An Improved Algorithm for the Detection of Diabetes using Iris Images

Abstract

With the advancement in computer technology, the computer vision system has found its way in various applications like object recognition, industrial defect inspection, and biometrics. It links to various fields using image processing. Iridology is also based on this method, in which the evaluation our internal organ is done by looking at an image of the iris. Iridology is based on the iris recognition process, in which the extraction of the iris code is done by using basic iris recognition steps after this image registration is done, and i.e. comparison of iris image with the iris chart. For this process, a number of methods have been introduced, which make iris recognition method acceptable. Therefore, it is really a scientific knowledge that can be used as a pre-diagnostic tool in several cases. In this work, a high-end research is going in a particular direction for the detection of Diabetes. The work is basically carried out in five steps, namely; iris image acquisition, image pre-processing which includes (iris localization/segmentation, iris normalization, and image enhancement). In proposed work image localization, segmentation, and enhancement of iris is based on Circular Hough Transform. Iris normalization is carried after the extraction region of interest and the segmented iris is then normalized based on Daugman’s Rubber sheet model. The novelty of the work lies in the facts that only Region of Interest is normalized rather than the methods where whole iris image was first normalized and then Region of Interest was extracted and also for the detection of diabetes both left and right eye images are used while in previous work only left eye images are considered. Also, the features are extracted from the iris images, by employing Discrete Cosine Transform (DCT) rather than Discrete Wavelet Transform (DWT). Finally, Support Vector Machine (SVM) is used for the classification. For good performance of Iris Recognition scheme, segmentation and normalization play an important role. The discussed method is tested on a database of 59 images containing 47 of diabetic patients and 12 being healthy. A comparative study with existing ones was carried out and simulation results demonstrate that an improved accuracy of 5.72% has been obtained with proposed technique.