HEp-2 Specimen Classification Using Multi-resolution Local Patterns and SVM

Abstract

A pattern recognition system was developed to classify immunofluorescence images of HEp-2 specimens into seven classes: homogeneous, speckled, nucleolar, centromere, golgi, nuclear membrane, and mitotic spindle. Root-SIFT features together with multi-resolution local patterns were used to capture local shape and texture information. Sparse coding with max-pooling was applied to get an image representation from these local features. Specimens were classified using a linear support vector machine. Leave-one-specimen-out experiments on the I3A Contest Task 2 data set predicted a mean class accuracy of 89.9%.