标题:Graph-Regularized Discriminative Analysis-Synthesis Dictionary Pair Learning for Image Classification
年份:2019
Corresponding author: Heyou Chang (cv_hychang@126.com)
problem:分析-综合字典对学习在图像去噪上个取得了良好的性能,但用于图像分类的字典对学习研究还不充分。
proposed method:graph-regularized discriminative analysis-synthesis dictionary pair learning (GDASDL),利用图约束保留了数据的局部几何结构。判别性:将标签信息与字典原子相关联来引入全局信息。信号的几何结构和鉴别结构对图像分类任务是必不可少的。
通过训练合成字典来很好地表示图像,并利用分析字典将图像的结构信息从原始图像空间转换到更具鉴别性的空间。
Related Work:
合成字典学习:ADL
分析字典学习:SDL
分析合成字典学习:ASDL
判别性(利用标签信息)分析合成字典学习:DASDL,第二项为分类误差项,将第i类稀疏投影到标签空间的第i维,P:线性投影矩阵,Y:标记矩阵
proposed method:GDASDL,捕获数据间的局部几何结构
The last two terms represent the discrimination constraint term and the graph-regularized term, which are used to improve the classification ability of the proposed model. P is a linear transformation. The discrimination constraint term enforces the signals from the same class to have similar representations and those from different classes to have dissimilar representations.
OPTIMIZATION:
为了简化:
有:
为了有效地求解:采用逐个原子更新策略。在第m步,除了第m对原子外,其他原子都保持固定。然后,我们可以分离出对第m个原子对的依赖关系