局部保持“字典对”学习算法及其应用

Locality Preserving Dictionary Pair Learning Algorithm for Classification and Recognition

字典学习(DL)方法近年来被广泛应用于解决各种计算机视觉领域的问题.现有的大部分字典学习算法均旨在学习一个综合型字典来表示输入信号,并使表示系数或表示误差具有一定的判别能力.这些字典学习算法大都需要对稀疏表示系数采用l_0或者l_1范数的约束,所以学习过程比较耗时。解析型字典学习的提出较为有效地解决了字典学习算法效率低的问题.在分类识别任务中,联合学习一个综合型字典和一个解析型字典正在成为一个热门的研究趋势,这不仅很大程度上降低了学习过程中的计算复杂度,而且在分类识别性能上也能有一定的提升。借鉴了最新提出的"字典对"学习思想,利用训练数据的局部结构信息,提出了局部保持的综合型-解析型"字典对"学习算法.在3个国际公开测试数据库(人脸识别库Extended YaleB、AR和图像分类库Caltech101)上的实验结果表明,局部保持的综合型-解析型"字典对"学习算法在准确率和效率方面都具有很好的性能.

Dictionary learning （DL） has been widely applied in various computer vision problems. Most existing DL methods aim to learn a synthesis dictionary to represent the input signal while enforcing the representation coefficients and/or representation residuals to be discriminative. However, the l〇 or ^-norm sparsity constraint on the representation coefficients adopted in most DL methods makes the learning phase time consuming. Hence,a new DL framework, namely analysis dictionary learning, has been proposed to deal with this problem. Besides, jointly learning a synthesis dictionary and an analysis dictionary has also drawn much attention recently, which can not only reduce the time complexity in the learning phase, but also lead to very competitive accuracy in a variety of classification and recognition tasks. This paper is motivated by the recently proposed dictionary pair learning algorithms. Our proposed novel Locality Preserving Discriminative Synthesis- Analysis Dictionary Pair Learning Algorithm integrates the local structure of training data〉 which is totally ignored in most applications of dictionary pair learning. We evaluate the proposed algorithm on three databases 〉 including two face databases （Extended YaleB and AR） and one image categorization database （CaltechlOl）. Based on the experiment results〉 we conclude that our algorithm can largely reduce the computational burden and improve the accuracies.