Image retrieval by pattern categorization using wavelet domain perceptual features with LVQ neural network

Abstract

For the efficient and cost effective management of large volume of images in textile industry, an effective retrieval system is expected. Textile (e.g., curtain) images of raw clothes have wide varieties of design patterns. Despite many research works in this area, only a few emphasize on complex pattern characteristics. Such patterns are horizontal, vertical, cross-stripes, leaves and flowers in curtain database. In this study, we propose a system that retrieves images based on wavelet domain perceptual features which mainly depend on edge and correlation characteristics of the wavelet sub-bands in the major directions (horizontal and vertical). In order to reduce searching time, we first catagorize various patterns using supervised learning vector quantization (LVQ) technique. Then for each category or group, a prototype vector is formed by averaging all classified feature vectors in it. For a typical query, the query key is first compared with a few prototype vectors to determine the expected category. Then the query key performs similarity comparisons with the population of that particular group and retrieves relevant images. Users have also the provision to select subsequent similar groups if any query fails to capture the correct group at first attempt. An experiment with a set of curtain images shows the effectiveness of the proposed features compared to conventional Gabor, pyramidal wavelet transform (PWT) or local binary pattern (LBP) features.