一种基于图的特征选择方法

在很多的机器学习和数据挖掘任务中,特征子集选择是重要的数据预处理步骤之一。提出一种基于图方法的无监督式特征选择方法(GBFS),构造一个以样本数据为顶点,数据间相似性作为边的图,再根据各特征的得分优先选择那些具有局部信息保持和全局区分能力的特征。实验结果表明,基于该方法选择的特征子集,在大多数情况下都能取得较好的分类效果。

基于图的关键词提取方法研究

关键词提取是自然语言处理(NLP)应用中的一个重要步骤,其目的是捕获文档的主题,关键词提取广泛应用于信息检索、文本挖掘和自然语言处理中.在关键词提取方法中,最近的研究主流是基于图的关键词提取方法.为此,研究了基于图的关键词提取方法,分析比较了各种基于图的方法,为今后的研究提供了指导,并鼓励开发新的基于图的关键词提取方法.

结合改进密度峰值聚类的LGC半监督学习方法优化

基于图的局部与全局一致性(LGC)半监督学习方法具有较高的标注正确率,但时间复杂度较高,难以适用于数据规模较大的实际应用场景。从缩小图的规模入手,提出一种全局一致性优化方法。使用改进后的密度峰值聚类算法,迭代地从数据集中筛选出多个中心点,以每个中心点为簇中心进行局部聚类,并以中心点为顶点构建图,实现基于LGC的半监督学习。实验结果表明,优化后的LGC方法在D31、Aggregation等数据集上具有较好的鲁棒性,在标注正确率和算法执行时间上优势明显。

增强现实系统建模方法研究

图象与几何模型相结合的增强现实技术是充分利用基于图象与基于几何模型两种建模方法各自优点的有效途径,近年来已成为研究热点。针对目前几乎没有成熟的增强现实系统的现状,本文在分析基于图象的关键技术、基于几何模型的关键技术、图象与几何模型相结合增强现实一般理论的基础上,阐明一个图象与几何模型相结合的增强现实系统的建模和实现方法,以使增强现实技术走向实用化。

Sinogram fusion-based metal artifact correction method

To solve the problem that metal artifacts severely damage the clarity of the organization structure in computed tomography（CT） images, a sinogram fusion-based metal artifact correction method is proposed. First, the metal image is segmented from the original CT image by the pre-set threshold. The original CT image and metal image are forward projected into the original projection sinogram and metal projection sinogram, respectively. The interpolation-based correction method and mean filter are used to correct the original CT image and preserve the edge of the corrected CT image, respectively. The filtered CT image is forward projected into the filtered image sinogram. According to the position of the metal sinogram in the original sinogram and filtered image sinogram, the corresponding sinograms PM^D （ in the original sinogram） and PM^C （ in the filtered image sinogram）can be acquired from the original sinogram and filtered image sinogram, respectively. Then, PM^D and PM^C are fused into the fused metal sinogram PM^F according to a certain proportion.The final sinogram can be acquired by fusing PM^F , PM^D and the original sinogram P^O. Finally, the final sinogram is reconstructed into the corrected CT image and metal information is compensated into the corrected CT image.Experiments on clinical images demonstrate that the proposed method can effectively reduce metal artifacts. A comparison with classical metal artifacts correction methods shows that the proposed metal artifacts correction method performs better in metal artifacts suppression and tissue feature preservation.

Image-based Water Level Measurement Method under Stained Ruler

This paper propoes the water level measuring method based on the image, while the ruler used to indicate the water level is stained. The contamination of the ruler weakens or eliminates many features which are required for the image processing. However, the feature of the color difference between the ruler and the water surface are firmer on the environmental change compare to the other features. As the color differeaces are embossed, only the region of the ruler is limited to eliminate the noise, and the average image is produced by using several continuous frames. A histogram is then produced on the height axis of the produced intensity average image. Local peaks and local valleys are detected, and the section between the peak and valley which have the greatest change is looked for. The valley point at this very moment is used to detect the water level. The detected water level is then converted to the actual water level by using the mapping table. The proposed method is compared to the ultrasonic based method to evaluate its accuracy and efficiency on the various contaminated environments.

High-resolution remote sensing image-based extensive deformation-induced landslide displacement field monitoring method

Landslide is one of the multitudinous serious geological hazards. The key to its control and reduction lies on dynamic monitoring and early warning. The article points out the insufficiency of traditional measuring means applied for large-scale landslide monitoring and proposes the method for extensive landslide displacement field monitoring using high- resolution remote images. Matching of cognominal points is realized by using the invariant features of SIFT algorithm in image translation, rotation, zooming, and affine transformation, and through recognition and comparison of characteristics of high-resolution images in different landsliding periods. Following that, landslide displacement vector field can be made known by measuring the distances and directions between cognominal points. As evidenced by field application of the method for landslide monitoring at West Open Mine in Fushun city of China, the method has the attraction of being able to make areal measurement through satellite observation and capable of obtaining at the same time the information of large- area intensive displacement field, for facilitating automatic delimitation of extent of landslide displacement vector field and sliding mass. This can serve as a basis for making analysis of laws governing occurrence of landslide and adoption of countermeasures.

用于有遮挡环境漫游的五维光场模型

将四维光场函数扩展为五维光场函数，保持了四维光场函数的优点，并且能用于有遮挡的环境，与五维完全函数相比，它不需要输入全景图的对应点。该方法会产生大量的图象数据，采用矢量量化的方法可以将数据大幅度压缩，而且能快速解压缩。介绍了五维光场的表示和采样、重采样的过程，最后给出了实验结果并讨论了后续的工作。