RepECN: Making ConvNets Better Again for Efficient Image Super-Resolution

Qiangpu Chen, Jinghui Qin, Wushao Wen, Antonio Fernández-Caballero, Byung-Gyu Kim · 2023

Read RepECN: Making ConvNets Better Again for Efficient Image Super-Resolution by Qiangpu Chen, Jinghui Qin, Wushao Wen, Antonio Fernández-Caballero, Byung-Gyu Kim (2023) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/pmc-10708766.txt.

Description

The authors (2023) present RepECN, a convolutional network that enlarges images quickly without giving up quality. Transformer models currently lead image super-resolution but are slow and bulky. Classic convolutional models run cheaply yet score lower. RepECN borrows the transformer's stage layout while dropping its costly attention step. Larger kernels give each block a wide view instead. A training trick called structural re-parameterization lets the blocks simplify at test time.

On standard test sets, the model matched or beat the best transformer system while running two and a half to five times faster. Its upscaling branch mixes neighbor interpolation with pixel attention to save parameters. A second branch feeds the network the high-frequency detail it must learn. The authors frame the result as proof that careful convolution design still competes. For phones and other tight hardware, that speed gap can decide whether super-resolution ships at all.

Catalog record

CREATORQiangpu Chen, Jinghui Qin, Wushao Wen, Antonio Fernández-Caballero, Byung-Gyu Kim
YEAR2023
COLLECTIONPAPERS
FORMATTXT
FILE SIZE42 KB
VERSION1.0
LICENSECC BY 4.0
RIGHTSCC BY
SOURCEEurope PMC

Source: https://europepmc.org/articles/PMC10708766

Browse the full library · About the Tower of Babel