data sets, and the results showed its effectiveness (Hiremath et al. 2017).

In addition to LBP, by using Gray Level Co-Occurrence Method (GLCM) and Euclidean distance method to identify the swietenia mahagoni wood defects, the optimal average accuracy reached 88.90% (Riana et al. 2021). The Improved-Basic Gray Level Aura Matrix (I-BGLAM) feature extraction method was proposed, and the back-propagation neural network classifier was used to realize the automatic classification of 52 kinds of wood (Zamri et al. 2016). The Lucy Richardson algorithm was used to enhance blurred images, the statistical feature extraction technology and SVM were used to identify 20 tropical wood species, and the classification accuracy reached 89.3% (Rajagopal et al. 2019). A new Gabor method for wood recognition was proposed, and the method was evaluated on the wood database of Zhejiang Agriculture and Forestry University (ZAFU), which show that the method had achieved progress relative to the state of the art (Wang et al. 2013). LBP and GLCM were used to extract features through X-ray tomography, and K-Nearest Neighbor (KNN) classifier was employed in order to classify six kinds of diffuse porous wood (Kobayashi et al. 2019). A method of combining Tamura and GLCM texture features was proposed, and BP neural network was used to classify the wood surface defects, thus improving the accuracy and robustness of recognition (Xie and Wang 2015).

The above-cited scholars took transverse sections as the object, but the radial and tangential sections also provide a lot of information. Barmpoutis et al. (2018) proposed a new approach for automating wood species recognition through multidimensional texture analysis. The accuracy performance of transverse-section classification was 91.47% and 84.2% for radial and tangential sections. Up to the present, there have been few studies on classification and recognition of high-value hardwood species from the transverse, radial, and tangential sections. In this paper, three-section images of 2 μm resolution were obtained by Micro CT, and 19 hardwood species with similar structure were classified and recognized. In this work, LBP deformation forms uniform LBP, rotation-invariant LBP, and rotation-invariant uniform LBP were fused with GLCM features to form three fusion features. Compared with a single texture feature, three fusion features combined with SVM and BP neural network showed the highest accuracy and the best recognition effect. Among them, the uniform LBP and GLCM fusion method was the optimum. Combined with SVM and BP neural network, it achieved the best classification effect on the transverse, radial and tangential sections.

EXPERIMENTAL

Materials

The experimental materials were taken from the Specimens Museum of the Shandong Jianzhu University (Jinan China). A total of 19 kinds of hardwood with similar micro-structure were selected, and the samples were cut to 2.5 mm × 2.5 mm × 25 mm for scanning. The Bruker Kscan1272 was used for scanning, and the voltage and current values of the radiation source were set to 50 kV and 200 μA. A total of 3000 images were obtained from each sample, 1000 for transverse sections, 1000 for radial sections, and 1000 for tangential sections. The size was 2042 × 1640 pixels, and the resolution was 2 μm. Table 1 shows the 19 high-value hardwood species.

Table 1. Information Related to the High-value Hardwood Species under Study

Dataset Construction

A total of 300 images were randomly selected from the three sections of the 19 wood species, 100 images for each section. The selected images were cut randomly into sub-images with size of 500 × 500 pixels, which were used to construct the three-section data sets. In sum, there were 5700 microscopic sub-images obtained, 1900 sub-images for each section.

Table 2 shows the microscopic images of transverse, radial, and tangential sections. A variety of structures are seen in the microscopic images, such as vessel pores, axial parenchyma, wood rays, etc.

Table 2. The Three-Section Images of 19 High-Value Hardwood Species

Feature Fusion

The surface of wood is distributed with different structures, which constitute the texture of wood, and the texture distribution of different sections varies greatly. In this work, a method is proposed, namely a texture feature fusion for uniform LBP, rotation-invariant LBP, and a rotation-invariant uniform LBP that is fused with GLCM. The fusion method is as follows:

To reduce the LBP dimension and shorten the calculation, 90% texture modes and information of the microscopic images were grouped to construct Uniform LBP (Ojala et al. 2002, 2013). The U value was limited to being no greater than 2, and calculation was carried out using Eq. 1:

(1)

To maintain the rotation of LBP, the circular neighborhood was rotated one circle clockwise to generate different binary codes. Take the minimum value in the codes as the Uniform LBP to describe the invariant rotation feature. The formula is as follows.

(2)

Combining with the above process, the Uniform LBP with invariant rotation feature is constructed, and the formula is as follows.

(3)

The radius P is equal to 8, and the feature dimensions of , , are 59, 36, and 10, respectively.

