结露和结霜现象的视觉特征提取与呈现模式识别

项目名称： 结露和结霜现象的视觉特征提取与呈现模式识别

项目编号： No.61502358

项目类型： 青年科学基金项目

立项/批准年度： 2016

项目学科： 计算机科学学科

项目作者： 朱磊

作者单位： 武汉科技大学

项目金额： 20万元

中文摘要： 地表结露和结霜现象的发生状态、持续时间以及表征呈现模式是近地面大气与土壤水分交换研究中的一类重要数据。然而，如何以自动的方式测量上述气象参数是天气现象观测中的难题。本课题探索使用图像特征来描述这两类现象视觉特性的方法，并实现对其气象参数的自动测量。内容包括：1）研究结露与结霜现象表征的视觉显著性机理，提出利用单类学习模型构建先验知识的方法，并以此知识引导显著性区域检测模型，以实现结露和结霜区域的自动提取；2）建立结露和结霜区域图像特征的静态量、时变量以及实时大气状态数据与现象演化过程的联系，并提出基于异类特征的联合判别模型，以识别现象发生的状态和测量现象的持续时间；3）提出利用中层特征编码结合多类型图像特征融合的方法来提高特征的描述能力，以解决现象表征呈现模式难分类的问题。本课题的研究成果为获取更丰富的结露和结霜现象信息提供了一种有效的途径，也为两类现象的自动观测提供一种可行的解决方案。

中文关键词： 结露和结霜现象；外观特征提取；显著性分析与检测；异类特征联合学习；中层特征编码

英文摘要： The occurrences, durations and visual patterns of dew and frost happened near land surface are a type of important data for the research on moisture exchanges between the nearby atmospheric and the soil. However, automatically measuring the aforementioned meteorological parameters of dew and frost is still a challenge in surface weather observation. This research explores the methods for closely associating image features with the visual appearance of dew and frost that occur in nature, and investigates the approaches for automatically identifying the meteorological parameters of dew and frost. The main contributions of this proposal include: 1) we locate the deposition regions of dew or frost in images with an innovative saliency detection method, which formulates the priors of the proposed attention model by an one-class learning procedure; 2) We propose to identify the occurrences of dew or frost as well as measure the dew or frost duration with a discriminant model, which is optimized by integrating the image features that are spatially and temporally extracted from the deposition regions with the real-time meteorological factors; 3) Recognition of dew and frost patterns is investigated. For a reliable classification of those patterns, the extracted low-level image features are enriched by a mid-level feature encoding technique, followed by a multi-channel feature fusion framework in a sophisticate manner. This study proposes an effective approach for obtaining more informative data of dew and frost than other methods. In addition, it also potentially provides a practical solution for automated dew and frost observation.

英文关键词： Dew and frost observation;Appearance-based features extraction;Saliency analysis and detection;Joint learning of heterogeneous features;Mid-level feature encoding