Robust Face Detection using Convolutional Neural Network

Abstract

Faces epitomize multifaceted dimensional meaningful visual stimuli which is a challenge for face detectors in detecting faces which is not in perfect conditions, a situation which happens often than not in real life, hence difficult developing a model for its recognition computationally. In this study, recognition rate, classification performance, estimation rate and preprocessing, and execution time of facial detection systems are improved. This is supported by the implementation of varied approaches. The face detection aspect is handled by the adaptation of Viola Jones descriptor and down-sampled by the Bessel transform which reduces feature extraction space to augment processing time. Gabor feature extractions were passed afterwards to extract thousands of facial features representing various facial deformation patterns. A deep convolutionary based Ada-boost hypothesis is carried out to select a few out of the many neurons features extracted to augment classification which are later fed into the classifier through a back-propagation algorithm. The convolutional neural network (CNN) make available for partial invariance to translation, rotation, scale, and deformation which extracts uninterruptedly larger features in a hierarchical set of layers. The results of the proposed approach were very encouraging and demonstrate superiority when compared with other state-of-the-art techniques.

Enhancing speaker verification accuracy with deep ensemble learning and inclusion of multifaceted demographic factors

“…The performance analysis of the proposed model for speaker verification is carried out in terms of equal error rate (EER) and minimum decision cost function (minDCF) [30]. EER is a performance metric representing the point where the false acceptance rate (FAR) and false rejection rate (FRR) are equal.…”

“…The minDCF is calculated by weighting the FAR and false rejection rate FRR according to each error type's cost. The performance of the proposed speaker verification system is compared with the existing work proposed by Yao et al [30] in terms of both EER and minDCF. Yao et al [30] presented a speaker verification system based on the temporal embeddings obtained from multiple streams.…”

“…The performance of the proposed speaker verification system is compared with the existing work proposed by Yao et al [30] in terms of both EER and minDCF. Yao et al [30] presented a speaker verification system based on the temporal embeddings obtained from multiple streams. The CNN classification model is implemented and validated on the same dataset adopted in our work.…”

“…The CNN classification model is implemented and validated on the same dataset adopted in our work. Table 4 illustrates the numerical value and comparative analysis for multistream CNN [30] and the proposed speaker verification method.…”