Neural architecture search for deep image prior. (August 2021)

• Record Type: Journal Article
• Authors: Ho, Kary; Gilbert, Andrew; Jin, Hailin; Collomosse, John
• Is Part Of: Computers & graphics. Volume 98(2021)
• Page Start: 188
• Page End: 196
• Publication Date: 2021-08
• Subjects: Machine learning -- Network architecture search -- Super-resolution -- Inpainting; Computer graphics -- Periodicals; 006.6
• DOI: 10.1016/j.cag.2021.05.013
• Languages: English
• ISSNs: 0097-8493

Abstract

Highlights: Representation and method for evolutionary neural architecture search of encoder-decoder architectures for Deep Image prior. Leveraging a state-of-the-art perceptual metric to guide the optimization. State of the art DIP results for inpainting, denoising, up-scaling, beating the hand-optimized DIP architectures proposed. Demonstrated the content-style dependency of DIP architectures. Graphical abstract: Abstract: We present a neural architecture search (NAS) technique to enhance image denoising, inpainting, and super-resolution tasks under the recently proposed Deep Image Prior (DIP). We show that evolutionary search can automatically optimize the encoder-decoder (E-D) structure and meta-parameters of the DIP network, which serves as a content-specific prior to regularize these single image restoration tasks. Our binary representation encodes the design space for an asymmetric E-D network that typically converges to yield a content-specific DIP within 10--20 generations using a population size of 500. The optimized architectures consistently improve upon the visual quality of classical DIP for a diverse range of photographic and artistic content.