Bilinear Factorization via Recursive Sample Factoring for Low-Rank Hyperspectral Image Recovery

Abstract

Low-rank hyperspectral image recovery (LRHSIR) is a very challenging task in various computer vision applications for its inherent complexity. Hyperspectral image (HSI) contains much more information than a regular image due to significant number of spectra bands and the spectral information can be considered as multiview. In this paper, a method of bilinear factorization via recursive sample factoring (BF-RSF) is proposed. Different from traditional low rank models with each data point being treated equally, the importance of each data point is measured by the sample factoring that imposes a penalty on each sample in our BF-RSF model. The sample factoring is a cosine similarity metric learnt from the angle between each data point and the principal component of the low-rank matrix in the feature space. That is, the closer a data point to the principal component vector, the more likely it is a clean data point. By imposing the sample factoring onto the training dataset, the outliers or noise will be detected and their effect will be suppressed. Therefore, a better low-rank structure of clean data can be obtained especially in a heavy noisy scenario, with the effect of noisy data points in modeling being suppressed. Extensive experimental results on SalinasA, demonstrate that BF-RSF outperforms state-of-the-art low-rank matrix recovery methods in image clustering tasks with various levels of corruptions.