Research On Discriminative Dictionary Learning For Image Classification

Image classification is an important research field of computer vision and pattern recognition.With the rapid development of mobile Internet and computer technology,image classification has been widely used in information security,medical treatment,industry and other fields.The theory and application of sparse signal representation have made a rapid development in the past years.It approximates the signal by a small number of atoms in an over-complete dictionary through a linear representation method,so as to realize the simple and effective representation of signals in order to facilitate the follow-up work.Dictionary learning aims to find a dictionary where signals in some ensemble have sparse representations.In image classification field,to enable discrimination for classification applications,the dictionary is usually designed to force signals of the same class to share a common support,e.g.,by allocating a unique sub-dictionary for each class.However,the common support requirement is too ideal and restrictive in view of the fact that in many classification/recognition applications,signals of the same class often have both common features and innovation features.In this paper,a new discriminative dictionary learning framework for classification,where a flexible dirty sparse model is employed.In contrast to the sub-dictionary based approaches that fails to scale for a large number of classes,the proposed approach considers a dictionary shared by all classes and thus is scalable with the number of classes.The dictionary learning framework is formulated into an optimization problem with designed regular terms to promote both the dirty structure and discrimination capability.An efficient and effective iterative algorithm based on the alternating direction method of multipliers(ADMM)is provided to solve the proposed optimization problem.The superior performance of the proposed approach is demonstrated in comparison with state-of-the-art dictionary learning based approaches for classification by conducting extensive experiments on various image classification tasks.