Video Face Recognition via Learned Representation on Feature-Rich key Frames
Abstract
Nowadays, facial verification based authentication has become an important area of research in law enforcement and surveillance applications to combat widespread occurrences of security threat incidents,. It is satisfactory that the existing methodologies have marked great verification results at equal error rate, but have poor results at lower fault acceptable rates Therefore further research is required to increase verification performance while lowering the fault acceptance rate.
This paper, introduces a new face matching algorithm, that constitutes 1) Discrete Wavelet Transform and Entropy calculations for getting feature-rich key frames of a video .Then,
2)A Deep Learning Architecture, that consists of a Stacked Denoising Sparse Autoencoder (SDAE) and deep Boltzmann machine (DBM); is helpful for feature extraction..finally After completion of all these steps 3) a multilayer Feed Forward Neural Network System has been used as a classifier to get proper verification result.
The output is analyzed on two different openly accessible databases, YouTube Video Faces and Android Mobile Phone Database a type of point and shoot challenge(PaSC).
Results shows that the algorithm is successful: 1) in achieving sharpen increase in performance compared to histogram based frames, arbitrary frames, or frame collection with no reference image eminence measures 2) joint feature learning in SDAE and sparse and low rank regularization in DBM contributes to improve the face verification rate. The suggested method yields the success rate of matching faces about 95% s at equal error rate for the You Tube Video Faces database, and it is possible to achieve even better results for PaSC database on the other hand.
Full Text:
PDFCopyright (c) 2018 Edupedia Publications Pvt Ltd
This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.
All published Articles are Open Access at https://journals.pen2print.org/index.php/ijr/
Paper submission: ijr@pen2print.org