基于B-CNN模型的细粒度分类算法研究

发布时间：2018-03-25 04:32

本文选题：细粒度分类　切入点：分类错误　出处：《中国地质大学(北京)》2017年硕士论文

【摘要】：近年来伴随着大数据的增长,图像分类任务得到了迅猛发展,分类任务也由二分类发展为简单的粗分类进而发展为细粒度分类。这里所说的粗分类大多数情况下是把不同的物体区分开来,比如把猫、狗、树、车分开,它们之间有着显著的差异,比较容易分类。而细粒度分类则是把同一个物体继续细分,比如不同类别的鸟,不同型号的飞机,人脸识别等等,它们之间的差异通常非常小,而且它们的拍照角度,光照,遮挡都会对分类造成更大的困扰,所以更加困难。随着深度学习的发展,图像分类任务变得越来越容易,人们对于图像分类的要求也在不断提高,细粒度分类任务应运而生,包括对花的品种、鸟的种类、人脸识别等。基于机器学习的图像分类任务包含两部分内容:图像特征提取,分类,特征提取的好坏直接关乎分类精度,特征维数越大,准确率越高,然而当特征维数太高时,又会导致占用内存较多,计算量大等问题。卷积神经网络常被人用来提取特征,一般而言,网络越深,特征越好,随着层数的加深,然而当网络到达一定深度后,通过增加深度提高准确率变得越来越难。因此研究者们提出了很多种方法用来解决这个问题,例如基于部件的摸型,但它的缺点是效率低,人工标记任务繁重;而B-CNN模型则同时解决了这两个问题,它只需要训练集样本的标签,不需要目标的边界框及部件的标注,同时又有较高的准确率。B-CNN模型虽然在一定程度上解决了细粒度分类任务的问题,但是它的训练和分类是针对所有样本的,对于部分易混淆的类不能有针对的区分,本文在B-CNN模型的基础上做了不同的改进尝试,其中,根据B-CNN模型得到的首次分类结果与真实标签做比对,从而利用CLR聚类算法将易分错、易混淆的类别进行聚类,再根据聚类后的结果将易混淆的类进行单独的再训练和再分类,取得了较好的效果。本文在CUB-200-2011、FGVC-Aircraft-2013b和Stanford-cars三个数据集上进行了实验,分类准确率分别由B-CNN模型的84.35%,83.56%,89.45%提高到了84.67%,84.11%,89.78%,验证了本文算法的有效性。另外,本文也做了一些其它实验,得到了一些结论,并提出了一些接下来可以继续改进的点。
[Abstract]:In recent years, with the growth of big data, the task of image classification has developed rapidly. The task of classification has also evolved from two categories to simple coarse classification and then to fine-grained classification. In most cases, the coarse classification is to separate different objects, such as cats, dogs, trees and cars. There are significant differences between them, which are easier to categorize, and fine-grained classification is about subdividing the same object, such as different types of birds, different types of aircraft, face recognition, and so on. The difference between them is usually very small. And their camera angle, illumination and occlusion all cause more trouble to classification, so it's more difficult. With the development of in-depth learning, the task of image classification becomes more and more easy, and people's requirements for image classification are also increasing. The task of fine-grained classification emerges as the times require, including the species of flowers, bird species, face recognition, etc. The task of image classification based on machine learning consists of two parts: image feature extraction, classification, and so on. The quality of feature extraction is directly related to the classification accuracy. The larger the feature dimension, the higher the accuracy. However, when the feature dimension is too high, it will lead to the problems of more memory and large computation. Convolution neural network is often used to extract features. Generally speaking, the deeper the network, the better the feature, as the number of layers deepens. However, when the network reaches a certain depth, it becomes more and more difficult to improve the accuracy by increasing the depth. For example, component-based modeling, but its disadvantages are low efficiency and heavy manual marking, while B-CNN model solves these two problems at the same time. It only needs the tag of training set samples, and does not need the boundary box of the target and the tagging of the parts. At the same time, there is a high accuracy. B-CNN model can solve the problem of fine-grained classification to some extent, but its training and classification are aimed at all samples, and some confusing classes can not be distinguished. On the basis of B-CNN model, this paper makes different improvement attempts, among which, according to the first classification result obtained by B-CNN model and the real label, the CLR clustering algorithm is used to cluster the easily error-prone and confusing categories. According to the clustering results, the confusing classes are retrained and reclassified separately, and good results are obtained. Experiments are carried out on the CUB-200-2011FGVC-Aircraft-2013b and Stanford-cars data sets. The classification accuracy is increased from 84.35% 83.56% of the B-CNN model to 84.67% 84.11% and 89.78% respectively, which verifies the validity of this algorithm. In addition, some other experiments are done, some conclusions are obtained, and some points that can be further improved are put forward.
【学位授予单位】：中国地质大学(北京)
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】