Using rival penalized competitive clustering for feature indexing in Hong Kong's textile and fashion image database
Abstract
Efficient content-based information retrieval in image databases depends on good indexing structures of the extracted features. While indexing structures for text retrieval are well understood, efficient and robust indexing structures for image retrieval are still elusive. We use the rival penalized competitive learning (RPCL) clustering algorithm to partition extracted feature vectors from images to produce an indexing structure for Montage, an image database developed for Hong Kong's textile, clothing, and fashion industry supporting content-based retrieval, e.g., by color, texture, sketch, and shape. RPCL is a stochastic heuristic clustering method which provides good cluster center approximation and is computationally efficient. Using synthetic data, we demonstrate the recall and precision performance of nearest-neighbor feature retrieval based on the indexing structure generated by RPCL.
Authors: Irwin King, Lei Xu, Laiwan Chan