项目名称: 基于区分型码本的图像表示的研究与应用
项目编号: No.61503145
项目类型: 青年科学基金项目
立项/批准年度: 2016
项目学科: 其他
项目作者: 王兴刚
作者单位: 华中科技大学
项目金额: 22万元
中文摘要: 基于特征集的图像表示是计算机视觉中十分核心且具有挑战性的问题,在诸多应用中发挥着重要作用。本项目拟提出一种新颖的基于区分型码本的图像表示方法,它具有三个主要优点:(1)采用了一种新的数据空间划分方式,使得图像表示更加简洁;(2)区分型码本自动学习样本各维度的权重,可以有效融合不同种类的特征;(3)能够在码本中融入人工标注的高层语义,使得图像表示更加高效。本课题研究焦点在于区分型码本的性质、学习方法、编码方法、图像表示、以及解决大规模数据下的图像分类、物体检测等应用问题。本课题中的研究有助于解决区分型聚类、弱监督学习、物体识别等机器学习、计算机视觉领域中的任务。另外,区分型码本是一种通用的码本表示方法,可以应用于文本、音频等多媒体数据的表示,并推动相关领域的研究发展。
中文关键词: 码本学习;区分型学习;图像表示;物体识别
英文摘要: Feature set based image representation is an important yet challenging problem in computer vision and machine learning, which plays a critical role in a wide spectrum of applications. In this project, we propose a novel image representation method based on discriminative codebook. Comparing to the state-of-the-art data representation methods using generative codebook, main advantages of discriminative codebook lie in the following three aspects: (1) it uses a new space partition strategy, which makes our codebook more compact; (2) it learns weights for each dimension of data, which helps to fuse different kinds of features; and (3) it contains high-level semantics. We focus on studying the characteristics of discriminative codebook, discriminative codebook learning approaches, coding approaches of discriminative codebook, image representation using discriminative codebook, and how to apply the proposed image representation for large-scale image classification, object detection etc. The research in this project helps to solve the fundamental problems in machine learning and computer vision, such as discriminative clustering, weakly supervised learning and object recognition. Besides, the flexibility of our discriminative codebook learning method makes it possible to be widely applied to text, audio and other types of data, and thus promotes the development in the related areas.
英文关键词: Codebook Learning;Discriminative Learning;Image Representation;Object Recognition