Dformer: Learning Efficient Image Restoration with Perceptual Guidance

Nodirkhuja Khudjaev, Roman Tsoy, S M A Sharif, Azamat Myrzabekov, Seongwan Kim, Jaeho Lee*
Opt-AI Inc.
LG Sciencepark, Seoul, South Korea 

Abstract

Image restoration tasks incorporate widespread realworld application. Apart from its significant practicability, generic deep image restoration methods still fail to handle complex tasks, like shadow removal, low-light enhancement, etc. This paper addresses the limitations of existing image restoration methods by introducing a novel deep architecture. The proposed model incorporates illumination mapping inspired by the Retinex theory within a double encoder-decoder network. Additionally, it utilizes a multi-head cross-attention mechanism to correlate input and reconstructed images to generate plausible and refined images. The proposed model employs a perceptual optimization strategy to tackle intricate restoration tasks effectively. It surpasses state-of-the-art methods in demanding tasks such as shadow removal, low-light image enhancement, and blind compress image enhancement, all while utilizing fewer trainable parameters. Our method is selected among the top solutions in the New Trends in Image Restoration and Enhancement’24 (NTIRE) challenge
for shadow removal, securing a top position without resorting to score-boosting techniques such as ensembling.

*Corresponding author