Sparse autoencoder based spatial pyramid facial feature learning

The spatial pyramid feature learning methods, such as Spatial Pyramid Matching (SPM) and Sparse Coding based Spatial Pyramid Matching (ScSPM), have achieved significant performance in image categorization. While most of these methods are still based on manual-design features, such as SIFT, HOG and LBP, which limits the representation of data. In this paper, we propose a novel Sparse Autoencoder based Spatial Pyramid Matching (SaSPM) method, which exploits unsupervised sparse autoencoder network infeatures learning and then builds a spatial pyramid structure. There are three main contributions in SaSP-M: Firstly, SaSPM is a learning method directly learning features from original data. Secondly, SaSPM is a full feedforward method in feature extraction, which is more efficient for on-line systems comparing with ScSPM method. Thirdly, we design patch-shared and patch-specific SaSP-M models to learn different local features separatively on well-aligned face images. It is proven that SaSPM outperforms the original spatial pyramid features in varieties of challenging data sets.