Inspecting rice seed species purity on a large dataset using geometrical and morphological features

Although there is a great interest in developing automatical machines for classifying rice seed varieties, it is still unclear if differences in performance of existing techniques come from better feature descriptors or if this is due to varying inter-class or intra-class among the examined species. In this paper, we present a novel method for inspecting purity of rice seed species from the largest number of rice species dataset. The proposed method is conducted utilizing both morphological and geometrical features extracted from high resolution RGB images. Particularly, we take into account relevant pre-processing techniques so that the collected seeds are normalized by their biological structure. As a consequent, the geometrical features at local part of a seed can measured precisely. In addition, whereas existing methods include a limitation number (or a few) of examined species, we construct a dataset a much larger number of species. Because of a sufficient number of species, we can analyze the dependence of a classification performance on similarities of species (or their distinguishable), or types of the extracted features. In the evaluations, we confirm that both morphological features and geometrical features are informative. Combinations of them achieve the highest performances. Extensive evaluations on several schemes of different classifiers as well as several sub-datasets which consist of varying similarity of species are taken into account. These evaluations confirm stability and feasibility of the proposed method.