Fast Nonlinear Filter-Based Local Phase Quantization for Texture Classification

Texture divides an image into subparts called regions. Texture analysis characterizes region using texture content. Local phase quantization is a conventional method, which gives highest classification accuracy. Bilateral filter is also a good approach to obtain smoothness of a digital image while preserving the edges. Also, fast bilateral filter is another approach for high dynamic range images. In this study, new approach is proposed by integrating local phase quantization and fast bilateral filter, which in turn results in good classification accuracy. Using different filter domain parameter and filter range parameter texture, features such as mean, standard deviation, entropy, skewness, and kurtosis are extracted. Finally, for classification, these features are given to k-nearest neighbor (K-NN) classifier. The new hybrid technique is for testing and training images from Brodatz database. The results of proposed technique are compared with conventional local phase quantization results. Best classification accuracy is obtained by hybrid method for different values of domain and range parameters with different window sizes.