Segmentation of SAR images using similarity ratios for generating and clustering superpixels

Abstract

The superpixels are groups of similar neighbouring pixels which are perceptually meaningful and representationally efficient segments. Among those existing superpixel generating algorithms, simple linear iterative clustering (SLIC) seems to be one of the simplest ones. Its simplicity is due to adaption of a distance measure which is a linear combination of colour and spatial proximity. It is this measure that is modified using a similarity ratio. This modified measure is used to label the pixels within the search areas for generating the superpixels. This generation phase is further augmented with a clustering phase based on the same formulated similarity metric, which clusters the superpixels into larger segments. It has been demonstrated that this modified version performs better in terms of boundary recall and undersegmentation error, and is more robust to the speckle noise than the one in SLIC. Moreover, the clustered segments formed by superpixels generated by this approach has better boundary adherence than those formed by superpixels generated by SLIC.