开集鞋底花纹分类算法研究

发布时间：2018-03-25 14:25

本文选题：鞋底花纹开集分类　切入点：置信度　出处：《大连海事大学》2017年硕士论文

【摘要】：鞋底花纹是刑事侦查的重要物证之一,为了发挥鞋底花纹对案件侦破的重大作用,公安人员需要建立鞋印库。当新的鞋底花纹自动添加到鞋印库时,会出现其可能属于鞋印库中的某个已知类别也可能不属于鞋印库中任何一种类别的情况。但是如果将现有效果很好的开集分类算法直接用于鞋底花纹,分类器的性能会下降很多,因此需要设计鞋底花纹的开集分类算法。基于此,本文提出了开集鞋底花纹分类算法研究,主要工作如下:1)给出了基于置信度的CSoftmax算法该算法针对目前Softmax分类算法应用于开集场景时存在的问题,通过引入置信度增大已知类别和新类别之间的概率分布差异,减小二者概率之间的重叠区域,以此提高开集分类算法的准确率。实验结果表明:本文算法的性能指标对于存在明显隔离带的鞋印数据集取得更好的效果,其中在包含9294幅鞋印图像的数据集上 AUC 达到了 76.33%。2)给出了基于距离对比的DKNFST算法针对零空间下已知类别和新类别的距离特点,本文不仅利用距离最近这一信息,还考虑距离最远的两个类别包含的有用信息,以此增大已知类别和新类别之间的分布差异。实验结果表明:本文算法在有明显隔离带的鞋印数据集上AUC达到了 74.16%。3)给出了基于流形一致性的MKNFST算法本文利用待检测样本在变换后的低维零空间,以及原始高维空间的流形一致特性,通过将待检测样本的两种空间特性融合来设计分类器,进一步提高分类的准确率。实验结果表明:本文算法在有明显隔离带的鞋印数据集上AUC达到了83.77%。
[Abstract]:Sole pattern is one of the important material evidence in criminal investigation. In order to play the important role of sole pattern in the detection of cases, public security personnel need to establish shoe print bank. When the new sole pattern is automatically added to the shoe print store, It may or may not belong to any of the categories in the shoeprint library. However, if the existing open-set classification algorithm is used directly for the sole pattern, The performance of the classifier will decrease a lot, so it is necessary to design an open set classification algorithm for sole pattern. The main work of this paper is as follows: (1) this paper presents the CSoftmax algorithm based on confidence degree. Aiming at the problems existing in the application of Softmax classification algorithm to open set scene, the confidence degree is introduced to increase the difference of probability distribution between known and new categories. In order to improve the accuracy of the open set classification algorithm, the experimental results show that the performance index of this algorithm is better for the shoe print data set with obvious separation belt. In the data set containing 9294 shoeprint images, the AUC reaches 76.33.2) the distance characteristic of known and new categories in zero space is given by DKNFST algorithm based on distance contrast. This paper not only uses the nearest distance information, Considering also the useful information contained in the two categories most distant, The experimental results show that the AUC of the shoeprint data set with obvious straps is 74.16.3) the MKNFST algorithm based on manifold consistency is given in this paper. Detection of samples in the transformed low-dimensional zero space, And the manifold consistency of the original high-dimensional space, the classifier is designed by merging the two spatial characteristics of the sample to be detected. The experimental results show that the proposed algorithm achieves 83.77 AUC on the shoeprint data set with obvious separation band.
【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】