A novel description based on skeleton and contour for shape matching

Abstract

In computer vision field, feature extraction plays a critical role in shape matching, image alignment, object recognition and tracking etc. Generally speaking, feature extraction consists of three steps: feature detection, feature description and feature matching. In the second step, the detected features (e.g. gray value, SIFT, Harris corners) are converted to vectors or the form that can be described mathematically such that feature can be matched correctly. How to construct an efficient descriptor to realize accurate shape matching under a variety of transformations is still a challenge. To this end, a novel shape descriptor based on skeleton for shape matching is proposed in this paper. Firstly, the image is smoothed with Gaussian filter to remove the influence of the noise. Secondly, the smoothed image is segmented with Fuzzy C-means Cluster (FCM) to obtain a binary image. Thirdly, the binary image’s skeleton is extracted with Medial Axis Transform (MAT), thus the skeleton’s endpoints and joint-points locations are acquired. Furthermore, the object’s contour is extracted with contour coding. In the construction of skeletal descriptor, the relative location vectors of the skeletal endpoints to each contour point are computed. Being similar to shape context, statistical histogram is constructed in log-polar coordinate. Consequently, shape matching is performed via two histograms’ similarity measurement. Experiments on standard MPEG7 dataset show that the proposed shape description method allows translation, scale and rotation invariance.