Retinal Blood Vessels Segmentation Based on Multi-Classifier Fusion

In this paper, we propose a supervised learning method based on multiple features and multi-classifier fusion to segment retinal blood vessels. We treat vessel segmentation as a pixel-wise classification problem, so feature extraction is a key step. Four types features, including contrast enhanced intensity, B-COSFIRE filter response, line strength and fused gray-voting result with the 2D-Gabor filter result, are extracted. For classifier decision, we fuse results of decision tree and AdaBoost algorithm classifier, because we find that both classifiers get different even complementary results in experimental process. Finally, the used method is evaluated in public DRIVE database, our approach achieves average classification accuracy of 95.91% and sensitivity is high 82.15 %.