Improved HMM by Deep Learning for Ear Classification

Abstract

 Ear recognition is one of the most relevant  applications of image analysis. It’s a true challenge to  build an automated system which exceeds human ability  to recognize ears. Humans do not identify the ears  ordinarily, so we are not skilled when we must deal with  a large number of unknown ears. The modern  computers, with an almost limitless memory and  computational speed, should overcome humans'  limitations. This work uses Ear classification problem  to improve Deviance Information Criterion- Structural  Hidden Markov Model (DIC-SHMM) by Convolutional  Neural Network (CNN). HMM is a strong model for  large size of features. While the CNN, as a most  important technology for deep learning, used for image  classification, to recognize persons by their ears as a one  of unique biometric physiological characteristics. Three  systems will be used to classify ear images, deep  learning for the original image directly, deep learning  for eigenvector as Principle Components Analysis  (PCA) of the original image to compare them with  proposed combining convolution layers of CNN with  improved HMM for the original Image. The proposed  system shows the best correction rate as 97.5%.