基于多教师引导的知识蒸馏图像去雾算法  

An Image Dehazing Algorithm Based on Multi-teacher Guided Knowledge Distillation

作　　者：崔智高 兰云伟 王念 张炜 CUI Zhigao;LAN Yunwei;WANG Nian;ZHANG Wei(Rocket Force University of Engineering,Xi’an 710025,Shaanxi)

机构地区：[1]火箭军工程大学,陕西西安710025

出　　处：《火箭军工程大学学报》2024年第5期35-43,共9页Journal of Rocket Force University of Engineering

基　　金：陕西省自然科学基金(2023-JC-YB-501)。

摘　　要：针对现有图像去雾算法容易导致去雾图像颜色失真等问题,采用多教师知识蒸馏方式,融合基于先验信息的图像复原方法与基于深度学习的图像去雾方法的互补优势,提出了一种基于多教师引导的知识蒸馏图像去雾算法。该算法选择两个预训练的去雾模型作为教师网络,并利用教师模型蕴含的知识指导学生网络的训练;此外,算法还通过构建一个多尺度学生网络以及设计特征注意残差密集块来提高学生网络的特征提取能力。结果表明:与其他去雾算法相比,所提算法在合成图像和真实场景图像上均表现出优异的去雾性能,在SOTS室外数据集上的峰值信噪比和结构相似度分别达到了23.57 d B和0.934。To solve the problem of serious color distortion caused by existing image dehazing algorithms,a single image dehazing algorithm based on multi-teacher guided knowledge distillation dehazing network(MGKDN)was proposed,which combined the complementary advantages of prior-based image restoration methods and deep learning-based image dehazing methods.In the proposed MGKDN,two published models namely principled synthetic-to-real dehazing network(PSD)and enhanced pix2pix dehazing network(EPDN)were used as teacher networks.Moreov-er,to build an efficient student network,a feature attention residual dense block,which improved feature extraction and keep the student network compact was adopted.Compared to other dehaz-ing algorithms,the proposed algorithm achieves excellent performance on both synthetic and real scene images.It achieves PSNR of 23.57 dB and SSIM of 0.934 on the SOTS outdoor dataset.

关 键 词：图像处理 图像去雾 知识蒸馏 多教师网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]