تمييز الوجوه اعتمادا على التحويلات الكونتورية ومصنف الشبكة العصبية متعددة الطبقات == Face Identification Based On Contourlet Transform And Multi - Layer Neural Network Classifier
Author name:
علي فتاح حسون
Supervisor name:
ماهر خضير محمود
General topic:
Electrical, Electronic and Communications Engineering
Specific topic:
Electronics and Communications Engineering
Degree:
Master
University:
Mustansiriyah University - College Of Engineering - Department Of Electrical Engineering
Language:
English
University location:
Baghdad
First pages:
34T534 - p.pdf
Abstract:
Face recognition plays an important role in many applications like security, banking, access to buildings, and surveillance …. etc.Because of the face image can significantly change due to changing in some factors, such as lighting conditions, facial expression, and others factors, therefore face recognition is still a challenging mission. Useful properties of the Contourlet Transform (CT) are exploited in this thesis to investigate more discriminant features to enhance the face identification performance. In this thesis, a face identification system is suggested based on CT, and Multi - Layer Neural Network (MLNN) Classifier. The main reasons behind using the CT are : First, the CT supports progressive data compression/expansion, hence it is used for image compression. Second, the features in human face are not just horizontal or vertical. CT is utilized for feature extraction because it is a genuine 2D transform that can capture the edge information in all directions. After decomposing an image by CT, low and high frequency coefficients of CT are calculated in different scales and various angles. Most significant information of faces is contained in coefficients, which is important for face identification. The frequency coefficients are utilized as an input feature vector for a neural network classifier. Simple feed forward MLNN neural network is used to achieve the identification process. The network parameters are tuned to optimal values, in order to produce fair comparison between different types of feature vectors. To evaluate the algorithm performance, seven standard databases, and a proposed database are used. Some of them are of high variability, to examine the algorithm robustness against variability. In addition, the proposed algorithm is also evaluated using a generated database composed from five databases. Then the suggested method is compared with other classical feature - based methods such as, wavelet, Principle Component Analysis (PCA), and CT - PCA. The results indicate that the CT - based method has better identification (ID) rate, and is faster than the Wavelet - based and the PCA - based methods. This is due to the high sparsity of the CT and its efficient capability of compression. The ID rates obtained for a tested databases are Japanese Female Facial Expression (JAFFE) database (100%), face 94 database (99.34%), proposed database (98.55%), ORL database (96.5%), Senthil database (92%), Yale database (90%), FERET database (76.2 %), and FEI database (72.57 %), and comparing with other systems, an average identification rate 90.64% for the CT, 85.71% for the Wavelet, 53.92% for the CT - PCA, and 46.7% for the PCA for all tested databases.Moreover, the characteristics of low and high frequency sub bands a are studied with their effect on ID rates. The simulation results indicate that the best identification rates are obtained when using only the low frequency sub - band for all tested databases
