A PCA/MDA scheme for hand posture recognition

Principal component analysis (PCA) and multiple discriminant analysis (MDA) have long been used for appearance-based hand posture recognition. In this paper, we propose a novel PCA/MDA scheme for hand posture recognition. The scheme is represented by two layers of nodes (classes). The first layer of nodes acts as a crude classification using PCA, and each input pattern is given a likelihood of being in the nodes of this layer. Then MDA is applied locally to the postures in each node of the first layer to give a precise classification of the postures. Each precise class is a node in the second layer. For training, unsupervised classification at the first layer can be obtained using expectation maximization (EM). For better training results, a feedback from each node in the second layer is introduced in the EM process. The experiments on a 100-sign vocabulary show a significant improvement from 57.0% to 63.5%, compared with the global MDA. If combined with a hidden Markov model (HMM) for movement modeling, about a 93.5% recognition rate is achieved for test data.