LGE-KSVD: Flexible Dictionary Learning for Optimized Sparse Representation Classification

Sparse representations have successfully been exploited for the development of highly accurate classifiers. Unfortunately, these classifiers are computationally intensive and subject to the adverse effects of coefficient contamination, where for example variations in pose may affect identity and expression recognition. We propose a technique, called LGE-KSVD, that addresses both problems and attains state-of-the-art results for face and gesture classification problems. Specifically, LGE-KSVD utilizes variants of Linear extension of Graph Embedding to optimize K-SVD, an iterative technique for small yet overcomplete dictionary learning. The dimensionality reduction matrix, sparse representation dictionary, sparse coefficients, and sparsity-based linear classifier are jointly learned through LGE-KSVD. The atom optimization process is redefined to have variable support using graph embedding techniques to produce a more flexible and elegant dictionary learning algorithm. Results are obtained for a wide variety of facial and activity recognition problems to demonstrate the robustness of the proposed method.