A novel approach of retina-based biometric authentication using multifractal analysis and wavelet fusion

These days, biometric recognition knows a full expansion, given the number of works that have been published in this field. When retinal images are addressed to be the biometric pattern due to their high degree of uniqueness and accuracy, researchers focus more on the vascular network segmentation, which is not an easy task. In the present work, we propose a new approach considering all the eye fundus texture to extract the main features leading to biometric authentication. This technique combines the well-known multifractal analysis to the image wavelet fusion. Since there exists different approaches to obtain multifractal attributes, we try in this work to compare two methods; one based on Legendre spectrum calculation, and the second relies on the detrended fluctuation analysis which is widely used in many fields to describe complexity and irregularity in signals and images. All relevant characteristics are extracted from Retina Identification Database (RIDB) and then assessed via Support Vector Machine (SVM) and K-Nearest Neighbor (KNN) classifiers. The results obtained from Legendre spectrum are slightly better than those of the Multi-Fractal Detrended Fluctuation Analysis (MFDFA) method, especially with the SVM classifier where the accuracy achieved 94%, while the MFDFA approach gives a rate of 90%. This is due to parasitic multifractality that can be induced by small offsets caused by automatic Region Of Interest (ROI) selection and fusion operations. By introducing the entropy as an additional parameter of classification in order to improve the previous results, we managed to achieve 99% and 95% classification rates for Legendre and MFDFA methods respectively.