Texture classification using wavelet transform

Abstract

This paper describes an algorithm for classifying textures using wavelet transforms, A set of subband features improves the classification performance and is used to lower the computational complexity. The performance of our algorithm classifies better than tree-structured wavelet transform methods with lower complexity. We show that the choice of wavelet basis is critical.I. INTRODUCTION Texture analysis has found many important applications in such areas as medical imaging, computer vision, and remote sensing. There have been many studies in texture classification based on statistical, structural and spectral properties of textures. Wavelet based methods are shown to be efficient in detection, classification and segmentation. A large class of natural textures can be modeled as quasiperiodic signals whose dominant frequencies are located in the middle frequency channels. To fully utilize the power of the wavelet transform, frequency selective features with great discrimination ability must be extracted. This paper is organized as follows. In section TI, we briefly review the theory of wavelets. Our proljosed feature selection is described in section 111. Experimental results are reported in section IV and we conclude in section V. WAVELETS AND FILTERBANKSThis paper is restricted to the 2-band orthonormal wavelet basis. A 2-band compactly supported orthonormal wavelet basis is characterized by a scaling filter, b(n), and the wavelet filter, h,(n), of finite lengths N and satisfymg the linear condition [ 11(1) n=O and for all kEZ the quadratic condition N -I JI=O The scaling function, e(t), and the wavelet function, q,(t), are defined by the dyadic difference equation i E { O,I} (3) n As can be seen from (3), both the scaling function and wavelet functions depend on the selection of ho(n) and hl(n).In the orthogonal case, the two filters are related as where N is the length of the filters. Therefore, the mother wavelet depends on the selection of the low pass filter coefficients h&). /?,(ti)= (-1)n/?,3(2?d-1-/2) (4) 0-7803-5491-.5/99/$10.00 0 1999 IEEE 1053 111. METHODOLOGY We develop our algorithm by addressing the three main issues of multichannel texture classification: 1) Feature extraction within each channel 2) Channel selection and 3)Feature combination among channels.For signal analysis and classification problems, energy concentration in a subband has been used as a criterion for its fiuther decomposition [2], i.e. at each level, subbands of maximum energy are identified and decomposed further. In these schemes, the energy map of the surviving subbands is used as the feature set.However, for classification purposes, a criterion based on the difference between signals of different classes is preferable. The desired signals may be quite similar in the high-energy bands and may contain significant information in the low energy bands. Instead of decomposing the subbands hrther based on some criterion. e.g. energy concentration, only one level of the wavelet transform decomposition is used and features are obt...

Natural rock images classification using spatial frequency measurement

“…We apply PCA to all sixteen classes of color rock images and select the principle component band, which corresponds to the highest eigenvalue. In the textural feature analysis phrase, we find and plot the relationship graph between each class of rock images and some statistical features [7]: F-norm, Variance, and original SFM, compared with modified version of SFM. The scatter plots in figure 3 to 7, the X-axis corresponds to the value of each feature and Yaxis refers to the classes of each rock image.…”

“…Second group is non-homogenous of dark rock or bright rock, which consist of less finer black grains (class No. 3,7,8,9,10,11,12 and 15), while the third Group is non-homogenous of bright rocks which consist of numerous finer black grains or large black grains (class No. 4,6,13,14 and 16).…”