Encoding Image Features
Abstract The characterization of rotation invariant is significant for representing HEp-2 cell images. To improve the classification performance, we propose two kinds of rotation invariant descriptors to characterize HEp-2 cells that are highly discriminative and descriptive with respect to their staining patterns. We firstly propose a rotation invariant textural feature of pairwise local ternary patterns with spatial rotation invariant (PLTP-SRI). The intensity gradients of our HEp-2 cells are weak, especially in the intermediate intensity cells, as shown in Fig.6.1. Local Binary Pattern (LBP) related features are sensitive to noise and smooth weak illumination gradients. To solve the problem, we replace the binary patterns by three-value patterns, which is more efficient than LBP for such a specific classification task. Furthermore, we propose a spatial pyramid structure based on patch-level rotation invariant LTPs to capture spatial layout information.Then, we integrate PLTP-SRI feature and BoW representation into a discriminative and descriptive image representation. Both features are respectively effective for capturing informative characteristics of the staining patterns in their own ways. While our proposed PLTP-SRI feature extracts local feature, BoW builds a global image representation. It is reasonable to extract multiple features for compensation. The combined feature can take the advantages of the two kinds of features in different aspects. We will demonstrate the validity of the proposed feature by experimental results consistently.