基于子编码和全编码联合惩罚的稀疏表示字典学习方法

Sub-coding and Entire-coding Jointly Penalty Based Sparse Representation Dictionary Learning

针对目前稀疏表示字典学习的惩罚函数版本不一且各有优势的问题,提出基于子编码和全编码联合惩罚的稀疏表示字典学习方法,该方法在字典学习的目标函数中同时加入子编码惩罚函数和全编码惩罚函数。子编码惩罚函数使得学习后的字典在稀疏表示识别时可以用子字典的重构误差和子字典上编码系数的大小来识别,全编码惩罚函数则能直接利用整个字典上的编码系数来识别,通过联合这两个惩罚函数可以获得非常好的识别效果。为了验证所提方法的有效性,在语音情感库和人脸库上与最新的基于字典学习的稀疏表示识别方法 DKSVD和FDDL进行对比,并与著名的识别方法SVM和SRC进行比较,实验结果显示所提方法具有更好的识别性能。

Currently, the penalty function of dictionary learning （DL） used for sparse representation classification has many versions and each of them has its own advantages. This paper presented a new dictionary learning method called Sub-coding and Entire-coding jointly penalty based dictionary learning, which jointly adds sub-coding penalty functions and entire-coding penalty functions into the dictionary learning objective function. Sub-coding penalty function makes the dictionary after learning can use its reconstruction error and sub-coding for classification, and entire-coding penalty function makes the dictionary after learning can directly use its whole coding for classification at the same time. By combining these two penalty function, good recognition effect can be got. The proposed method is extensively evaluated on emotion speech database and face database in comparison with famous DL based sparse representation classification methods DKSVD and FDDL, and other famous recognition method SRC and SVM. The experimental results show that proposed method has better recognition performance.