Research On Discriminative Dictionary Learning Method For Image Classification

With the rapid development of artificial intelligence and the widespread use of intelligent terminals,image data has grown rapidly.Compared with other information forms,the information carried by image data is more vivid.Therefore,image data,as an important interaction medium between humans and computers,has become the main information carrier of the internet and is widely used in various fields.In this age of massive images,how to simplify data expression to effectively obtain image information has attracted more and more researchers to explore the field of image classification.Inspired by the sparse coding mechanism of the human visual system,sparse coding that represents signals as sparse linear combinations of dictionary bases(that is,dictionary atoms)has been successfully applied to image processing,visual recognition,and machine learning.Given a set of signals,sparse coding aims to find a suitable dictionary so that each input signal can be well approximated by linearly combining dictionary atoms.Therefore,it is particularly important to learn a suitable dictionary from the training data,and the study of dictionary learning,which can be used for classification,has also become a research focus and a difficult point in the field of image classification in recent years.Based on dictionary learning and sparse coding,this thesis proposes a supervised dictionary learning method called class-oriented discriminative dictionary learning(CODDL).The purpose of this method is to use the structure information and category information of the sample to learn a discriminative dictionary and use it for image classification.The method emphasizes the class discrimination ability of dictionary atoms and sparse coding coefficients,and uses the category information of the samples to constrain the dictionary and coding coefficients,thereby enhancing the discrimination between the learned dictionary and coding coefficients.At the same time,the criterion of maintaining sparsity is adopted in the encoding process,and a sparse constraint term of quadratic form is introduced to simplify the objective function.With a comprehensive consideration of multiple optimization objectives and an adjusted sparsity constraint term,the objective function is concise and easy to solve.The adopted suitable optimization strategies also simplify the process of algorithm solving.What's more,with the discriminative representation coefficients,a simple and efficient classification scheme is designed for image classification.The algorithm proposed in this thesis has undergone extensive experiments on multiple classification tasks,and has achieved the best classification accuracy.In addition,this thesis conducts a comprehensive analytical experiment of the algorithm to verify the effectiveness of the algorithm.Finally,the effectiveness of the algorithm is summarized.