Feature extraction and description for retinal fundus image registration / Roziana Ramli

Roziana , Ramli (2020) Feature extraction and description for retinal fundus image registration / Roziana Ramli. PhD thesis, Universiti Malaya.

	PDF (The Candidate's Agreement) Restricted to Repository staff only Download (220Kb)
	PDF (Thesis PhD) Download (23Mb)

Abstract

Retinal fundus image registration (RIR) is performed to align two or more fundus images. A general framework of a feature-based RIR technique comprises of preprocessing, feature extraction, feature descriptor, matching and estimating geometrical transformation. The RIR is mainly performed for super-resolution, image mosaicking and longitudinal study applications to assist diagnosis and monitoring retinal diseases. Registering image pair from these applications involve a combination of challenges such as overlapping area and rotation between images. The challenges of the overlapping area and rotation can be addressed at feature extraction and feature descriptor stages of the feature-based RIR technique, respectively. To address the overlapping area, a reliable and repeatable anatomical information such as retinal vessels is required. However, finding the feature points on retinal vessels can be challenging due to noises with similar structure representation as the vessels. For rotation, a distinctive descriptor is necessary to characterise the feature points on retinal vessels that are lack of textural information and exhibit repetitive patterns in the local patches of fundus image. Therefore, this study proposed new feature extraction and feature descriptor methods for feature-based RIR technique to address these issues. The proposed feature extraction method extracts the feature points on retinal vessels by considering the characteristics of the retinal vessels and noises. The proposed feature descriptor method characterises the feature points with statistical properties obtained from the surrounding region of the feature points. The proposed work is tested on five public datasets, namely, CHASE_DB1, DRIVE, HRF, STARE and FIRE. Aspects of the evaluation include evaluating the extraction accuracy of the proposed feature extraction method and the registration accuracy of the proposed feature-based RIR technique. Experimental results show that the proposed feature extraction method attained the highest overall extraction accuracy (86.021%) and outperformed the existing feature extraction methods; Harris corner (41.613%), SIFT (16.164%), SURF (18.929%), Ghassabi’s (28.280%) and D-Saddle (20.509%). The registration success rate of the proposed feature-based RIR technique (67.164%) is also outperformed the existing feature-based RIR techniques; Harris-PIIFD (3.731%), GDBICP (27.612%), Ghassabi’s-SIFT (12.687%), H-M 16 (16.418%), H-M 17 (19.403%) and D-Saddle-HOG (11.940%). The influence of the overlapping area and rotation on the proposed feature-based RIR technique are significant but the weakest among the evaluated feature-based RIR techniques.

Actions (For repository staff only : Login required)