基于混合范式多任务学习的图像稀疏表示

受视觉表示和多任务学习的研究结果启发，发现传统模型约束项所获得的稀疏表示或过于冗余或过于严格要求信息共享，为寻找一种折中且更加有效的特征表示方法，提出基于混合范式多任务学习的图像稀疏表示学习框架。该框架以多特征的类别信息作为先决信息对特征进行组划分。选择L2，1和L1混合范式做约束惩罚函数约束，其中L2，1范式，在特征组内提取同种特征相关共享信息，L1范式在多特征组之间去相关，选择竞争性更强的特征种类。提出的学习框架不仅实现了多特征联合，而且充分考虑了不同特征之间的互补表示能力又消除了冗余。实验结果表明，由该框架学习得到的稀疏表示不仅可以达到稀疏要求，同时也实现了较好的分 类性能，证明了混合范式算法对提取图像关键本质信息的有效性。

Multi-Task Sparse Learning Based on Mixed Norm for Image Ｒepresentation

Motivated by the recent success of visual representation and the problem of conventional approaches which were too correlated or too strict constraint，this paper proposed a framework of decorrelating mixed norm multi-task sparse learning representation to seek an effective and intermediate feature representation approach．By leveraging side information about multiple features to group all kinds of features，L2，1and L1 mixed-norm regularizations were used as penalty function，for L2，1-norm's feature sharing information among a feature group and L1-norm's sparsity and competition among unrelated feature group． This framework simultaneously enforced multi-features joint and considered the complementation among decorrelated features． Extensive comparison experiments on data set show that the proposed method can obtained idea sparsity and is competitive to the stateof-the-art methods for image classification，which demonstrated the efficiency of L2，1 and L1 mixed-norm algorithm for extracting key information of an image．