Contrast context histogram—An efficient discriminating local descriptor for object recognition and image matching

Abstract

In this paper, we propose a new invariant local descriptor, called the contrast context histogram (CCH), for image matching and object recognition. By representing the contrast distributions of a local region, it serves as a distinctive local descriptor of the region. Our experiments demonstrate that contrast-based local descriptors can represent local regions with more compact histogram bins. Because of its high matching accuracy and efficient computation, the CCH has the potential to be used in a number of real-time applications.

Introduction

Invariant local descriptors constructed from images have been proposed as a way to solve many problems, such as in image matching [1], object recognition [1], [2], [3], and many other vision-based applications [4], [5], [6], [7]. The idea is to detect the invariant local properties of salient image corners under a class of transformations, and then establish discriminating descriptors for these corners. Descriptors provide robust representations of local image regions, even under partial occlusion. The basic problem is how to find the relevant information required to encode the local signatures.

Descriptors of local features have received considerable attention in recent years. For example, Freeman and Adelson [8] developed steerable filters, in which filters of arbitrary orientations are synthesized from linear combinations of pixel derivatives in particular directions. Belongie et al. [2] proposed a feature description called the shape context, which is a histogram of edge points with respect to a reference point under the log–polar coordinate. Lowe [1] introduced a scale invariant feature transformation (SIFT) descriptor that is invariant to scale and rotation. In this approach, keypoints are computed through the detection of scale-space extremes in a series of difference-of-Gaussian (DoG) images. Local descriptors are built up for each keypoint based on a weighted histogram of gradient orientations from a patch of pixels in its local neighborhood. In Refs. [3], [9], it is shown that SIFT is one of the most effective matching approaches when scale and viewpoint changes occur in images. Various extensions of SIFT have been proposed. For example, Ke and Sukthankar proposed PCA-SIFT [10], which applies principal components analysis (PCA) [11] to a normalized gradient patch, instead of using SIFT's smoothed weighted histograms. The gradient location-orientation histogram (GLOH) [3] computes the SIFT descriptor for a log–polar location grid and then reduces the size of the descriptor with PCA. The primary focus of these extensions is to provide more distinctive and compact descriptors to improve the matching accuracy and speed.

In this paper, we propose a novel invariant local descriptor called the contrast context histogram (CCH) for image matching and object recognition. Our primary motivation is to develop a descriptor that is computationally fast, requires fewer histogram bins to represent a local region, and can achieve a good matching performance. CCH exploits the contrast properties of a local region, instead of storing the weighted edge orientation histograms of salient corners like the SIFT approach. Rotation and linear illumination changes are considered to make the CCH robust against geometric and photometric changes. Compared to the approaches such as SIFT (PCA-SIFT and GLOH) that require computing the gradient orientations of all the pixels in a region, CCH is more efficient to compute since it only evaluates the intensity differences between the center pixel and the other pixels in a region. Therefore, CCH is potentially more suitable for real-time applications such as augmented reality [5]. In the experiments, we use CCH descriptors to represent cluttered scenes and objects, and evaluate the method's effectiveness.

The remainder of the paper is organized as follows. In the next section, we describe the construction of CCH descriptors from the salient corners of images. Section 3 discusses the implementation of the CCH approach. The experiment results are reported in Section 4. Finally, in Section 5, we present our conclusions.