الفهرس 
IntroductionOnly 14 pages are availabe for public view
 |  | from | 51 |  |  |
| | | from | 51 | | |
Abstract
Face and image recognition systems using modem image processing techniques are proposed in this thesis. In the image recognition system we proposed new invariant features that are used in image recognition. In order to prove the activity of these features we used it to perform the face or non-face image classification. The face recognition system includes two-basic processes. The first one is the face detennination process, and the second one is the face recognition.
Two sets of pose invariant features that depend on Fourier and Wavelet Transforms (FWT) are proposed and implemented. The translation invariance is achieved by translating the origin of the Cartesian coordinate system to the center of mass of the pattern. The scale invariance is obtained by transformation the pattern of the image into the polar coordinate. Finally the rotation invariance is achieved using the one-dimensional Fourier transform of the image in polar coordinates. We used Wavelet transform because its coefficients represent the pattern at different resolution levels. Then we propose two-sets of feature by thresholding the Wavelet coefficients, which can be used in pattern recognition. We apply our feature extraction methods on different images such as a tree, a moon, a flower, and Saturn.
One of the applications of our image recognition approaches, which is the face or non-face image classification, is proposed and implemented using neural network. When we fad the features that are extracted by the second approach to three layers Fast Back Propagation Neural Network (FBPNN), we obtained 1000/0 classification rate. We used in this step the (Olivetti Research Laboratory) ORL face database with a large number of different images.
We proposed a new method for full-face determination. Based on Principal Component Analysis (PCA) we used the first eigenvector to extract the face part in a color and grayscale image. Then we proposed a new method depending on Hotelling transform, which reposes the face image to an interesting pose to cancel the error that produces ITom the rotation and translation factor. The experiment showed that the facial part could be successfully extracted using PCA face determination ITom the color and grayscale image and aided us to overcome the disadvantage of the feature
extraction methods. Also experimental results by showed successively that Hotelling Transform could cancel the effect of rotation factor.
A recognition method of human face using statistical analysis feature extraction and a neural network algorithm is proposed. The face recognition process includes the preprocessing step, feature extraction step and recognition step. In the preprocessing step we detected the edges of the face image using the Sobel algorithm. Then we propose a new method to transform the two-dimensional black and white image to one-dimensional vector. That helps us to reduce the dimension of processing by data by III 0 % reduction factor. In the feature extraction step we extract seven features based on the statistical analysis, whish are the mean, median. mean absolute deviation, standard deviation, mod of distribution, third moment of distribution (skewness), and the peak period of Fourier transform. In the recognition step we use the two layers Fast Back Propagation Neural Network (FBPNN). Experimental results gave 75% recognition rate and showed that the recognition rate become less sensitive to facial expression (open/closed eyes, smiling! non-smiling), facial details (glasses! no glasses), and the large values of tolerance or rotation.
New features based on texture analysis of facial skin that are used in face recognition are proposed. We analyzed the region of face that does not contain hair or edges like the region offorehead or the region of the skin over the cheek (zygomatic) bone. To determine this region we propose a new point detection method. In order to enhance the activity of statistical features in face recognition we combined these features with our texture features, (Energy, Entropy, and Homogeneity) whose extracted depends on the concurrence matrix of the predetermined face region. Experimental results showed that the recognition rate was enhanced to &5%.
The introduction, literature survey, and the objective of the thc:sis are presented in Chapter One.
Chapter Two discusses the most important techniques that are used in face recognition.
Chapter Three explains the principle of artificial neural network and its application in image processing.
Chapter Four presents the wavelet transforms in some details and two proposed approaches for invariant image classification based on Fourier Wavelet Transform of the image in polar coordinates. Also it explains the implementation of these features in face or non-face image classification.