In this chapter, we have proposed a novel framework for automatic staining pattern classification of HEp-2 cells. The characteristic of rotation invariant is essential for HEp-2 cell classification, since most of the HEp-2 cells in the image present large orientation variances. To this purpose, we incorporate two kinds of rotation invariant descriptors to represent HEp-2 cells. The first is a rotation invariant textural feature based on ternary patterns and considering spatial layout information. It is highly descriptive, robust and rotation invariant. Additionally, we exploit the BoW framework to encode dense SIFT features. To incorporate spatial information, spatial pyramid structure is introduced to further improve the robustness to rotational variances.
Both features are respectively effective for capturing informative characteristics of the staining patterns in their own ways. While our proposed PLTP-SRI extracts local feature, BoW builds a global image representation. They are fused into a powerful image representation by taking the advantages of the two kinds of descriptors in different aspects. The final classification is performed by a linear SVM classifier.
The proposed algorithm was fully evaluated on two HEp-2 cell datasets: the ICPR2012 dataset and the ICIP2013 training dataset. The effectiveness of the proposed algorithm is proven by the experimental results. In particular, it significantly outperforms the winner of the ICPR’12 contest.