Meta-learning based relation and representation learning networks for single-image deraining

doi:10.1016/j.patcog.2021.108124

Meta-learning based relation and representation learning networks for single-image deraining

2021-12-01

14

PRE

AI

DOI：10.1016/j.patcog.2021.108124

Original

Original request for help

Save

Abstract

En 中文

Single-image deraining is a kind of computer vision task that aims to restore the image that be degraded by rain streaks, which motivates existing methods to either directly translate the rainy image to its clean one, or indirectly learn the rain residual based on the prior information. However, both methodologies harm the generalization ability due to the limited diversity of the training samples, comparing with the endless varieties of the real-world rainy images. Such fact inspires us to take the merit of meta-learning and propose a meta-learning based representation learning network to learn the transferable embed dings of the rainy/clean images, while their discrepancies are characterized by the relation vector, which is generated by the subsequent meta-learning based relation learning network. These networks are leveraged into the meta-learning based deraining network (MLDN) to enhance the generalization ability by removing the latent relation vector from the transferable embedding of the rainy image and generate high-quality deraining result. Superior performance is achieved by MLDN, which has averaged 4 % better than the state-of-the-arts. (c) 2021 Elsevier Ltd. All rights reserved.

Keywords:

Meta-learning

Relation network

Single-image deraining

Representation learning

Meta-learning based relation and representation learning networks for single-image deraining

Abstract

AI Summary

Journal

Organization