A Comparative Study of Face Recognition and Detection Mechanisms Through deep and Machine Learning and Handcrafted Features
TL;DR
A comparative study between handcrafted and automated feature extraction method has been provided, showing that the performance with HE is the best in the handcrafted approach.
Abstract
In this paper, a comparative study between handcrafted and automated feature extraction method has been provided. The handcrafted method has been based over local binary pattern (LBP) as feature extraction technique. The histogram equalization (HE), multi-scale retinex (MSR), and a difference of Gaussian (DOG) have been used as a preprocessing technique to improve the image quality. The results of the handcrafted approach have been shown that the performance with HE is the best. In the automated part, ALEXNET has been used as convolutional neural network (CNN) architecture. The standard gradient descent with momentum (SGDM) has been used as the optimizer, because the results were better when it has been used. The results of the automated part have been shown how the layers activation functions works. In the automated part, the training and test accuracy have been evaluated and compared between different databases. The accuracy has achieved up to 100% in Face94 and Face grimace databases as the best accuracy in the CNN approach.