Recently, the great development of deep learning has prompted many neural networks for single image dehazing to occur. How-ever, due to the ill-posed nature of haze, an excellent charac-teristics representation capacity is still challenging. In this paper, we propose a lightweight yet effective senet-residual (SE-Res) multiscale end-to-end neural network named SRM-Net. Inspired by the remarkable performance of residual networks, we intro-duce a SE-Res structure which is an improved residual framework with an embedded SE unit to obtain feature maps. These maps pass through a multiscale mapping layer which can aggregate characteristics in different receptive fields. Notably, the utilization of all point-wise convolutions in the SRM-Net leads to fewer parameters for training, and the reuse of feature maps makes it more lightweight. Through extensive numerical experiments on three datasets including real hazy images, synthetic indoor and outdoor hazy images, the proposed SRM-Net achieves superior performances on subjective visual results and objective evaluation metrics compared to the state-of-the-art methods.
Paper
Full text
SRM-Net: An Effective End-to-end Neural Network for Single Image Dehazing
Semantic Scholar · Computer Science · 2019
Abstract
Recently, the great development of deep learning has prompted many neural networks for single image dehazing to occur. How-ever, due to the ill-posed nature of haze, an excellent charac-teristics representation capacity is still challenging. In this paper, we propose a lightweight yet effective senet-residual (SE-Res) multiscale end-to-end neural network named SRM-Net. Inspired by the remarkable performance of residual networks, we intro-duce a SE-Res structure which is an improved residual framework with an embedded SE unit to obtain feature maps. These maps pass through a multiscale mapping layer which can aggregate characteristics in different receptive fields. Notably, the utilization of all point-wise convolutions in the SRM-Net leads to fewer parameters for training, and the reuse of feature maps makes it more lightweight. Through extensive numerical experiments on three datasets including real hazy images, synthetic indoor and outdoor hazy images, the proposed SRM-Net achieves superior performances on subjective visual results and objective evaluation metrics compared to the state-of-the-art methods.