Hybrid deep Ensemble for Fine-Grained Race Estimation

Race identification has made advances over the past few years, and finds application in numerous areas including surveillance, law enforcement, and even in administrative policy. It is, yet, limited to the groups with major diversifications like Asian, African, Caucasian, and Indian and is still under-explored for sub-racial categorization probably due to the unavailability of sufficient data. This article is aimed at estimating the recognized sub-groups of the ‘Indian’ class. An exclusive dataset (North East India Dataset) has been created for this purpose. Also, a hybrid ensemble blending the concepts of Bagging and Stacking (Super Learner) with minute customizations has been put forward that trains on specialized augmented patterns of data along with intermediate features. Extensive experiments have been conducted and the findings reveal the potential effectiveness of the method especially in lowering the loss. The model has also been validated against another benchmark dataset, FairFace. On the North East India and FairFace set, the proposed model achieves an accuracy of 91% and 80% respectively, outperforming the base models by at least 4%. The proposed ensemble performs better than the standard classical methods such as Bagging and Boosting (AdaBoost) on both datasets.