Research On Predictive Discrimination Dictionary Learning Network

Recently,with the great development of artificial intelligence,image recognition technology has gradually penetrated daily life and brings great convenience,which is one of the most important applications in the field of artificial intelligence.Classification based on dictionary learning(DL)is a sustained attention in the image recognition areas.DL is important representation learning method.Most previous studies on dictionary employ generative model to reconstruct an input image with a sparse linear combination of atoms.Moreover,sparse codes and overcomplete dictionary of the DL model are also outstanding features for classification tasks,which brings classification based on DL methods into sight of researchers.Classification of images based on DL relies on a well-learned dictionary in which atoms could exhibit many interesting and representative patterns.Many DL based classification algorithms have achieved remarkable performance.However,due to the sparse coding stage is always computationally expensive,few of them has been widely employed in problems of large-scale.This thesis presents a novel discriminative dictionary learning network using both generative and feed-forward neural network(FNN)models to efficiently learn a dictionary with both discriminative and descriptive atoms.High flexibility of our model allows fast inference of the label of a test sample and speeding up solving a sparse representation.More importantly,the learned atoms not only illustrate interesting properties,but also are very discriminative,which results in competitive performance in the classification tasks.Moreover,a novel dreaming phase are introduced in this thesis to improve the robustness of the model for unknown pattern in known class.Many experiments verified effectiveness of our method.The highlights and main contributions of this thesis are as follows:1.The proposed discriminative dictionary learning network is novel framework for DL based image classification by attaching a co-trained FNN to DL based classification algorithms,which is used to predict discriminative sparse codes and improve the computation efficiency.2.The proposed discriminative dictionary learning network has a design of a new structure of dictionary in this thesis,which consists of descriptive atoms and label specific atoms.Based on the dictionary structure,labels of samples are integrated into sparse codes.Without extra classifiers or computation of reconstruction error,which are used by previous classification based on DL method,the label of each sample can be read from sparse codes directly.3.Two implements of discriminative dictionary learning network framework are given in this thesis,which are based on L0 norm and L1 norm constraint,respectively.4.This thesis provides detailed analysis of the problem of unknown pattern in know class and develop a new training process referred to dreaming for the problem.At the same time,it presents implementation on the proposed DDLN models.Dreaming improves the prediction ability of FNN for unknown pattern in know class with virtual samples generated by reasonable code and the dictionary.