基于多重迁移注意力的增量式图像去雾算法  

Incremental Image Dehazing Algorithm Based on Multiple Transfer Attention

作　　者：韦金阳 王科平[2] 杨艺[2] 费树岷[3] Wei Jinyang;Wang Keping;Yang Yi;Fei Shumin(Zhengzhou Hengda Intelligent Control Technology Company Limited,Zhengzhou 450000,China;School of Electrical Engineering and Automation,Henan Polytechnic University,Jiaozuo 454003,China;College of Automation,Southeast University,Nanjing 210096,China)

机构地区：[1]郑州恒达智控科技股份有限公司,河南郑州450000 [2]河南理工大学电气工程与自动化学院,河南焦作454003 [3]东南大学自动化学院,江苏南京210096

出　　处：《系统仿真学报》2024年第4期969-980,共12页Journal of System Simulation

基　　金：河南省科技攻关(232102210040)。

摘　　要：为提高深度神经网络去雾算法对增补数据集的处理能力,并使网络差异化处理重要程度不同的图像特征以提高网络去雾能力,提出一种基于多重迁移注意力的增量式去雾算法。通过自编码器形式的教师注意力生成网络提取标签和雾霾的多重注意力,作为特征迁移媒介网络的标签约束网络训练,形成与教师注意力尽可能相近的迁移媒介注意力,并将其融入学生去雾网络的特征中,提高学生去雾网络的去雾能力;通过增量式训练方法提高学生去雾网络对增补数据集的处理能力。结果表明:所提算法对ITS、OTS以及真实雾图上皆具有较好的处理能力,在保证去雾图像像素结构完整、颜色不失真的同时具有较好的去雾效果,算法处理后的图像在主观视觉效果和客观评价指标上皆优于对比算法。In order to improve the processing ability of the depth-neural network dehazing algorithm to the supplementary data set,and to make the network differently process the image features of different importance to improve the dehazing ability of the network,an incremental dehazing algorithm based on multiple migration of attention is proposed.The teacher's attention generation network in the form of Encoder-Decoder extracts the multiple attention of labels and haze,which is used it as the label of the characteristic migration media network to constrain the network training to form the migration media attention as close as possible to the teacher's attention.The attention is integrated into the characteristics of the student's dehazing network to improve the dehazing ability of the student's dehazing network.The incremental training method is used to improve the processing ability of students'dehazing network to the supplementary data set.The results show that the proposed algorithm has good processing ability on ITS,OTS and real hazy images,and has good dehazing effect while ensuring the integrity of pixel structure and color distortion of the dehazing image.The image processed by the algorithm is superior to the contrast algorithm in subjective visual effect and objective evaluation index.

关 键 词：深度学习 图像去雾 迁移注意力 增量式训练 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]