Fast convolution technique for non-separable oversampled lapped transforms
| Autor: | Kota Horiuchi, Shogo Muramatsu |
|---|---|
| Rok vydání: | 2016 |
| Předmět: |
Overlap–add method
Theoretical computer science Discrete-time Fourier transform Fast Fourier transform 02 engineering and technology 01 natural sciences Convolution 010309 optics Generative model Acceleration 0103 physical sciences 0202 electrical engineering electronic engineering information engineering 020201 artificial intelligence & image processing Convolution theorem General-purpose computing on graphics processing units Algorithm Mathematics |
| Zdroj: | APSIPA |
| Popis: | In this study, an acceleration technique for non-separable oversampled lapped transforms (NSOLT) is proposed. NSOLT is a non-separable transform which simultaneously satisfies the oversampled, overlapped, symmetric, tight-frame, real-valued and compact-supported property. From the non-separability, NSOLT is valid for sparsely representing edges and textures with oblique directions as a linear generative model. In this paper, a fast convolution technique for NSOLT is proposed by using fast Fourier transform (FFT) and frequency-domain multirate processing. The analysis and synthesis processing for 2-D and 3-D data is evaluated through GPGPU implementation, and the processing speed of the proposed structure is compared with the conventional lattice-based one to show the effectiveness. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|