Diabetic Retinopathy Classification of Retinal Image using Invariant Histogram of Oriented Gradients (RIHOG) and Improved Fuzzy Rule-Based Classifier (IFRB)

Abstract

Diabetic retinopathy is a main complication of diabetes. Analysing and diagnosing huge amount of images by manual process requires large time. So it makes the necessity of automated techniques. The Diabetic Retinopathy (DR) stage is classified into three classes in this study. They are, mild, moderate/severe and normal Non- Proliferative Diabetic Retinopathy (NPDR).

A rotation invariant histogram of oriented gradients (RIHOG) image descriptor is proposed in this work. The drawbacks of histogram of oriented gradients (HOG) are rectified by this RIHOG and this technique has high sensitivity to rotation of an image. The pixel magnitude values are only considered in HOG but neighbouring pixels magnitude are not considered by this. The relative magnitude and orientation between pixels with its neighbouring pixels are accumulated in this RIHOG method to reduce image rotation sensitivity.

Features are selected using factor analysis method. Improved Fuzzy rule based classifier (IFRB) is used to classify input feature vector in first mechanism. Moderate/severe, Normal and mild NPDR are classified by IFRB.The experimentation is performed by comparing classification of Normal vs moderate/severe NPDR and Mild vs moderate/sever NPDR. 85% of accurate results are proposed by this method. When compared to HOG, high accurate classification and texture classification is shown by proposed RIHOG.