Applications of Artificial Neural Networks to Facial Image Processing

During the past 20 years, artificial neural networks was successfully applied for solving signal processing problems. Researchers proposed many different models of artificial neural networks. A challenge is to identify the most appropriate neural network model which can work reliably for solving realistic problem. This chapter provides some basic neural network model and efficiently applying these models in facial image processing problem. In detail, three techniques : a hybrid model of combining AdaBoost and Artificial Neural Network (AANN) to detect human faces, a local texture model based on Multi Layer Perceptron (MLP) for face alignment and a model which combines many Neural Networks applied for facial expression classification are present. This case study demonstrates how to solve face recognition in the neural network paradigm. Each of these techniques is introducted as follows: Technique 1 an approach to combine adaBoost and artificial neural network for detecting human faces: The human face image recognition is one of the prominent problems at present. Recognizing human faces correctly will aid some fields such as national defense and person verification. One of the most vital processing of recognizing face images is to detect human faces in the images. Some approaches have been used to detect human faces. However, they still have some limitations. In the research, some popular methods, AdaBoost, Artificial Neural Network (ANN) were considered for detecting human faces. Then, a hybrid model of combining AdaBoost and Artificial Neural Network was applied to solve the process efficiently. The system which was build from the hybrid model has been conducted on database CalTech. The recognition correctness is more than 96%. It shows the feasibility of the proposed model. Technique 2 local texture classifiers based on multi layer perceptron for face alignment: Local texture models for face alignment have been proposed by many different authors. One of popular models is Principle Component Analysis (PCA) local texture model in Active Shape Model (ASM). The method uses local 1-D profile texture model to search for a new position for every label point. However, it is not sufficient to distinguish feature points from their neighbours; i.e., the ASM algorithm often faces local minima problem. In the research, a new local texture model based on Multi Layer Perceptron (MLP) was proposed. The model is trained from large databases. The classifier of the model significantly improves accuracy and robustness of local searching on faces with expression variation and ambiguous contours. Achieved experimental results on CalTech database show its practicality.