Neural Architecture Search for Image Dehazing

Abstract

Manual design of deep networks require numerous trials and parameter tuning, resulting in inefficient utilization of time, energy, and resources. In this article, we present a neural architecture search (NAS) algorithm—AutoDehaze, to automatically discover effective neural network for single image dehazing. The proposed AutoDehaze algorithm is built on the gradient-based search strategy and hierarchical network-level optimization. We construct a set of search space layouts to reduce memory consumption, avoid the NAS collapse issue, and considerably accelerate the search speed. We propose four search spaces <inline-formula><tex-math notation="LaTeX">$\text{AutoDehaze}_{B}$</tex-math></inline-formula>, <inline-formula><tex-math notation="LaTeX">$\text{AutoDehaze}_{U1}$</tex-math></inline-formula>, <inline-formula><tex-math notation="LaTeX">$\text{AutoDehaze}_{U2}$</tex-math></inline-formula>, and <inline-formula><tex-math notation="LaTeX">$\text{AutoDehaze}_{L}$</tex-math></inline-formula>, which are inspired by the boat-shaped, U-shaped, and lateral connection-based designs. To the best of authors knowledge, this is a first attempt to present an NAS method for dehazing with a variety of network search strategies. We conduct a comprehensive set of experiments on Reside-Standard (SOTS), Reside-<inline-formula><tex-math notation="LaTeX">$\beta$</tex-math></inline-formula> (SOTS) and Reside-<inline-formula><tex-math notation="LaTeX">$\beta$</tex-math></inline-formula> (HSTS), D-Hazy, and HazeRD datasets. The architectures discovered by the proposed AutoDehaze quantitatively and qualitatively outperform the existing state-of-the-art approaches. The experiments also show that our models have considerably fewer parameters and runs at a faster inference speed in both CPU and GPU devices.