【作者】 方敏；

【导师】 王年； 张艳；

【作者基本信息】 安徽大学 ， 信号与信息处理， 2020， 硕士

【摘要】 足迹作为现场遗留率较高的痕迹,不仅在刑侦领域有着举足轻重的地位,而且在鞋业生产、医疗等领域有着广泛的应用。然而针对足迹分类问题,传统的研究方法通常依赖相关足迹专家提取具有物理意义的足迹特征,但是因相关专家的经验限制与提取特征方法不同,使得不同研究者得到的特征好坏难以界定,另外对于足迹一些抽象特征,如足迹各部位的关系等也难以提取。随着深度学习的发展,卷积神经网络在许多领域取得了显著的效果。相关足迹研究者也不断将深度学习方法应用于足迹分类,同时也面临一些问题,一方面同一人大量的足迹样本获取难度较大,而一个或几个足迹样本往往获取较简单;另一方面足迹图像又属于细粒度图像范畴,不同人足迹间的差别比较细微,一般难以辨别其所属类别。针对上述足迹研究存在的不足,本文通过元学习方法使模型在任务空间学习大量的任务分布,从而提高模型对样本较少新任务的泛化能力。主要研究内容如下:(1)按采集规范采集足迹数据并进行预处理。基于光学足迹采集仪器,从静态和动态角度,本文分别采集了静态114人和动态142人的两个光学足迹数据。为了让模型获得丰富的元知识,本文采用63人的压力足迹图像以及部分来自南京相关部门的光学足迹图像扩充了数据集。为了获得高质量足迹样本,对采集的足迹样本进行去噪、标准化处理,以减小数据分布的差异。(2)为了克服小样本学习的关系网络存在的不足,提出了一种通道-空间注意力的足迹分类算法。该方法将通道注意力与空间注意力相结合,用于挖掘样本更丰富的信息,通过在不同卷积层嵌入联合注意力机制,既可以从多通道特征图中获得关注度高的特征图,也可以从单张特征图上获得不同区域的特征信息。最后有效地将两种信息融合,从而提高特征的表达能力。本文在小样本常用数据集上进行了大量的实验,表明了该方法的有效性,然后将该算法应用到光学足迹分类中,通过将左、右足分开训练及测试,实验取得了较好的结果。(3)由于光学足迹样本量较少,而通道-空间注意力模型包含大量的训练参数,可能导致网络参数无法得到最优值,因此本文提出了双线性足迹分类网络。该网络针对足迹细粒度问题,在特征提取阶段使用了双线性网络结构,该方法利用平移不变性进行足迹图像局部像素交互建模,获得不同区域的重要信息。最后利用欧氏度量进行样本相似性计算。特别地,对于多个支持样本,通过在特征空间计算样本的均值以得到包含丰富特征信息的类原型表示。本文在足迹数据集上进行大量实验,结果表明该算法具有较好的性能。更多还原

【Abstract】 The footprint is the trace that has a high legacy rate on the scene,which not only has a pivotal position in the field of criminal investigation,but also has a wide range of applications in the production of shoes,medical and so on.However,for footprint classification,the traditional classification research often rely on the experienced footprint experts that extract the features of the physical significance of the footprint.Duo to their relevant experience is limited,and the method for feature extraction also is different,so that different researchers get the features that is difficult to define good or bad,in addition,some abstract feature relationship between different regions of the footprint is hard to be extracted.With the development of deep learning,convolutional neural network has achieved remarkable performance in many fields.Relevant footprint researchers have also applied the deep learning methods to solve the footprint classification,at the same time,there are some problems.On the one hand,it is difficult to obtain a large number of footprint samples from the same person,while it is often easier to obtain one or several footprint samples.On the other hand,the footprint is fine-grained image,and the difference between the footprints of different people is more subtle,so that it is generally difficult to distinguish the category.Aiming at shortcomings existing footprint research,this paper can learn a large number of task distributions in the task space based on the meta-learning,so as to improve the generalization ability of the model to new tasks with few samples.The main research contents are as follows:(1)The footprint data was collected and preprocessed according to collection specifications.From a static and dynamic perspective,this paper collected static 114 and dynamic 142 people’s optical footprint data respectively based on the optical footprint acquisition instrument.In order to acquire more meta-knowledge,pressure footprint images of 63 people and some optical footprint images of people from relevant departments in Nanjing were used to expand the data set in this paper.In order to obtain high-quality footprint samples,the collected footprint samples are de-noised and standardized to reduce the difference in data distribution.(2)In order to overcome the shortcomings of relation network of few-shot learning,a footprint classification algorithm of channel-spatial attention is proposed.This method combines channel attention with spatial attention to mine the rich information of the samples.By embedding the joint attention mechanism in different convolutional layers,the focused feature map can be obtained from the multi-channel feature map,while some information of different regions can be obtained from the single feature map.Finally,the two kinds of information are fused effectively to improve the ability of feature expression ability.In this paper,a large number of experiments were carried out on common datasets,which showed the effectiveness of the method,and this algorithm was applied to the optical footprint classification,and the good results were obtained by training and testing the left and right feet separately.(3)Due to there are few optical footprint samples,and the channel-spatial attention model contains a large number of training parameters,so that the network parameters cannot get the optimal value.This paper proposes a bilinear network for footprint classification.For finegrained footprint,the bilinear network structure is used during feature extraction,this method can model between footprint image pixels in translation invariant,and obtain different regional important information.Finally,the similarity between different samples are computed using Euclidean.In particular,for several support samples,class prototype representation is obtained by computing mean of a few labeled samples belonging to a class,and the class prototype contains rich feature information.In this paper,a large number of experiments are carried out on the footprint data set,and the results show that this algorithm has better performance.更多还原