A Tensor Singular Value Decomposition Accelerator for Hyperspectral Imaging Applications

This paper presents a hardware accelerator for tensor singular value decomposition (t-SVD) for hyperspectral imaging (HSI) applications. A parallel hardware architecture was developed to simultaneously process multiple slices of a tensor. Additionally, blockwise singular value decomposition (SVD) was used to reduce the number of hardware resources. Implemented on an AMD Virtex UltraScale+ VCU118 FPGA platform operating at 200 MHz, the accelerator can perform t-SVD on a $256 \times 256 \times 256$ tensor within 0.0564 ms. Two HSI application datasets with sizes of $(145 \times 145 \times 224)$ and $(240 \times 143 \times 191)$, respectively, were used to validate the accelerator, demonstrating an average $65 \times$ speedup compared with the software implementation on an Intel i7-7700K CPU operating at 4.2 GHz.