Progressive Deep Non-Negative Matrix Factorization Architecture with Graph Convolution-based Basis Image Reorganization

Abstract

Deep non-negative matrix factorization is committed to using multi-layer structure to extract underlying parts-based representation. However, the basis images obtained by continuous depth factorization is too sparse, resulting in too fragmented parts reflected by the basis image. This makes the number of factorization layers limited and the underlying local feature representation is inaccurate. Therefore, we propose a novel progressive deep non-negative matrix factorization (PDNMF) architecture that adds a basis image reconstruction step to the successive basis image factorization steps. This helps the basis image in depth factorization to maintain better robustness of feature representation. In the reconstruction step, the attribute similarity graph (ASG) is constructed to describe the semantic expression ability of each basis image. With the help of the ASG, the basis image enhances its own semantic integrity through graph convolution without drastically destroying its representation. The evaluation in image recognition shows that the recognition accuracy of the proposed PDNMF improves with the increase of layers. Our method outperforms the state-of-the-art deep factorization methods in image recognition.