Binary wavelet transform-based completed local binary pattern texture descriptors for classification of microscopic images of hardwood species

This paper presents an approach for generating a binary wavelet transform-based completed local binary pattern (BWTCLBP) texture descriptor to improve the classification accuracy of microscopic images of hardwood species. Firstly, gray-level slicing method is used to obtain eight (b₀–b₇) bit planes from grayscale image. Then, the two-dimensional binary wavelet transform (2D-BWT) decomposes each of the most significant bit-plane (b₇) images up to the fifth scale of decomposition. The texture descriptors are then acquired from each of the subimages up to the five scales of decomposition. Further, two variants of support vector machine (SVM), linear SVM and radial basis function kernel SVM, were employed as classifiers. The classification accuracy of the proposed and existing texture descriptors was compared. The BWT-based uniform completed local binary pattern (BWTCLBP^u2) texture descriptor achieved the best classification accuracy of 95.07 ± 0.72% at the third scale of decomposition. The classification accuracy is produced by linear SVM classifier for full feature (1416) vector data. In order to overcome the effect of curse of dimensionality, the minimal-redundancy–maximal-relevance feature selection method is employed to select the best subset of feature vector data. This approach has resulted in improved classification accuracy of 96.60 ± 0.80% (450) by linear SVM classifier.