Shape retrieval using angle-wise contour variance

Abstract

In this study, we propose a geometric feature set for 2D shape retrieval. Conventional Hough feature gives the edge locations along with angle and creates Hough table if there are multiple intersections at borders. In this paper, a statistical way to represent the relation of repeating contours at each angle around the shape centroid is presented. The main contribution of this paper is to use the standard deviation of repeating contours. We calculate the angle between the shape centroid and each point on the contour. For each integer angle value, three features were extracted: the number of contour repetitions, the average distance of the points at that angle to the centroid, and the standard deviation of the points at the same angle. Thus, a 2D image was represented by a constant sized matrix, regardless of its size. In the case of similarity between two images, instead of merging features within a single expression, the algorithm picked the feature with the highest similarity rate for that comparison. We tested the proposed method on MPEG-7, Kimia99, ETH-80 datasets for a benchmark with the state-of-the-art. It outperformed most of the recent methods in terms of retrieval rate.

Quadrant-based contour features for accelerated shape retrieval system

“…The proposed method decreases the feature vector size by extracting different types of features but with much less amount per feature. In [21], a shape is represented by 1080 feature samples. We will decrease this number to a very small amount.…”

“…The study [21] gives out 90 numbers of each three features n α , d α and σ α for each quadrant. We introduce three new features by using the features from [21].…”

“…We briefly explain the background of this study. In [21], the authors defined three features by using the contour of a shape. There are 360 imaginary vectors originate from the CoM of the shape and rotating around the clock so that there is a vector traveling in the direction of each integer-valued angle.…”

“…In our previous study [21], we introduced angle-wise contour statistics for shape representation. The method was extracting three features for each angle in range [0, 359 • ] by taking the shape's center of mass (CoM) as the origin.…”

“…This paper focuses on the time consumption problem. We extract the same features as in [21]. Next, we divide the shape into four quadrants.…”