A Bidirectional Deep Manifold Model for Image Representation
Yang Shangming, Liu Yongguo, Li Qiaoqing · Frontiers in artificial intelligence and applications · 2018
Deep learning model has been extensively applied to the area of data classification and clustering in recently years because of its high similarity to human neurons in computational performance. In this paper, we propose a new deep model, multi-layer manifold based nonnegative matrix factorization with partial neural connections for image representation. In this model, a bidirectional multi-layer NMF decomposition for both basis and encoding vectors are conducted to capture structures in high dimensional data space. The connections of neurons in the learning are constrained in a neighborhood so that the similarities of elements in the same cluster can be learned. Test results on two different image datasets confirm that the proposed method can learn very good performance for image clustering tasks.