Convex hull-based distance metric learning for image classification

Abstract

Distance metric learning is an important topic in visual classification tasks. Learning an appropriate distance measure can greatly improve the performance of image recognitions. In this paper, we model an image set as a regularized convex hull and propose a distance metric learning method for image classification based on regularized convex hulls. In particular, a regularized point-to-convex hull distance metric (RPCHD) and a regularized convex hull-to-convex hull distance metric (RCHCHD) are introduced to measure the distance between the query image object and the existing image sets for image classification tasks. The coefficients in distance metric are solved by an approximate optimization strategy. Two SVM-like distance metric learning models are constructed and transformed into the standard support vector machines to learn distance metric matrix in RPCHD and RCHCHD. Positive and negative sample pairs are proposed to represent distance metric matrix. Experiments on three image databases show that our proposed RPCHD and RCHCHD can effectively improve the performance of image classification.