Effective image retrieval by shape saliences

Abstract

Content-Based Image Retrieval (CBIR) systems havebeen developed aiming at enabling users to search and retrieve images based on their properties such as shape, color and texture. In this paper, we are concerned with shapebased image retrieval. Here, we discuss a recently proposed shape descriptor, called contour saliences, defined as the influence areas of its higher curvature points. This paper introduces a robust approach to estimate contour saliences by exploiting the relation between a contour and its skeleton, modifies the original definition to include the location and the value of saliences along the contour, and proposes a new metric to compare contour saliences. The paper also evaluates the effectiveness of the proposed descriptor with respect to Fourier Descriptors, Curvature Scale Space and Moment Invariants.

Contour salience descriptors for effective image retrieval and analysis

“…Costa's algorithm [6] can be more efficiently implemented (in time proportional to the number of pixels) by using the image foresting transform (IFT) [7]-a graph-based approach to the design of image processing operators based on connectivity [8].…”

“…For a given set S of seed pixels, the IFT can provide the simultaneous computation of the Euclidean distance transform in the cost map C and of the discrete Voronoi regions in the root map R [8]. This operator requires an Euclidean adjacency relation A and a path-cost function f euc defined for any path pZ hp 1 , p 2 ,.,p n i in the graph as q 2AðpÞ0 ðx q Kx p Þ 2 C ðy q Ky p Þ 2 % r 2 ;…”

“…Note that, the main idea is to find for every image pixel p n a path P Ã ðp n Þ from a seed pixel p 1 2S, such that f euc ðP Ã ðp n ÞÞ is minimum. The exact Euclidean distance transform will depend on the appropriate choice of r, as demonstrated in [8]. However, for most practical situations involving eight-connected curves, such as contours and skeletons, rZ ffiffi ffi 2 p is enough [16].…”

“…Its correctness for weaker conditions that are applied to only optimum paths in the graph is presented in [8].…”

“…The narrow band is used to reduce as much as possible cross-influence of opposite parts of the curve, which come close to each other. Torres et al [7] presented a more efficient way to compute shape saliences using the image foresting transform [8] and a contour salience descriptor for image retrieval [9] and analysis [10]. In [9] and [10], the contour salience descriptor was compared with several other shape descriptors, including Image the popular curvature scale space [11,12] and the recently proposed beam angle statistics [13,14].…”