Topological Property-based Clustering of Self-Organizing Feature Maps(Pattem Recognition)

スポンサーリンク

概要

• 論文の詳細を見る

• The Self-Organizing Feature Map(SOFM) that is one of unsupervised neural networks is a very powerful tool for data clustering and visualization in high-dimensional data sets by projecting the original data sets into a low-dimensional regular gird. When the number of SOFM grids is large, similar nodes of grids, referred to as 'prototypes','clusters', and 'neurons', need to be grouped. In previous work, three concepts are mainly used to group the similar prototypes: 1) manual visual-inspection requiring human intervention, 2) clustering technologies, such as agglomerative clustering, and 3) supervised techniques, such as Learning Vector Quantization(LVQ). This paper proposes an automatic clustering method of SOFM without supervised techniques, which means performing clustering with only given data sets without any additional information, such as labels. Previous clustering-based work using agglomerative clustering, k-means, and Gaussian mixture models did not consider characteristics of the SOFM preserving topology in the sense that it keeps similar property nearby, as they cluster only weights of maps produced after training of SOFM. This paper use mode-seeking approach on distance matrices of map based on the assumption that these distances are approximately inversely proportional to the density of the data and modes-based tentative clustering on SOFM using distance between coordinates and weights of maps to keep topology preservation. Moreover, this paper estimates mode-based intra-cluster distance to finally merge the clusters, as definitive clustering. In our implementation applied to texture segmentation using Brodatz texture images and artificial data, the proposed method improves precision rates than previous clustering-based method by keeping topology preservation, and reduces computational times by using weights of only modes to estimate the intra-cluster distance

• 社団法人電子情報通信学会の論文
• 2006-11-17

著者

• Jung Keechul

  Hci Lab School Of Media College Of Information Technology Soongsil University

• Jung Keechul

  Hci Lab. School Of Media College Of Information Science Soongsil University

• Park Anjin

  Hci Lab. School Of Media College Of Information Science Soongsil University

関連論文

• Fast Dimension Reduction of 3D Voxel for Hand Gesture Recognition(Biometrics2)
• Efficient 3D Voxel Reconstruction using Precomputing method for Gesture Recognition(Gestures)
• Topological Property-based Clustering of Self-Organizing Feature Maps(Pattem Recognition)

スポンサーリンク