Face detection with clustering, lda and NN
Hiroyuki Kobayashi, Qiangfu Zhao · 2007
In this paper, we study neural network (NN) based face detection. The main purpose is to reduce the complexity of the NN detector and thus speedup training and detection through linear dimensionality reduction. Neither principle component analysis (PCA) nor linear discriminant analysis (LDA) is good for this purpose. PCA often reduces descriptive and discriminative information together. On the other hand, LDA maps all data into a (C - 1)-dimensional feature space, where C = 2 for face detection. Since face detection is highly non-linear, classification in 1-dimensional space is clearly not enough. In this paper, we propose a new method for face detection. In this method, the problem is first changed to a multi-class problem using clustering. A modified LDA (m-LDA) is then proposed to extract useful features. The NN is used to make the final decision. Here, m-LDA is used to minimize the within-cluster variance between all "face" clusters; and maximize the between-cluster variance between all "face" clusters and all "non-face" data points. The feature space so obtained has a dimensionality less than that of the original problem. To validate the proposed method, we conducted several experiments with four methods, namely the proposed method, NN, PCA+NN, and LDA+NN. Results show that the proposed method can provide lower false positive and false negative errors for test images.