Cell image classification by a scale and rotation invariant dense local descriptor

This work tackles the problem of indirect immunofluorescence images classification. In particular, a dense
local descriptor invariant both to scale changes and to rotations is proposed to classify six classes of
staining patterns of the HEp-2 cells. In order to provide a compact and discriminative representation, the
descriptor combines a log-polar sampling with spatially-varying gaussian smoothing applied on the gradients images in specific directions. Bag-of-Words is finally used to perform classification. Experimental
results on the dataset provided in the recent contest hold in 2014 at ICPR show very good performance.