Locality-constrained discriminative learning and coding

This paper explores the enhancement by locality constraint to both learning and coding schemes, more specifically, discriminative low-rank dictionary learning and auto-encoder. Previous Fisher discriminative based dictionary learning has led to interesting results by learning more discerning sub-dictionaries. Also, the low-rank regularization term has been introduced to take advantage of the global structure of the data. However, such methods fail to consider data's intrinsic manifold structure. To this end, first, we apply locality constraint on dictionary learning to explore whether the identification capability will be enhanced or not by using the geometric structure information. Moreover, inspired by the recent advances from auto-encoders for learning compact feature spaces, we propose a locality-constrained collaborative auto-encoder (LCAE) for feature extraction. The improvement from applying locality to dictionary learning and auto-encoder is evaluated on several datasets. Experimental results have demonstrated the effectiveness of locality information compared with state-of-the-art methods.