Transposition-based Architecture of 2-D Non-separable Oversampled Lapped Transforms

This paper proposes an embedded implementation architecture of 2-D Non-separable Oversampled Lapped Transform (NSOLT). NSOLT yields a frame for image representation which satisfies the symmetry and overlapping property. Since the redundancy can be set to an arbitrary rational value, high performance image processing can be realized while saving memory consumption. This work mainly concerns the implementation issue. In general, non-separable systems require more computational resources than separable ones. However, NSOLT can be implemented by a separable manner thanks to the order factorable property. Based on the fact, a transposition-based architecture is proposed, and the significance is verified through the estimation of required hardware resources and its processing speed. The hardware module performances are evaluated for Xilinx XC7Z020-1CLG484 (Zynq).