A New Approach in Combining Fisher’s Linear Discriminant and Neural Network for Face Detection

This paper presents a new method for combining Fisher’s Linear Discriminant (FLD) and Multi Layer Perceptron (MLP) for face detection. The input patterns are first clustered into 10 face and 10 non-face clusters using K-means algorithm. Then, the FLD coefficients are calculated to obtain the optimal projection of face and non-face images. For each 19×19 pixel window, only 19 FLD coefficients are selected and presented to a MLP classifier. The salient point of our approach, comparing with similar works, is the utilization of K-means, FLD and a simple neural network without need of preliminary transformations, such as Principle Component Analysis (PCA). The proposed method has been successfully tested on a data set completely different from the training data set, containing 970 face and 11547 non-faces. The results of experimentations exhibit an error rate of 1.34% on faces and 2.03% on non-faces, i.e. an average error rate of 1.98%, a very interesting result considering the small number of FLD coefficients and the simple structure of network, which make it an appropriate choice for real-time applications. Key-words: Face detection, Neural Networks, Fisher Linear Discriminant, K-means, Face recognition