| 摘要: |
| 图像是信息的重要承载形式。雾霾的出现降低了图像采集设备采集到的图像质量，容易出现色彩暗淡、对比度和饱和度降低、细节信息丢失等问题，直接影响了有用信息的表达和利用。目前对图像去雾的研究多采用深度学习的方法，卷积神经网络代替了人工特征提取方式，取得了优于传统算法的去雾效果，但普遍存在着对真实世界雾霾图像和清晰图像对的依赖。无监督学习的方法带来了新的解决思路。从监督学习和无监督学习的角度对有代表性的深度学习图像去雾算法进行分类，归纳了常用的数据集、评价指标，概括分析了有影响力的去雾模型的核心思想，总结了各算法的优缺点和适用场景。针对目前工作存在的不足，探索了下一步研究的方向。 |
| 关键词: 深度学习 图像去雾 监督学习 无监督学习 |
| DOI：10.20079/j.issn.1001-893x.220402002 |
|
| 基金项目:广西科技重大专项(桂科AA210770071) |
|
| Research progress of image dehazing algorithms based on deep learning |
| LI Ya,WANG Lie,TENG Sihang,CUI Lijuan,LAN Zhengjie |
| (School of Computer and Electronic Information，Guangxi University，Nanning 530004，China) |
| Abstract: |
| Image is an important form of information.The appearance of haze reduces the quality of images collected by image acquisition equipments.Problems such as dull color,reduced contrast and saturation,and loss of detail information are prone to occur,which directly affect the expression and utilization of useful information.At present,the researches on image dehazing mostly use deep learning methods.The convolutional neural networks replace artificial feature extraction methods and have achieved better dehazing effects than traditional algorithms.But there is widespread reliance on real-world haze image and clear image pairs.The methods of unsupervised learning bring new solutions.The representative deep learning image dehazing algorithms are classified from the perspective of supervised learning and unsupervised learning,and the commonly used datasets and evaluation indicators is summarized.The core idea of the influential dehazing models is summarized and analyzed.The advantages,disadvantages and applicable scenarios of each algorithms are summarized.In view of the shortcomings of the current work,the direction of further research is explored. |
| Key words: image dehazing deep learning supervised learning unsupervised learning |
