基于内容的云图检索技术研究

发布时间：2018-03-19 16:16

本文选题：卫星云图　切入点：基于内容的云图检索　出处：《宁波大学》2017年硕士论文　论文类型：学位论文

【摘要】：气象卫星技术的迅速发展带来了气象云图数据的爆炸性增长,传统依靠人工标注文本检索的云图检索方法越来越不能满足气象需求。基于内容的图像检索作为一种管理海量图像数据的检索技术表现出非常高效的检索性能,本文在分析卫星云图特点的基础上结合基于内容的图像检索技术进行了基于内容的云图检索的关键技术与应用研究,主要工作内容如下:(1)基于分数阶达尔文粒子群优化算法(FODPSO)与FCM聚类算法的云层信息提取研究。传统FCM聚类方法进行云层信息提取时,算法的收敛容易受初始聚类中心的影响,陷入局部最优解。FODPSO采用自然选择机制全局寻找最优解,能极大的避免陷入局部最优值,利用全局寻优性能非常好的分阶达尔文粒子群优化算法优化模糊C均值初始聚类中心。并用改善初始聚类中心的模糊C均值聚类算法进行云系信息的提取,为下阶段有效特征的提取做好准备。(2)多通道卫星云图云图融合研究。不同通道获取的云图蕴含了不同天气特征,融合多通道的云图能够形成包含多种天气特征信息的云图,利用这样的云图进行检索能够匹配出更加准确的历史相似云图。本文提出基于NSST与自适应PCNN相结合的方法进行红外和可见光卫星云图融合。实验结果表明融合云图中云系特征明显,含有丰富的边缘、纹理细节信息,具有更高的清晰度,蕴含了更多的气象信息。运用融合云图进行检索,能获得比单一类型云图更好的检索效果。(3)准确有效表示云图“内容”的特征提取研究。基于内容的云图检索技术的基础是提取能够表示云图“内容”的有效特征。根据云图固有的特性,在提取云层信息的基础上提取灰度、纹理和形状三种可靠的底层特征表示云图“内容”。灰度特征利用云图的灰度直方图方法提取,纹理特征采用具有良好抗噪声和灰度平移不变的LTrP算子提取。形状特征采用具有尺度、旋转和平移不变性的Krawtchouk矩来提取。(4)云图多特征决策融合检索研究。在多特征决策融合检索过程中,特征权重的选择直接影响着检索的准确性,传统通过人工分配不同特征权重进行融合的方法需要进行非常多次的实验才可能找到比较好的检索效果,但随着融合特征种类的增多,人工权重的确定会越来越变得没有效率。本文依据不同特征检索结果的相似度量得分排序曲线下的面积作为特征权重,从而自适应确定每种特征的权重。本文研究表明特征权重的大小与相似性度量排序得分曲线下的面积呈负相关,越好的特征其得分曲线下的面积越小,越差的特征面积越大。
[Abstract]:The rapid development of meteorological satellite technology has brought explosive growth of meteorological cloud map data. The traditional cloud image retrieval method based on manual tagging text retrieval can not meet the meteorological requirements. As a retrieval technology for managing massive image data, content-based image retrieval has a very efficient retrieval performance. Based on the analysis of the characteristics of satellite cloud images, the key technologies and applications of content-based image retrieval are studied in this paper. The main work is as follows: (1) based on Fractional Darwin Particle Swarm Optimization (FODPSO) and FCM clustering algorithm, cloud information extraction is studied. The convergence of traditional FCM clustering algorithm is easily affected by the initial clustering center. Fall into the local optimal solution. FODPSO uses natural selection mechanism to find the global optimal solution, which can greatly avoid falling into the local optimal value. The fuzzy C-means initial clustering center is optimized by using the hierarchical Darwinian particle swarm optimization algorithm, which has very good global optimization performance, and the cloud information is extracted by using the fuzzy C-means clustering algorithm which improves the initial clustering center. To prepare for the next phase of effective feature extraction. 2) Multi-channel satellite cloud image fusion research. Different channels of cloud image contains different weather features, the fusion of multi-channel cloud image can form a cloud image containing a variety of weather characteristics. The retrieval of this kind of cloud image can match a more accurate historical similar cloud image. This paper presents a method of infrared and visible satellite cloud image fusion based on NSST and adaptive PCNN. The experimental results show that the fused cloud can be fused. The features of the cloud system are obvious in the picture. Contains rich edge, texture details, higher clarity, more meteorological information. The research on feature extraction for accurately and effectively representing "content" of cloud image can be obtained. The basis of content-based cloud image retrieval technology is to extract valid features that can represent "content" of cloud image. Based on the inherent characteristics of the cloud map, On the basis of extracting cloud information, three reliable underlying features, texture and shape, represent the "content" of the cloud image. The gray feature is extracted by using the gray histogram method of the cloud image. The texture feature is extracted by LTrP operator with good noise resistance and invariant gray level translation, and the shape feature is based on scale. In the process of multi-feature decision fusion retrieval, the selection of feature weights directly affects the accuracy of retrieval. The traditional method of fusion by manually assigning different feature weights requires a lot of experiments in order to find a better retrieval effect, but with the increase of the types of fusion features, The determination of artificial weights will become more and more inefficient. This study shows that the size of feature weight is negatively correlated with the area under the similarity measure ranking score curve, and the smaller the area under the score curve is, the bigger the feature area is.
【学位授予单位】：宁波大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP751

【相似文献】