Morphological and Texture Features for HEp-2 Cells Classification

Abstract

This paper describes our texture descriptor ensemble aimed to compete for the Cell Level classification task (Task 1) in the "Contest on Performance Evaluation on Indirect Immunofluorescence Image Analysis Systems", hosted by the I3A Workshop on Pattern Recognition Techniques for Indirect Immunofluorescence Images. Our system is based on the combination of 4 descriptors based on Local Binary Pattern (LBP) and 1 morphological feature set: the multiscale Pyramid LBP, Local Configuration Pattern, Rotation Invariant Co-occurrence among adjacent LBP, Extended LBP and finally Strandmark morphological features. From each image a total of 2643 features are extracted. The corresponding 5 feature sets are classified using Support Vector Machines and results are combined according to the sum rule. By using a 10-fold cross validation testing protocol, the proposed ensemble obtains 60.9% of accuracy, outperforming many state-of-art stand-alone texture descriptors as well as other ensembles.