A fusion representation for face learning by low-rank constrain and high-frequency texture components

Abstract

• An effective technique for extracting the high-frequency texture component of image is designed. • A learning framework is proposed to reconcile the subspace learning procedure with texture information. • Experimental results show our approach can learn a better low-rank structure and achieve better clustering performance. Face image processing has always been an important field of artificial intelligence. In machine learning tasks, samples are vectorized into a data matrix. Hence, different faces lie in several respective low-rank subspaces , which can be effectively captured by Low-Rank Representation (LRR) learning. However, most technologies usually use the raw data directly but ignore the weight among different information of the image itself. Following the intuition that human beings distinguish different people through the lineaments and wrinkles of faces which are indeed the texture components of face images, it is believed that representation learning will benefit from introducing the texture components of samples into the information fusion. Inspired by Multi-view learning models which can reconcile the knowledge gained from different views, a fusion framework is proposed in this letter, named High-Frequency texture components Low-Rank Representation (HFLRR). In HFLRR, the high-frequency texture components of samples will be extracted by Fast Fourier Transformation and Butterworth filtering firstly. Subsequently, the texture information will be introduced into the subspace learning procedure. An effective optimization solution is presented for this framework and the experimental results show that the proposed algorithm outperforms the state-of-art methods on several real-world face datasets.