Eye Detection Using Ensemble of Weak Classifiers Based on Correlation Filter

Abstract

This work proposes a novel system for detecting racial landmarks in images using an ensemble of correlation-based filters known as Inner Product Detector (IPD). This work has three main contributions: i) the usage of a bootstrap aggregating algorithm (bagging), to produce a ensemble classifier with higher accuracy when compared with the original IPD detector; ii) a new discriminant function based on the highest IPD mean value calculated from samples positively classified in a voting scheme; iii) and a study to assess the influence of class unbalance over the system performance. The proposed method was evaluated on the BioID and LFPW datasets, achieving an average accuracy of 93.3% in the BioID for both eyes, at 10% of the interocular distance, and accuracies of 85.2% and 81.6% for the left eye and right eyes respectively, on the LFPW database, at 10% of the interocular distance. Since it can detect the eyes at approximately 70 FPS in a Matlab implementation, the proposed method is also fast enough to be used in real time applications. These results were compared to the ones in the state of the art in eye detection - which include methods using deep learning - in terms of accuracy and computational complexity.