Factorizing square kernels into two no-activation 1D kernels cuts parameters and FLOPs by 32 to 62 percent in small super-resolution networks, with PSNR wins more frequent at 4x scale but SSIM consistently dropping.
In: Pro- ceedings of the British Machine Vision Conference
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Enhancing Frequency for Single Image Super-Resolution with Learnable Separable Kernels
Factorizing square kernels into two no-activation 1D kernels cuts parameters and FLOPs by 32 to 62 percent in small super-resolution networks, with PSNR wins more frequent at 4x scale but SSIM consistently dropping.