Graph-based adaptive and discriminative subspace learning for face image clustering

Current graph-based subspace clustering methods have achieved some results for the clustering of face images. The core of those methods lies in graph learning. However, they still have the following problems when learning the graph. Firstly, the graph learning processes of those methods do not consider the alignment of the images. It is well known that the obtained images of the same category may not be aligned due to different devices and shooting angles. The unaligned images used for graph learning directly affect the accuracy of the resulting graph. Hence, the graphs obtained by these methods are not accurate. We know that the inaccuracy of the learned graph will directly reduce the clustering performance of those methods. Secondly, they believe that important features, redundant features, and noise play the same contribution in the process of graph construction and feature representation. Redundant features and noise are not beneficial to graph reconstruction and feature representation, and even cause the learned graph to be inaccurate. Thirdly, the intrinsic structural correlation between samples is rarely considered for graph learning, which makes it difficult for the learned graph to reflect the structural correlation, then a good clustering performance cannot be obtained. To address those problems, this paper proposes a graph-based adaptive and discriminative subspace learning method (GADSL). In GADSL, image alignment is introduced and unified with subspace learning under the same graph learning framework which helps reduce the impact of different shooting equipment. Besides, GADSL can adaptively assign large weights to the important features and small weights to the unimportant features by introducing the weighting matrix. Moreover, in order to consider the correlation between samples, the structural consistency constraint is introduced into the subspace learning process so that the intra-class difference decreases and the inter-class difference increases. The experimental results show that GADSL used for face image clustering has achieved better clustering performance than many state-of-the-art methods.