Hierarchical Ensemble of Global and Local Classifier for Texture Classification

In this paper, we propose a novel hierarchical ensemble classifier for texture classification by combining global Fourier features and local Gabor features. Specifically, in our method, global features are extracted from images firstly by 2D Discrete Fourier Transform. Then, real and imaginary components of low frequency band are concatenated to form a single feature set for further processing. Gabor wavelet transform is exploited for local feature extraction. Firstly, Gabor wavelets are used to extract local features from the whole image. Then, these features are spatially partitioned into a number of feature sets, each corresponding to a local patch of the image. After the above processes, an image can be represented by one Global Fourier Feature Set (GFFS) and multiple Local Gabor Feature Sets (LGFSes). These feature sets contain different discriminative information: GGS contains global discriminative information and each LGFS contains different local discriminative information. In order to make full use of all these diverse discriminative information, we propose multiple component classifiers by applying Fisher Discriminant Analysis (FDA) on GFFS and each LGFS, respectively. At last, we combine them into one ensemble by weighted sum rule.