Noise Learning Based Dictionary Learning Algorithm For Image Classification

With the background of increasing amount of data and the dimension of data samples,dictionary learning has become one of the most popular topic in the field of machine learning and computer vision in recent years,due to high complexity of signal processing but low-dimensional corresponding representatives of these signals.The basic theory of dictionary learning applied to image signal processing is that the natural images have corresonding sparse representations,therefore,dictionary learning methods utilize a set of over-completed bases(dictionary)to estimate the linearly representation of image signals,where the corresponding representation can be obtained under some specific conditions of sparsity with accurate mapping to the original signal.To some extent,the general mathematical optimization model of dictionary learning is acquired from the perspective of signal recovery/reconstruction.Besides,previous researches of dictionary learning are generalized to many applications due to the robustness of algorithms,such as signal recovery/reconstruction,unsupervised clustering,regression,recognition,etc.Some state of the art in classification researches also exploit dictionary learning frameworks as an essential section,where the performances of classication tasks are well-determined by the learned dictionary.Generally,there are two methods to learn a good dictionary.The first is a predetermined/given analysis dictionary,such as wavelet basis,discrete cosine transform(DCT),etc;the second is to learn a corresponding dictionary from the given training data.It can be demonstrated that,a learned dictionary from training data can greatly improve the recognition performances in a specific application compared to a given dictionary,especially for the supervised dictionary learning of image classification tasks.With the development of general dictionary learning models,the linear combination of fidelity terms and regularization terms is usually designed for optimization of the corrsponding representations and the learned dictionary.The fidelity term is utilized to map the value and distribution of the reconstruction residuals,and the regularization term provides effective and efficient prior knowledge for the distribution of signals or other data characteristics.In the proposed methods,we assume that there is additive white Gaussian noise accompanied with general image signals in training datasets,therefore,the fidelity term adopts the-norm as the constraint;also,we exploit the group regularization method as the regularization term,which can obtain a more accurate optimal solution and more efficient framework due to the effective prior knowledge.Group regularization combines a regularization term of-norm for mapping the corresponding representation distribution,with a cross-label suppression term to enlarge the difference among the corresponding representations which belong to different labels,as well as a mathematical operator of Laplacian matrix named N-cut in spectral clustering to shorten the difference among the corresponding representations of the same label.However,in the classification tasks of high-demsional or complex-distributed image signals,dictionary shows its limitation.Therefore,this paper proposes the consideration of the learned noise signal via the framework of Alternating Direction Method of Multipliers to separate dictionary and nosie,with the design of two innovative corresponding denoising classifiers for better performances.Compared to some classic previous work,the proposed algorithm has an obvious improvement in the task of classification based on five natural image signal datasets in the experiments.Although the previous work considered that noise in datasets may consist of some information of labels,the experimental results of this paper demonstrate that learning a dictionary without noise is probably to have a better recognition performance in image classification than a dictionary containing noise.