Research on feature extraction and filter independence in ICA-based texture classification
Xiaohong Xu, YANG Xuezhi, Jun Tao Gao, DeMei Yang · 2010
A new feature extraction method was proposed by exploring non-gaussianity of ICA coefficients. Features with respect to asymmetry and sparsity of ICA coefficients were extracted,which were shown to be robust to outliers and achieve better classification performance than previously reported ICA features. Furthermore,the impact of filter independence on various texture features was investigated,and the relationship between the non-gaussianity of ICA coefficients and the discriminating power of features were further revealed.