Key-Point Detection Based on Sparse Coding

Abstract

<p>Numerous key-point detectors such as Harris corner, SIFT, SURF, proposed for detection of corners, blobs, or junctions in images, may demonstrate good performance under geometric changes, yet the very nature of depending on a human-defined structure could make these detectors inflexible to different contexts. Additionally, dramatic non-uniform lighting variations could also degrade their performance. Thus, we aim to develop an integrated key-point detection solution that is flexible to changed scenarios, and robust with images featuring simultaneous geometric as well as non-uniform illumination transformations.</p> <p>To take on this seemingly insurmountable task, we took a progressive approach, splitting the development of the solution into two stages. Firstly, a novel Sparse Coding based Key-point detector (SCK) is proposed to address simultaneously the flexibility and illumination problems, a predicament hindered the performance of almost all the detectors reported in the literature. In SCK, a key-point is mainly determined by a complexity measure calculated from a sparse representation of its surrounding block. Relying on no human pre-designed structure, SCK is expected to be very flexible. Meanwhile, the use of normalization in the sparse coding step results in its robustness with non-uniform lighting change. In the second stage, a Scale and Rotational Invariant Sparse Coding based Key-point detector (SRI-SCK) is introduced, taking the invariance of geometric transformation (scale and rotation) of images into consideration while maintaining other properties of SCK. In SRI-SCK, an automatic means to select key-points’ sizes/scales is proposed. Its rotational invariance is fundamentally created with a combination of multiple rotated versions of the sparse coding dictionary. Along with vigorous verifications, experiments on public datasets have revealed the advantages of SCK and SRI-SCK over other techniques.</p> <p>Finally, we explore SRI-SCK’s application in retinal image registration, a very challenging topic in medical domain, which supports ophthalmologists to diagnose and monitor advancement of several eye disorders such as glaucoma, diabetic retinopathy, and age-related macular degeneration (AMD). Essentially, an automated image registration solution that incorporates SRI-SCK is proposed to fully exploit its flexibility, scale, rotational, and non-uniform lighting invariance. Experiments in different datasets have verified the remarkable performance gain achieved by the solution compared with state-of-the-art works.</p>