Segmentation, registration,and selective watermarking of retinal images

Abstract

In this dissertation, I investigated some fundamental issues related to medical image segmentation, registration, and watermarking. I used color retinal fundus images to perform my study because of the rich representation of different objects (blood vessels, microaneurysms, hemorrhages, exudates, etc.) that are pathologically important and have close resemblance in shapes and colors. To attack this complex subject, I developed a divide-and-conquer strategy to address related issues step-by-step and to optimize the parameters of different algorithm steps. Most, if not all, objects in our discussion are related. The algorithms for detection, registration, and protection of different objects need to consider how to differentiate the foreground from the background and be able to correctly characterize the features of the image objects and their geometric properties. To address these problems, I characterized the shapes of blood vessels in retinal images and proposed the algorithms to extract the features of blood vessels. A tracing algorithm was developed for the detection of blood vessels along the vascular network. Due to the noise interference and various image qualities, the robust segmentation techniques were used for the accurate characterization of the objects’ shapes and verification. Based on the segmentation results, a registration algorithm was developed, which uses the bifurcation and cross-over points of blood vessels to establish the correspondence between the images and derive the transformation that aligns the images. A Region-of-Interest (ROI) based watermarking scheme was proposed for image authenticity. It uses linear segments extracted from the image as reference locations for embedding and detecting watermark. Global and locally-randomized synchronization schemes were proposed for bit-sequence synchronization of a watermark. The scheme is robust against common image processing and geometric distortions (rotation and scaling), and it can detect alternations such as moving or removing of the image content.