Pixel Clustering for Face Recognition

This work proposes a theoretical framework for an unsupervised feature extraction called Pixel Clustering. The main idea is based on the clustering of the pixels in order to mitigate the multicollinearity issue and a new feature is extracted for each cluster of similar pixels. This allows to define feature extraction techniques by setting just three parts: (1) defining pixel vectors in the training set, each pixel vector is a representative for a pixel on every training image, (2) a clustering algorithm for the pixels vectors, (3) finally it is performed a linear combination of the pixel into a cluster, in order to create a single feature per cluster. The framework also makes it possible to create simpler, computationally cheaper and more general new implementations of well known feature extraction methods such as Waveletfaces. Two extraction methods are implemented and tested in three face datasets. Test results are compared to the traditional Eigenfaces and others state-of-art feature extraction methods for face recognition. The proposed method achieves up to 38% higher face recognition rate than Eigenfaces, if few classes are used for training the projections.