Efficient approximate checkerboard K-SVD for resource constrained embedded systems

K-Singular Value Decomposition (K-SVD) is one of the key techniques in many signal processing and machine learning applications, such as image denoising, principal component analysis and dictionary learning. It is challenge to implement real-time K-SVD on resource constrained embedded systems, due to its exponentially increased computational complexity for large scale image or data dimensions. Approximate computing, as a trending paradigm in saving size, weight, and power (SWaP) in computation, trades the accuracy with complexity, so as to enable energy efficiency in modern computing systems. In this work, an approximate K-SVD based Dictionary Learning is studied primarily with sub-divided data blocks using reduced precision. The scalability of checkerboard K-SVD performance and arithmetic approximation are quantified with reconstructed image quality and computational complexity. It enables the efficient implementation by showing up to 16x estimated speed up and 98% reduction in memory foot-print using checkerboard division, and averagely 37% memory foot-print savings with approximate computing.