Exploiting the fine-grained similarity of a large-scale rice species using shape motif discovery

In this paper, we propose a method to exploit trait similarity of a large-scale of rice species. Particularly, fine-gained challenges are handled when we distinguish highly similar species/varieties of rice. To this end, we attempt to construct a large-scale dataset images of ninety rice species. The collected images are normalized in term of both the size and biological structures of a rice seed. We then convert 2-D boundary of a rice seed image as a normalized time series. To find structure and regularities of each species from the collected datasets, we consider the problem of discovering shape-based motifs, which are approximately repeated shapes within (or between) rice species. Distance of a two seeds in each species is calculated using Dynamic Time warping algorithms. The k-level motif is specified by ordering the DTW distances. Two applications of the motif discovery are presented. The first one specify and visually distinguish two similar rice species. This distinguishing suggests valuable features to boot performances of recognition techniques for the rice species classification. The second application presents a visual classification by constructing a similarity map of a large number of rice species. We produce a similar map of ninety species of rice seeds using shape-based motif discovery techniques.