Based on the GLCM, five features were obtained, including energy, contrast (CON), correlation (COR), entropy (ENT), and inverse difference moment (IDM). According to these features, feature data sets of 64, 41, and 15 dimensions were constructed for each image. Comparing with traditional LBP, this method is lightweight, and it reflects the IDM of texture images in gray distribution, the differences among textures, the similarity of rows’ and columns’ gray values, the complexity, and roughness of textures (Haralick 1979).

Fig. 1. Feature fusion

In this paper, uniform LBP, rotation-invariant LBP, and rotation-invariant uniform LBP were fused with GLCM texture features, and are expressed as , , , as shown in Fig. 1. The method of texture feature fusion can integrate all kinds of information in the image and improve the classification accuracy (Yusof et al. 2010; Yan et al. 2021).

Classification

According to the three fused features, namely , , and , the BP neural network and SVM were used for classification. The data of training set and testing set were 70% and 30%, respectively.

The BP neural network was set up as a three-layer network structure, including an input layer, a hidden layer, and an output layer. The number of features of each image in the input layer was 59, 36, 10, 64, 41, 15, the number of neurons in the hidden layer was 5, 10, 15, 20, 25, 30, 35, 40, and the classification accuracy of 19 species was in the output layer. The input layer to hidden layer activation function was tanh, and the hidden layer to output layer activation function was Purelin. The training times were 1000 and the learning rate was 0.01, the error rate was 0.00001, the momentum parameter was 0.01, and the minimum performance gradient was 1e-6. The structure of BP neural network is shown in Fig. 2.

The SVM was based on LIBSVM software package, and the Radial Basis Function (RBF) was applied. The value range of penalty factor C and kernel parameter g was 2-10-210 (Chih and Lin 2011). The number of features of each image imported was 59, 36, 10, 64, 41, and 15. These numbers were randomly arranged, and 5-fold cross validation was used to determine the optimal values of C and g. Then the classification accuracy for the 19 species was output.

Fig. 2. Classification structure of BP neural network and SVM

RESULTS AND DISCUSSION

BP Neural Network

During the classification by BP neural network, the number of neurons in the hidden layer affects the classification accuracy. Table 3 shows the maximum classification accuracy and time point corresponding to the number of optimal hidden-layer neurons with different features. The increase of feature dimension leads to more time required for classification. In the classification of single feature, the accuracy of in transverse, radial, and tangential sections was 94.73%, 86.77%, and 86.39%. The accuracy of in three-section was 93.50%, 86.28%, and 85.96%. The accuracy of in three-section was 95.78%, 88.19%, and 88.05%. In the classification of single feature, the training time used by was shorter, but the accuracy was low. exhibited the highest classification accuracy in transverse, radial, and tangential sections, and the classification effect was relatively good.

After feature fusion of the classification of transverse, radial, and tangential sections, the accuracy of increased by 0.7%, 3.75%, and 1.68% in comparison to , the accuracy of increased by 1.23%, 3.19%, and 0.88% in comparison to , and the accuracy of increased by 1.58%, 3.21%, and 2.03% in comparison to , respectively. These findings indicate that the classification result of was the best, and the classification accuracy for the three-section set of images was able to reach over 90%.

Table 3. Classification Results of BP Neural Network

As shown in Fig. 3, the accuracy of transverse sections classification using the above 6 feature classification methods were much higher than that of radial and tangential sections. Therefore, the classification results of transverse sections should be considered in the process of tree species classification. The results of wood species classification on the three sections showed that the accuracy of the method was the lowest, and the accuracy of the combination after feature fusion was the highest. After feature fusion, the classification accuracy of , , was generally higher than that of single feature such as , , . As the increase of feature dimension can improve the classification accuracy, the results considering both classification accuracy and time point after extension of time showed that, for the 19 kinds of high-value hardwood species and 6 feature classification methods, performed the best in accuracy and time.

Fig. 3. Classification results of the three sections (transverse, radial, and tangential) by BP neural network

Support Vector Machine (SVM)

The classification results of 19 high-value hardwood by SVM classifier are shown in Table 4. In the classification of single feature, the accuracy of in transverse, radial, and tangential sections was 94.92%, 90.52%, and 90.05%, respectively. The accuracy of in the three sections was 93.68%, 90.70%, and 89.82%, respectively. The accuracy of in the three sections was 95.27%, 91.40%, and 91.40%, respectively. In the classification of single feature, the training time used by was shorter, but the accuracy rate was low. achieved the highest classification accuracy in transverse, radial, and tangential sections, and the classification effect was relatively good.

After feature fusion of the three sections, the classification accuracies of were increased by 1.04%, 0.70%, and 1.17% compared to , and the accuracies of