鞋印图像多标签聚类算法研究

发布时间：2019-01-02 11:55

【摘要】：鞋印图像是犯罪现场的重要证物之一,对大量鞋印图像进行自动分类管理可以提高案件侦破效率。现有的针对鞋印图像的聚类算法都集中在单标签聚类问题上,但是鞋印图像库存在类别间差异较小的情况,而磨损和残缺等因素更模糊了类别之间的差别,即没有明显的隔离带将不同类别很明显的分隔开;同时,一幅鞋印图像往往包含了一种或多种不同类型的花纹图案,所以存在不同鞋印图像包含相同类型花纹的情况,因此,一幅鞋印图像可以根据其所含有花纹图像的种类划分到不同的类别。本文在分析鞋印图像特点的基础上,提出了鞋印图像多标签聚类算法。主要工作如下:1)提出了基于改进模糊C-均值聚类算法的鞋印图像多标签聚类算法本算法改进了模糊C-均值聚类算法中的隶属度矩阵,在考虑图像与聚类中心关系的同时,考虑图像与图像之间的关系,提出了适应于鞋印图像的多标签聚类算法。算法从高密度点开始聚类,利用改进后的隶属度矩阵确定图像与类别间的关系,实现鞋印图像多标签聚类。本算法在实际的鞋印数据集上的F-measure值达到了 79.09%。2)提出了基于随机游走的鞋印图像多标签聚类算法算法首次将随机游走的思想运用到对鞋印图像相似度矩阵的优化中,首先从每幅图像开始聚类,然后进行类别的合并,最后对多标签图像再标记,实现鞋印图像多标签聚类,本算法既考虑了图像之间的关系,也考虑了类别之间的关系。本算法在实际的鞋印数据集上的F-measure值达到了 78.34%。3)提出了基于概率潜语义分析的鞋印图像多标签聚类算法本算法应用概率潜语义分析模型对鞋印图像基元语义词汇表进行学习,得到图像与潜在主题之间的概率分布矩阵,建立图像之间的关系,经过单标签聚类和多标签聚类两个过程,实现了鞋印图像数据库的多标签聚类。本算法在实际的鞋印数据集上的F-measure值达到了 73.61%。
[Abstract]:Shoeprint image is one of the important evidence of crime scene. The automatic classification management of a large number of shoeprint images can improve the efficiency of case detection. The existing clustering algorithms for shoe printing image are focused on the single label clustering problem, but the difference between categories is small, and wear and tear and incomplete factors make the difference between categories more blurred. That is, there is no obvious separation between the different categories; At the same time, a shoe print image often contains one or more different types of pattern patterns, so there are different shoe printing images containing the same type of pattern, so, A shoe print image can be divided into different categories according to the type of pattern it contains. On the basis of analyzing the characteristics of shoe printing image, this paper proposes a multi-label clustering algorithm for shoe print image. The main work is as follows: 1) A new multi-label clustering algorithm based on improved fuzzy C-means clustering algorithm is proposed. This algorithm improves the membership matrix of fuzzy C-means clustering algorithm. Considering the relationship between image and clustering center and the relationship between image and image, a multi-label clustering algorithm for shoe printing image is proposed. The algorithm starts with high density points and uses the improved membership matrix to determine the relationship between image and category to realize the multi-label clustering of shoeprint image. In this algorithm, the F-measure value on the actual shoe print data set reaches 79.09.2) A new multi-label clustering algorithm based on random walk is proposed. The idea of random walk is applied to the shoe print image for the first time. In the optimization of likelihood matrix, First, each image is clustered, then the category is merged. Finally, the multi-label image is re-labeled to realize the multi-label clustering of shoeprint image. This algorithm considers the relationship between images as well as the relationship between categories. This algorithm has a F-measure value of 78.34.3 on the actual shoe print data set.) A multi-label clustering algorithm for shoe print image based on probabilistic latent semantic analysis is proposed. This algorithm applies probabilistic latent semantic analysis model to shoe print image. The pixel semantic vocabulary is learned, The probability distribution matrix between the image and the potential topic is obtained, and the relationship between the images is established. Through the two processes of single label clustering and multi-label clustering, the multi-label clustering of shoeprint image database is realized. The F-measure value of this algorithm on the actual shoe print data set is 73.61.
【学位授予单位】：大连海事大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：D918.91;TP391.41

【相似文献】