融合空域与频域特征的能见度监督学习  

Supervised learning for visibility combining features in spatial domain with that in frequency domain

作　　者：许茜[1] 李岩[1] 郝红卫[2] 

机构地区：[1]北京科技大学计算机与通信工程学院,北京100083 [2]中国科学院自动化研究所数字内容技术与服务研究中心,北京100190

出　　处：《计算机应用》2014年第3期892-897,共6页journal of Computer Applications

基　　金：国家自然科学基金资助项目(61105018;61175020);国家公益行业(气象)科研专项(GYHY201106039;GYHY201106047)

摘　　要：大气能见度不仅影响海路空交通运输和居民出行,而且是空气质量检测的主要指标。现有基于图像处理的能见度评测方法存在采用固定公式计算、稳定性差、对适用环境要求苛刻等问题。能见度监督学习测量方法直接从观测场景图像中提取空域边缘相关特征与频域能量分布特征构成高维特征向量,无需人工设立目标物或对观测场景建模。它通过k最近邻(kNN)选取与待测图像相似样本进行支持向量回归机(SVR)的训练,动态构建图像特征与能见度之间的学习模型,将各种能见度影响因子隐藏于模型内。对自然场景的测量实验结果表明,该方法的测量正确率最高可达96.29%,且具有良好的稳定性和实时性,操作简单,便于大规模推广。Atmospheric measurement not only impacts marine, land, and air transportation and resident trip, but also is a leading indicator of air quality. The existing visibility estimators based on image processing have problems such as constant computational formulae, poor stability and stringent requirement of application environment. Visibility measurement with supervised learning extracted features related to image edge in spatial domain and features of energy distribution in frequency domain to constitute the high-dimensional feature vector directly from observed scene images, and needed no manual object installing and modeling of the observed scene. It trained Support Vector Regression （SVR） model on the samples that were similar to test image and chosen by k-Nearest Neighbor （kNN）, dynamically established the learning model between image features and visibility, and hid various impact factors of visibility in the model. The experimental results of natural scene show that the accuracy of the method can be as high as 96.29%, and moreover, it has good stability and real-time and simplicity of operation so that it is propitious for generalization in large scale.

关 键 词：大气能见度 空域 频域 傅里叶变换 周向谱 监督学习 

分 类 号：TP391.413[自动化与计算机技术—计算机应用技术]