The features pool consists of the gray-level of green channelof RGB retinal image (green), group of five features based onthe green gray-level (f1, f2, f3, f4, f5), group of eight featuresbased on Hu moment-invariants (Hu1, Hu2, Hu3, Hu4, Hu5,Hu6, Hu7, Hu8) and the gray-level of the vessels-enhancedimage (Ive) 7. Most of the vessels segmentation approachesextract and use the green color image of RGB retinal imagefor further processing since it has the best contrast betweenvessels and background so it’s taken as feature. The five graylevel based features group is presented by Marin et al. 9 andits features describe the gray-level variation between vesselpixel and its surrounding.
The Hu moment-invariants are bestshape descriptors which are invariant to translation, scale androtation change. So they are used by the second group ofeight features to describe vessels have variant widths andangles. The vessels-enhanced image 9 is better enhancingblood vessels while removing the bright retinal structuresas optic disc and exudates, so it’s used for computing thegroup of eight Hu moment invariants based features and itsgray-level (Ive) is the new added feature to the previousfeatures pool as first improvement. These features are simple,better discriminate between vessel and non-vessel classes andneedn’t be computed at multiple scales or orientations.
Thefeatures computation is more detailed in 7.B. Correlation Based Feature Selection HeuristicIt’s a heuristic approach for evaluating the worth or meritof a subset of features 11. The main premise behind thisselection method is that the features that are most effective forclassification are those that are most highly correlated with theclasses (intensifiers and dissipaters), and at the same time areleast correlated with other features. The method is thereforeused to choose a subset of features that best represent thesequalities.
The best individual feature based on the followingmerit metric: