A Study On Discriminative Dictionary Learning For Sparse Image Representation And Classification

Dictionary learning aims at obtaining an over-complete dictionary D via optimization and then each input sample can be reconstructed by using the highly-relevant atoms in D to sparsely represent data,which can be utilized for image representation and classification.Dictionary learning has been attracting considerable attention in recent years due to its strong representation and classification ability.However,most existing discriminative dictionary learning models still suffer from the following shortcomings.First,the procedure of seeking the dictionary D and sparse codes S is not adaptive to different datasets,which also cannot explicitly establish the interconnections between the dictionary atoms and sparse coefficients.Second,the error reconstruction procedure does not conside extracting salient latent features,which ignores the distinguishing salient features and hence may lead to the degraded classification ability.Third,most existing methods utilize the shallow dictionary learning mechanism,which cannot seamlessly integrate the deep neural networks with the end-to-end deep learning model.These will lead to insufficient learning of the representations and directly affect the classification accuracy.To address the aforementioned shortcomings,in this thesis,we will propose three innovative solutions and will also verify their effectiveness on real-world image datasets.The major contributions of this thisis are summarized as:(1)To obtain the neighborhood relationship between the sparse coefficients during the dictionary learning and sparse coding procedures,we propose a new learning framework called Discriminative Local Sparse Representations by Robust Adaptive Dictionary Pair Learning.The propose algorithm integrates the locality-adaptive sparse representations,robust projective dictionary pair learning and discriminative sparse coding into a unified model.The L2,1-norm is also used to make the representation learning process efficient and robust to noise.In addition,to improve the discriminating ability of the coding coefficients and atoms,a discriminating function is incorporated to ensure high intra-class compactness and inter-class separation in the sparse coding space.(2)To extract the salient latent features during the dictionary learning to improve the data representation and discrimination ability,we propose a new learning framework called Structured Twin-Incoherent Twin-Projective Latent Dictionary Pair Learning algorithm.Our model seamlessly integrates the twin-incoherence discriminative learning,salient features extaction and twin-projective structured dictionary pair learning are into a unified model.As a result,our formulation clearly unifies the abilities of salient feature extraction,representation and classification.In addation,the twin-incoherence constraint on the sparse codes and salient features can explicitly ensure the discrimination ability over them,i.e.,high intra-class compactness and inter-class separation.(3)Existig shallow dictionary learning methods have the limited representation learning ability,and how to design an end-to-end model that seamlessly integrates the shallow dictionary learning with deep learning is unclear.To this end,we propose a new framework called Deep Convolutional Dictionary Pair Learning Network.Our framework seamlessly integrates the convolutional neural network and dictionary pair learning are into a unified model,which is an end-to-end deep dictionary learning framework.As such,it can extract the hierarchical features and deep information from input data.Furthermore,the stochastic gradient descent algorithm is used for optimizing parameters of the network,which can accelerate the training of the convolutional dictionary pair learning network.(4)Extensive simulations on several image databases demonstrate the effectiveness of our proposed models for image representation,classification and recognition.