基于增强型金字塔及图像超分辨率去雾网络  

Based on image super-resolution and boosted feature pyramid dehazing network

作　　者：王科平[1] 肖梦临 WANG Keping;XIAO Menglin(School of Electrical Engineering and Automation,Henan Polytechnic University,Jiaozuo 454003,China)

机构地区：[1]河南理工大学电气工程与自动化学院,河南焦作454003

出　　处：《兵器装备工程学报》2023年第11期299-307,共9页Journal of Ordnance Equipment Engineering

摘　　要：使用单一卷积神经网络去雾算法容易存在对比度偏低、细节信息丢失和去雾不完全等缺陷。为了解决上述问题,提出了一种增强型金字塔模型和图像超分辨率并联去雾网络结构。增强机制作用于特征金字塔图像重建过程,用以提升去雾图像信噪比。通道注意力将编码器提取的特征信息映射到解码器,赋予每个通道不同权重,以此提高去雾效率。超分辨率网络补充更多高频特征细节,提升去雾图像的清晰度。实验证明,增强型金字塔及超分辨率网络具有较强的去雾能力,性能优于其他方法,有效抑制单一的卷积神经网络输出图像分辨率下降问题。Using a single convolutional neural network dehazing algorithm is prone to low contrast,loss of detail information and incomplete dehazing.In order to solve the above problems,a boosted pyramid model and image super-resolution parallel demisting network structure are proposed.The boost algorithm acts on the feature pyramid image reconstruction process to improve the signal-to-noise ratio of the defogging image.Channel attention maps the feature information extracted by the encoder to the decoder,giving each channel different weights,so as to improve the efficiency of dehazing.The super-resolution network adds more high-frequency feature details to improve the clarity of the dehazing image.Experiments show that the boosted pyramid and super-resolution network have strong dehazing ability,and their performance is better than other methods,which can effectively suppress the degradation of the output image resolution of a single convolutional neural network.

关 键 词：图像去雾 增强机制 超分辨率 特征金字塔 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]