一种基于加速坐标下降的大规模图像分类算法研究  

A RESEARCH ON A LARGE-SCALE IMAGE CLASSIFICATION ALGORITHM BASED ON ACCELERATED COORDINATE DESCENT

作　　者：王智勇[1] 

机构地区：[1]四川商务职业学院,四川成都610113

出　　处：《计算机应用与软件》2014年第4期208-213,共6页Computer Applications and Software

基　　金：国家自然科学基金项目(61203254;F030403)

摘　　要：随着大规模图像分类数据集的出现,设计一种可扩展的、高效的多类分类算法成为目前一个重要的挑战。基于迹范数正则惩罚函数,提出一种新的大规模多类图像分类的可扩展学习算法。把具有挑战性的非光滑优化问题重构为一个带l1正则惩罚的无穷维优化问题,进而设计一个简单而有效的加速坐标下降算法。此外,展示了如何在量化的密集视觉特征的压缩域中进行高效的矩阵计算,该压缩域有100 000个例子,1 000多维特征和100多类图片。最后在图像网的子集"Fungeus"、"Ungulate"和"Vehicles"上的实验结果表明,所提出方法的性能明显优于目前最先进的16高斯Fisher向量方法。With the emergence of large-scale image classification datasets,to design a scalable and efficient multi-class classification algorithms becomes an important challenge now. We introduce a new scalable learning algorithm for large-scale multi-class image classification,which is based on the trace-norm regularisation penalty. We reframe the challenging non-smooth optimisation problem into an infinite-dimensional optimisation problem with regular l1regularisation penalty,and further design a simple and efficient accelerated coordinate descent algorithm. Moreover,in the paper we show how to perform efficient matrix computation in compressed domain of quantified dense visual features,the domain scales up to 100,000 examples,over 1,000-dimensional features,and more than 100 categories. Promising experimental results on＂Fungus＂,＂Ungulate＂,and ＂Vehicles＂subsets of ImageNet are presented,where we show that our approach performs significantly better than the state-of-the-art approach of Fisher vectors with 16 Gaussians.

关 键 词：大规模图像 多类分类算法 L1范数 压缩域 坐标下降算法 Fisher向量 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]