Using Multi-Descriptors for Real Time Cosmetic Image Retrieval

Abstract

Cosmetic Image Retrieval (CIR) is a methodology for searching and retrieving images from Cosmetic Image Collection (CIC). There are numerous cosmetic brands whose types are similar to others. In addition, there are not trivial to retrieve cosmetic images because of its complexity and duplicative shape, as well as detail of various cosmetic items. We present a method for CIR using multi-descriptors, combining global and local features for image descriptors. Along with integrating a Scale-Invariant Feature Transform (SIFT) and Critical Point Filters (CPFs) to achieve accuracy and agility in CIR processing, called CPF level 9 & SIFT. SIFT is used for detailed-image, such as cosmetic image, to reduce the time complexity for extracting keypoints. On the other side, CPF will filter only for the critical pixel of the image. From the experiment, our method can reduce computation time by 50.46% and 99.99% by using SIFT and CPF respectively. Moreover, our method is preserved efficiency, measured by precision and recall of CPF level 9 & SIFT, which is as high as the precision and recall of SIFT.