COURSENAME
Face Recognition using HMM with SVD Coefficients
Summary
I propose an approach for face recognition system based on Singular Value Decomposition (SVD) method to extract face features and Hidden Markov Model (HMM) for classification. To test the system three datasets will be used - the color FERET, SC face – Surveillance cameras face database, and CMU Multi-PIE. The system will be implemented and tested using Matlab.
Problem Statement
Manifold modern application require an accurate and robust face recognition functionality. In the presence of noise, occlusions, and dynamic change of background scene face recognition becomes a challenging problem. Majority of implementations use Principal Component Analysis (PCA) method which is not very effective approach in a sense it does not provide invariance over scale and lighting conditions change. Therefore, to overcome limitation of this approach I propose a technique based on using seven states of Hidden Markov Model (HMM) as classifier and Singular Value Decomposition (SVD) to extract face features.
Related Works
Cao and Yang (109-112) in their work introduced SVD technique to improve face recognition and proved its effectiveness. In 2008 Miar-Naimi and Davari (46-57) proposed fast and efficient 7 state HMM with SVD coefficients face recognition approach and showed has very accurate recognition rate. Similar works are performed by Dinkova et al (144-149).
Methodology
First, data should be preprocessed and in order to increase the speed of processing images will be cropped to the size of approximately 256x256 or 64x64 and turned to grayscale representation. Some preprocessing techniques as smoothing, information reduction, and noise-removal filters such as order-statistic filter will be applied to improve the performance of the system.
Second, useful features of images should be extracted. For this, 2D images will be transformed into 1D sequence as HMM requires one-dimensional observation sequence (Dinkova, Georgieva and Milanova 144-149). Image will be divided into T blocks which are overlapping and have width W equal to the original and height L different from original one. T is found according to
T= H-LL-P+1
where P is overlapping size and equal to L-1. It is stated that more overlapping between consecutive blocks notably improves the performance but is more computationally expensive (Miar-Naimi and Davari 46-57). Next step is to apply SVD to each of these block. as:
Xm×n=Um×m*∑m×n*Vn×nT
where U and V are mutually orthogonal matrices and ∑ is a diagonal matrix with singular values. The main property of SVD in image recognition is its stability as singular values contain algebraic image properties and contain information about noise level, energy, rank of the matrix, etc.
Obtained SVD coefficients are continuous values there is an infinite number of possible observation vectors which cannot be directly applied to HMM and therefore, should be quantized (Miar-Naimi & Davari, 2008). If X=(x1,x2, , xn ) is vector with continuous elements and in order to quantize it into Di levels, first, find difference between successive values:
∆i=xi max-xi minDi
and then quantize:
xi quant=xi-xi min∆i
Then Markov model is applied on 1D observational sequence. For this seven hidden states of Markov model are used which are: hair, forehead, eyebrows, eyes, nose, mouth, and chin (Figure 1).
Figure 1: HMM model with 7 states for a face image (Miar-Naimi and Davari 46-57).
Design of experiment
Preprocessing techniques, feature extraction and quantization, and Markov model will be implemented using Matlab. Data will be divided into training and testing sets in a proportion of 80% to 20% respectively. Block and feature extraction, and quantization will be applied on test set and classification accuracy will be calculated. To obtain best model image size, filtering, and quantization parameters will be tuned and different features will be used.
Data/ Test cases
For this project three different datasets will be used: the color FERET, SC face – Surveillance cameras face database, and CMU Multi-PIE. FERET database comprises 14, 126 images belonging to 1564 sets and includes 1199 subjects and 365 duplicate sets (the image for the same individual was taken in two different days over year). This data was obtained in 15 sessions during the period from August 1993 to July 1996. SCface contains 4160 static images in visible and infrared spectrum of 130 individuals taken in uncontrolled indoor environment with five surveillance cameras of varying qualities. Thus, the images were collected in close to real-world conditions over a five-day period. Moreover, using 130 individuals for the dataset significantly lowers the change of recognizing by chance (1/130 = 0.8%). Images are in lossless 24-bit color JPEG format and have cropped size of 1600x1200 pixels. Finally, Multi-PIE dataset consists of more than 750,000 images of 337 persons obtained during four sessions over the period of five months. Each image has size of 3072x2048 pixels. 15 view points and 19 illumination conditions were set up to taken a range of facial expressions such as neutral, smile, disgust, and scream.
Work Cited
Cao, Danyang and Bingru Yang. "An Improved Face Recognition Algorithm Based On SVD".
Computer and Automation Engineering (ICCAE), 2010 The 2nd International Conference on 3 (2010): 109-112. Print.
Dinkova, Petya, Petia Georgieva, and Mariofanna Milanova. "Face Recognition Using Singular
Value Decomposition and Hidden Markov Models". Mathematical Methods in Science and Mechanics 144-149. Print.
Grgic, Mislav et al. Scface - Surveillance Cameras Face Database. 2011. Database. Zagreb,
Croatia.
Miar-Naimi, H. and P. Davari. "A New Fast and Efficient HMM-Based Face Recognition
System Using A 7-State HMM Along with SVD Coefficients". Iranian Journal of Electrical & Electronic Engineering 4.1 (2008): 46-57. Print.
The CMU Multi-PIE Face Database. 2010. Database.
The Color FERET Database. 2011. Database. U.S.