Neural Architecture Search for Deep Image Prior
This work addresses image restoration for photography and art by incrementally improving DIP through automated architecture optimization.
The authors tackled the problem of enhancing unsupervised image restoration tasks (de-noising, in-painting, super-resolution) using Deep Image Prior (DIP) by applying neural architecture search (NAS) to optimize network structures, resulting in improved visual quality with convergence typically in 10-20 generations.
We present a neural architecture search (NAS) technique to enhance the performance of unsupervised image de-noising, in-painting and super-resolution 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.