Dictionary Learning Method Driven By Small Training Samples For Sparse Redundant Representation

Sparse representation of signals has been a hot topic in recent years.This paper mainly studies the learning algorithm of KSVD dictionary for sparse representation.Sparse representation is the use of atomic linear approximation signals in a few overcomplete dictionaries,and the better the ability to select the appropriate dictionary representation,while the dictionary trained and learned by KSVD algorithm is adaptive and highly matches the original signal,which effectively improves the sparse representation effect of signals.Based on the study of sparse representation and KSVD dictionary learning,this paper proposes a KSVD dictionary generation algorithm based on small samples in view of the problem that the training dictionary with large samples consumes a long time in KSVD algorithm.It is found that the training dictionary with small samples will not reduce the quality of the generated dictionary,but can effectively reduce the algorithm time.At the same time,aiming at the redundancy of over-complete dictionary,this paper proposes a method to eliminate invalid atoms in the dictionary,improve the utilization rate of atoms in the dictionary,reduce the footprint of the algorithm and improve the efficiency of the algorithm.In chapter 4,this paper uses an improved method to experiment with image denoising.Compared with the traditional method based on KSVD dictionary learning,the sparse representation of KSVD dictionary learning based on small samples is feasible in this paper.Compared to traditional sparse representation of the effect,the method under the condition of the guarantee not reduce the effectiveness of said can greatly reduce the computing time,at the same time,this paper puts forward the minimization of dictionary atomic number,eliminate invalid dictionary atoms,can generate small dictionary of storage requirements is not high,although the dictionary atoms reduced,but will not reduce the effectiveness of the representation of sparse representation.