Methods of Classification

Various texture classification methods are used in this research,

1.    Grey Level Co-occurrence Matrices (GLCM)

The GLCM studies the relationship of grey pixel pairs in the image. A secondary statistical value can be extracted as the features, 5 features are commonly used as textural features in texture classification problems, i.e. Contrast, Correlation, Energy, Entropy and Homogeneity. Other than the second-order features, the raw GLCM can be used as the features as well after downsizing.

2.    Gabor Filters

Gabor filters are also known as Gabor wavelets, it is a signal processing method that analyzes the frequency domain rather than spatial domain. The Gabor filters can be generated with different orientation and radial center frequency (or scale). The features can be downsized using singular value decomposition (SVD).

3.    Covariance Matrix

The covariance matrix represents the covariance between different values. Here the covariance matrix is generated from different feature images. The feature images are images or matrices with the same size and are generated using feature extraction algorithms, such as edge-based derivatives, GLCMs and Gabor filters.

4.    Verification-based Recognition

The verification-based algorithm calculates each training sample as a template. The test sample will be verified against all templates. The templates accept the test sample as the same species if it falls under the threshold. The species with highest templates accepted will be considered as the winning species and as the result.

An Overall System

An overall system is shown in Figure 2. After the image acquisition of the wood sample, the image of the wood sample will passed through image pre-processing, feature extraction and classification to obtain the results.

Figure 2. Overview of the system

Methods of Classification

For feature extraction, the various texture classification methods used in this research are as follows:

1.      Grey Level Co-occurrence Matrices (GLCM)

The GLCM studies the relationship of grey pixel pairs in the image. A secondary statistical value can be extracted as the features, 5 features are commonly used as textural features in texture classification problems, i.e. Contrast, Correlation, Energy, Entropy and Homogeneity. Other than the second-order features, the raw GLCM can be used as the features as well after downsizing.

2.      Gabor Filters

Gabor filters are also known as Gabor wavelets, it is a signal processing method that analyzes the frequency domain rather than spatial domain. The Gabor filters can be generated with different orientation and radial center frequency (or scale). The features can be downsized using singular value decomposition (SVD).

3.      Covariance Matrix

The covariance matrix represents the covariance between different values. Here the covariance matrix is generated from different feature images. The feature images are images or matrices with the same size and are generated using feature extraction algorithms, such as edge-based derivatives, GLCMs and Gabor filters.

4.      Verification-based Recognition

The verification-based algorithm calculates each training sample as a template. The test sample will be verified against all templates. The templates accept the test sample as the same species if it falls under the threshold. The species with highest templates accepted will be considered as the winning species and as the result.

Outcome of Research

The comparison of the experimental results show that the covariance matrix using feature images generated using Gabor filters has the highest accuracy of 85% compared to other techniques on six species of wood as shown in Table 1.

 Table 1. Comparison of results for different texture classification methods

The raw GLCM is tested to be significantly faster than the Gabor filters and is therefore selected for implementation in the embedded platform where the comparison of time is shown in Table 2.

 Table 2. Comparison of computational time on different platforms for raw GLCM

The system is also tested on a sample application that allows the user to load an unknown wood image and query for the closest match from the database as shown in Figure 3.

Figure 3. Sample application on searching for the most similar wood samples

 Conclusion

The research has shown the usefulness of several texture classification methods on the problem of wood species recognition with the covariance matrix using feature images generated using Gabor filters providing the best accuracy. However, due to the slower computation of Gabor filters, the raw GLCM is implemented onto the embedded system for faster computation. The research show the wood species recognition can be achieved using texture classification method and can be deployed onto the embedded systems.

Acknowledgement

This research is partly funded by Malaysian MOSTI ScienceFund 01-02-11-SF0019.