【作者】 刘健；

【导师】 熊惠霖；

【作者基本信息】 上海交通大学 ， 控制科学与工程， 2018， 硕士

【摘要】 近年来,深度学习受到了越来越多的关注和研究,同时推动了很多图像识别相关领域的实际应用。本文以深度学习为基础,研究了它在细粒度图像识别中的应用,包括细粒度生物图像识别和人体再识别。细粒度生物图像识别根据图像中的生物目标进行分类;人体再识别技术是视频监控的基础,任务是对不同摄像头下的人体图像进行匹配。人体图像之间差异很小,区分不同个体可以看作广义的细粒度图像识别任务。本文主要分为以下两个部分。1、本文设计了层次分类网络对生物图像进行识别。不同的生物在分类学上对应于不同的层级,可以组织成树形结构。由于以往的方法大都忽略了类别间的层次信息,仅仅在最低层级分类。针对此问题,本文设计了一种层次分类网络。该网络从上至下进行分类,每个子分类器只对同一父节点下的类别分类。下层识别的种类差异更小,以此学习图像中更精细的特征。实验结果表明,相较于忽略类别间信息的扁平分类网络,层次分类网络的运算量几乎没有增加,分类准确率得以提高。2、在人体再识别算法中,本文利用了简化的双线性卷积神经网络,并在全局特征中融入了精细的局部特征。双线性卷积神经网络最初是为解决细粒度的图像识别问题提出的,提取到的双线性向量维度较高,不利于后续计算。为了解决此问题,本文采用了其简化方法进行降维。针对主干网络特征图的尺寸较小、无法充分保留人体局部特征的情况,我们添加了额外的浅层网络,来获取更精细的局部特征,并融合到主干网络的全局特征中。在网络的目标函数上,本文选取了直方图损失函数,该方法对样本的利用率更高,而且避免了人体再识别中常用损失函数中的参数设定。实验结果表明,本文的方法比相关方法性能有明显提升。更多还原

【Abstract】 In recent years,deep learning has received more and more attention and research,promoting practical application in many fields about image recognition at the same time.This paper based on deep learning,researched its application in fine-grained visual categorization,including fine-grained biology image recognition and person re-identification.Fine-grained biology image recognition classifies image according to biology object in it;Person re-identification is the base of video monitoring,and it aims to match person images from different cameras.There are few differences in person images,so distinguishing them can be considered the task of generalized fine-grained visual categorization.This paper is divided into following two sections.1、This paper proposed one hierarchical classification network to recognize biology images.Different species correspond to different levels in biological taxonomy,and we can organize them in a tree structure.Because previous methods almost neglected hierarchical information between different species and only classified in the lowest level.For this problem,this paper proposed one hierarchical classification network.This network classifies from top to bottom,and every sub-classifier only classifies species having the same father node.The differences between species recognized by lower-level classifier is fewer,so the classifier can learn finer feature.The results of experiment show,compared to the flat classification network neglecting the information between different species,the computation of hierarchical classification network did not increase,but the classification accuracy was improved.2、For person re-identification,this paper applied compact bilinear convolutional neural network and merged finer local features into global feature.Bilinear convolutional neural network was initially proposed to handle fine-grained visual categorization problem,and the bilinear vector extracted by it is usually of high dimension,so it is not convenient for follow-up computation.To tackle this problem,this paper applied compact method to reduce feature’s dimension.For the situation that feature maps from main network are small and they cannot preserve person local features fully,we added extra shallow sub-network to get finer features and merged them to global features from main network.As object function,this paper chose histogram loss,because this function has higher data utilization and avoids setting parameters of the ordinary loss function in person re-identification.The result of experiment showed,this paper’s method improved performance on related methods.更多还原