Advanced leaf image retrieval via Multidimensional Embedding Sequence Similarity (MESS) method

A novel method for shape analysis and similarity measurement is introduced based on a time series matching approach. It applies to shapes represented through one-dimensional signals and has as objectives to utilize efficiently the provided information and to optimize the shape matching process. The new technique is tested on boundaries from leaf images, after their conversion into 1D sequences using either the Centroid Contour Distance (CCD) or the Angle code (AC) measurements. In the core of the new method lies the 'time delay'-based transformation of a given 1D sequence to an ensemble of vectors embedded in a multivariate phase space. The resulting point set is considered as representative of the leaf identity. Inter-leaf comparisons are carried out in a pairwise fashion by employing the multidimensional, Wald-Wolfowitz, statistical test for the 'two-sample problem', which implicitly performs shape matching and similarity quantification. The comparative experimentation shows that the complexity of our method is moderate, while the leaf retrieval performance, compared to that achieved by standard matching procedures usually employed with the CCD and AC representations, is greatly improved.