Abstract :

The paper suggests an ensemble training method that utilizes a quick adaption of the Fast Adaptive Multivariate Empirical Mode Decomposition (FA-MVEMD) and a convolutional neural network (CNN) for periocular recognition. Initially, the periocular images go through pre-processing. After pre-processing, a single-level discrete wavelet transform decomposes the image, resulting in the LL, LH, HL, and HH coefficients. The LL coefficients undergo extra decomposition into K quantity of Intrinsic Mode Functions (IMFs) using the Fast Adaptive Multivariate Empirical Mode Decomposition (FA-MVEMD). The Intrinsic Mode Functions (IMF), along with the residual and HH coefficients, form K+2 ensemble images. During the training stage, the CNN model is trained with ensemble images. During the testing phase, the K+2 ensemble images are created on the test image. These K+2 ensemble images are classified as yielding K+2 outcomes. The actual classification result is obtained based on the maximum number of similar matching results. The evaluation was performed using metrics such as UBPIr, AR, CASIA Iris rank 1, rank 2, equal error rate (EER), and area under the curve (AUC). The proposed approach achieves rank-1 recognition accuracy of 93.47%, 98.23%, and 98.03% on UBPIr, AR, and CASIA Iris datasets, respectively.