Geometric corner extraction in retinal fundus images

Abstract

This paper presents a novel approach of finding corner features between retinal fundus images. Such images are relatively textureless and comprising uneven shades which render state-of-the-art approaches e.g., SIFT to be ineffective. Many of the detected features have low repeatability (<; 10%), especially when the viewing angle difference in the corresponding images is large. Our approach is based on the finding of blood vessels using a robust line fitting algorithm, and locating corner features based on the bends and intersections between the blood vessels. These corner features have proven to be superior to the state-of-the-art feature extraction methods (i.e. SIFT, SURF, Harris, Good Features To Track (GFTT) and FAST) with regard to repeatability and stability in our experiment. Overall in average, the approach has close to 10% more repeatable detected features than the second best in two corresponding retinal images in the experiment.

A low-dimensional step pattern analysis algorithm with application to multimodal retinal image registration

“…For each feature point, the intensity gradient vectors within its neighbors are collected in histograms to form a descriptor of 128 dimensions. However, the algorithm fails to identify adequate, stable, repeatable, and uniformly distributed feature points in multimodal retinal images [6,18,20], and it is more suitable for monomodal image registration [6]. Therefore, enhancement methods are proposed.…”

“…It follows by performing bilateral matching. However, the Harris corners are not uniformly distributed [9,18] and the repeatability rate is poor when the scale changes between images go beyond 1.5 or in the presence of pathologies in the retina [9]. To circumvent the problems, Harris method is replaced with an uniform robust SIFT (UR-SIFT) [9] method.…”

“…The repeatability expresses the reliability of a detector for finding the same features under different viewing conditions. We exploit the geometric corner extraction method [18] to locate hypotheses of robust corner features based on edges from the edge maps. In the method, edges are extracted using a strip fitting algorithm [19].…”

“…Due to the well-known fragmentation of edge maps of real images, missing edge pixels will result in the fracturing of an edge. To circumvent this problem, a post-processing step [18] is applied to identify and fix broken edges with end-points and angles of close proximity. The purpose is to eventually remove edges that are isolated or insignificant, e.g.…”

“…edges < 5 pixels which are mainly noise. The subsequent step is to locate intersecting points from connecting edges which are called geometric corners [18]. Geometric corners are different from appearance-based corners such as Harris corners, as geometric corners are always true corners where each geometric corner comes from an intersecting point of two edges ℓ 1 and ℓ 2 .…”