Showing papers by "Michihiko Minoh published in 1997"
TL;DR: The clustering method repeatedly extracts mutually close samples as a cluster and leaves isolated noises unclustered, so the produced clusters are less affected by noises than those of C-Means.
Abstract: Partitional clustering methods such as C-Means classify all samples into clusters. Even a noise sample that is distant from any cluster is assigned to one of the clusters. Noise samples included in clusters bias the clustering result and tend to produce meaningless clusters. Our clustering method repeatedly extracts mutually close samples as a cluster and leaves isolated noises unclustered. Thus, the produced clusters are less affected by noises than those of C-Means. Because clusters can be obtained analytically by our method, repeated trials to avoid local minima are not necessary. The method is shown to be effective for extracting straight lines from images in the experiments.
6 citations
22 Aug 1997
TL;DR: A segmentation method based on velocity of the movements of a drawing pen shows that the segmented parts of the artistic drawing correspond to the physical objects to be represented.
Abstract: We propose a segmentation method of an artistic drawing into parts, each of which represents a physical object. Our method is based on velocity of the movements of a drawing pen. The artistic drawings are segmented into parts, each of which is drawn with almost constant velocity of the pen movements. We conducted experiments and the results show that the segmented parts of the artistic drawing correspond to the physical objects to be represented.
1 citations
TL;DR: This method improves the accuracy of identification and reduces the number of samples required in a practical operation, compared with conventional methods using a single feature, and an integration of features based on a simple linear connection is more effective than the Dempster-Shafer probability model.
Abstract: This paper proposes a method of identification of an individual by integrating multiple features, such as front facial image, walking image, and vocal features. This method improves the accuracy of identification and reduces the number of samples required in a practical operation, compared with conventional methods using a single feature. These three features are integrated at a similarity level in the proposed method. Identification experiments using 33 individuals show that the proposed method can identify 81% (maximum) of the individuals, compared with 60% (maximum) by a conventional method. This paper also shows that an integration of features based on a simple linear connection is more effective than the Dempster-Shafer probability model.
1 citations