Ensemble single image deraining network via progressive structural boosting constraints

Abstract

Image deraining is an extensively researched topic in low-level computer vision community. During the past, sufficient works have been proposed to address this problem. Though great improvements these methods have achieved, no derain network can confidently declare that it can solve rain removal problem perfectly. Single complex model may lead to overfitting, while simple model is too weak to achieve clear result. Therefore, in this paper, inspired from classic boosting idea, we have proposed an effective ensemble derain framework to aggregate multiple simple weak drain models to obtain a strong derain model. Cascade structural weighting-map are computed for adaptively emphasizing the quality of local derain regions. Struct-absolution losses are proposed to account for pixel-wise and local region-wise differences, and to facilitate embedding boosting idea into network training. The comprehensive experiments on public derain datasets and high-level vision tasks validate that our proposed model which just utilizes three generally weak derain subnets can achieve much better performance than compared state-of-the-art methods. Our ensemble framework has enough capacity that any state-of-the-art DL-based models can be taken as sub-modules to solve rain removal of multiple types within a single framework.