The Study Of Discriminative Dictionary Learning Algorithms For Robust Classification

Sparse representation and classification based on dictionary learning is an important issue in the area of data mining,signal processing and pattern recognition,and has been successfully applied in image recognition and computer vision.Note that most of the existing dictionary learning methods suffer from some common drawbacks,i.e.,these methods are sensitive to noise;the classification phase of these methods is time-consuming;the inter-class discrimination of these methods should be enhanced;these methods have a limited supervised information;and these methods have a low scalability caused by the complex training phase.To address these issues,we in this dissertation propose three novel discriminative dictionary learning frameworks.The simulations on face recognition,object recognition and machine fault classification will show the effectiveness of our proposed techniques.The major contributions are summarized as follows:?1?A semi-supervised label consistent dictionary learning algorithm is proposed.By incorporating the semi-supervised learning into dictionary learning framework to use both labeled and unlabeled data for discriminative dictionary learning,and using the achieved discriminative sparse codes as adaptive reconstruction weights for predicting the labels of unlabeled data,the discrimination and robustness can be enhanced;By using the multi-class classifier which is trained based on original samples to classify each testing sample,the extra sparse reconstruction process which is widely-used in most of the existing dictionary learning method can be avoided,the efficiency can therefore be improved;?2?A latent label consistent dictionary learning algorithm is proposed.By incorporating a latent vector and a structured label consistent error term to reduce the disturbance between inter-class atoms and similarity among within-class sparse codes,the discrimination can be improved;By fully considering the structure of data and decomposing given data into a sparse reconstruction part,a salient feature part and an error part,the robustness can be enhanced;By using the multi-class classifier which is trained based on the salient feature part to efficiently classify each testing data,the extra sparse reconstruction process can be avoided,and the efficiency can be improved;?3?An analysis discriminative dictionary learning method is proposed.We propose an analysis discriminative dictionary learning approach.The proposed analytical incoherence promoting term can enhance the discrimination of the learned dictionary and enhance the robustness of the proposed method.A robust and sparse L_2,1-norm is adopted to replace the time-consuming L₀ or L₁-norm regularization,note that the L_2,1-norm is easier to optimize,so the efficiency of training can be improved.An analytical sparse codes extraction term and an analytical classification term are also proposed for classification,thus each testing data can be classified by its approximate coding coefficient,and the efficiency of classification phase can be improved.