Face recognition (FR) is considered to be a typical machine learning problem. Among all FR algorithms, popular models include classical linear models such as eigen face, nonlinear models such as manifold learning, and sparse representation-based classification (SRC). Different machine learning techniques have been combined and tested on this problem over the past two decades. Structured sparse representation (SSR) was more recently developed to address the ambiguity from equally good linear representations of the input in SRC; the better approach minimizes the number of bases in the representation. Another well-known machine learning component is dimension reduction (DR), which has been proven to be able to reduce complexity and to increase accuracy in discriminative models.

The paper combines SSR and DR and claims up to two percent improvement over similar approaches based on three FR databases. Besides this new practice, equation 14 could be one main contribution; however, although section 5 has a sentence explaining why it is better than the original formula, it needs further proof in my opinion. Another worth-noting achievement is that the new approach works well with small training samples (three per subject). Thus, it is worth a try when you don’t in general have a large dataset for convolutional neural networks.