【作者】 贾世杰；

【导师】 孔祥维；

【作者基本信息】 大连理工大学 ， 信号与信息处理， 2013， 博士

【摘要】 基于内容的图像分类是指根据图像的视觉特征对图像进行自动语义分类,需要克服类内差异、遮挡、姿态变化和背景干扰对分类带来的不利影响,目前是计算机视觉领域最具挑战的课题之一。在电子商务领域,基于内容的商品图像自动分类能够为交易双方提供快速商品查询、确定商品的置放策略及进行用户感兴趣商品的智能推荐,从而有力提高电子商务市场的整体效能,是电子商务智能化的迫切要求。本文主要使用判别式分类模型研究基于内容的商品图像分类方法,具体工作如下：(1)为实现按照某种感兴趣信息(如女士皮鞋是圆头还是尖头,T恤衫是圆领口还是V型领口等)或商品类型对在线商品进行快速自动分类,研究了基于图像类描述与图像-类最近邻分类方式的商品图像分类方法。这种方法对每一个商品图像类建立类统计描述模型,在特征空间计算测试图像与每一类统计模型(类描述)的距离,将距离最小的图像类作为最终的分类结果。具体从两个方面构建商品图像类描述,实现图像-类最近邻分类。①全局特征法。采用具有互补特性的塔式梯度方向直方图和塔式关键词直方图全局特征构造商品图像基于特征分布参数的类描述符和基于特征分级匹配的类描述符；然后通过计算测试图像描述符与各类类描述符之间的距离(图像-类距离)实现商品自动分类。计算过程简单,分类性能比现有相关文献有一定提升。②局部特征法。为克服构建全局特征过程中量化误差的影响,局部特征法将商品图像及商品图像类看做是若干独立同分布局部特征的无序集合,采用图像-类最近邻方式实现商品图像分类。为快速实现图像-类距离的计算,本文在对每类的局部特征描述子进行多级聚类,通过设定聚类级数和类过滤比例能够灵活平衡分类正确率与分类速度。(2)构建图像类描述需要较大数量的已标记样本。针对已标记(训练)样本数量较少的情况,本文采用基于数据驱动的核函数构建方法,在词包(Bag Of Words, BOW)模型的基础上,设计了一种基于加权二次卡方(Weighted Quadratic Chisquared, WQC)巨离的直方图核函数,使用具有核技巧的支持向量机进行商品图像分类。对于训练样本较少情况下的图像分类,基于WQC直方图核函数方法有着较明显优势。(3)考虑到商品图像分类具有类别数量多、类内变化大、分类依据多样等复杂性,研究了多特征联合方法以提高商品图像分类性能。①多核联合方法。为避免传统多核学习中繁琐而困难的联合优化问题,提出了基于(去中心化)核经验校准的商品图像分类方法；②多分类器联合方法。本文建立了基于异构强分类器决策层联合的商品图像分框架,提出了基于支持向量机二级级联的商品图像分类算法。本文所提出的两种多特征联合方法能充分利用特征的互补特性,比传统多特征联合方法更能有效地提高商品图像分类性能。更多还原

【Abstract】 The aim of content-based image classification is to implement semantic classification automatically based on the visual features. However, some adverse effects, such as the within-class variation, obscure, pose variation, and background interference, are hard to overcome. Therefore it is still a challenging problem in the field of computer vision. Automatic product image classification can effectively improve the overall effectiveness of the E-commerce market, such as quick product querying, determining the placement strategy and conducting product intelligent recommendation. Consequently, it is a critical requirement of E-commerce intelligent. This dissertation focuses on the content-based product image classification with discriminative classification model. The main research work is as follows:(1) For the service of automatic real-time online product classification with some specific information of interest (such as round or pointed of lady shoes, round or V-neckline of T-shirts, etc.) or product categories, the product classification schemes are developed based on the class-specific descriptors and image-to-class nearest neighbor classifiers, in which each product image category is modeled statistically, the category nearest to the query product image in the feature space is chosen as the final classification result. Two kinds of approaches are proposed for class descriptor construction and image-to-class nearest neighbor classification:①Global feature based schemes. With two global complimentary features PHOG (Pyramid Histogram Of Gradient) and PHOW (Pyramid Histogram Of visual Words), CDDP (Class-specific Descriptor with Distribution Parameter) scheme and CDHFM (Class-specific Descriptor with Hierarchical Feature Matching) scheme are constructed, respectively. The image-to-class distances are calculated between the descriptors of the test product image and each class-specific descriptor for automatic product image classification. The procedure is simple and the classification performances are prior to the relative literature.②Local feature based scheme. In this scheme, all the product images and image classes are regarded as orderless sets of local descriptors and image-to-class nearest neighbour classifier is employed for product image classification. Local feature descriptors of each category are hierarchically clustered to speed the calculation of image-to-class distances, and the trade-off between classification accuracy and speed can be achieved flexibly through the set of clustering level numbers and the class filter ratio. (2) To construct the class-specific descriptor, the labeled samples are required to be in a quantity sufficient for good performance. In the case of product classification with a small number of labeled samples, data-driven kernel building methods are explored and a Weighted Quadratic Chi-squared (WQC) histogram kernel function is designed to combine with BOW (Bag Of Word) model. With the kernel based support vector machines, the proposed histogram kernel function offers superior performances with small training samples.(3) Taking into account the complexity of product image classification, such as the big number of categories, large within-class variation, multiple classification bases, multiple features combination methods are designed to boost classification performances.①Multiple kernels combination. To avoid the tedious and difficult joint optimization process, a (decentralized) kernel empirical aligment based scheme is proposed.②Multiple classifier combination. A framework is built with decision-level fusion of heterogeneous strong classifiers, and a scheme of two-layer SVM classifiers cascading is proposed for product image classification. The proposed multiple kernel and multi-classifier combination methods can take the advantage of the complementary features, and perform much better than the traditional combination methods for product image classification.更多还原