The Study On Texture Feature Extraction Method Based On LBP And GLCM

With the development of the Internet and multimedia technology, extensive use of digital images has resulted in the fact that image processing and analysis is increasingly becoming important. Texture feature is a low level feature of images and can synthetically reflect the changes of gray-value, spatial distribution and structural information. Thus extracting effective texture features plays an important role in image retrieval. Feature extraction and classification for images are two stages of image classification. After many years of research, the similarity measurement method has become mature and therefore, this thesis focuses on the image texture feature extraction technology.Relative phase is a new feature extraction technology and in the transform domain of an image, phase information shows the positions of the varying and the modulus of coefficients reflect the strengths of the varying. LBP can reflect the microstructure of an image, which is widely used to classify texture images and can obtain higher classification accuracy. The thesis puts forward LRPBP(Local Relative Phase Binary Pattern) based on LBP and relative phase. In LRPBP, relative phase information is obtained by Gabor transform, and then LBP is used to extract texture features. The experimental results show that LRPBP can get higher classification accuracy than LBP.An image contains a lot of information and a feature of an image can only extract a part of information. Image can be classified better based on multi-features. Therefore, we combine LBP and LRPBP in this paper in order to improve the image retrieval accuracy. Experiment results show that the classification effect of texture image based on multi-features is better than that of single feature.Moreover, the thesis proposes a new method of rotation invariant texture classification based on Gray-Level Co-occurrence Matrix(GLCM)in wavelet domain. In this method, GLCM is calculated from low pass subband in the wavelet transform domain and the statistical features of low pass and high pass subbands are also calculated. In the end, the two results are combined together as texture features to classify texture images. Experiments show that the feature extraction method based on GLCM and statistical features can achieve better classification results in rotation invariant texture classification.