Research On Image Feature Extraction Based On Sparse Representation

Computer vision,machine learning,image processing and other fields will involve classification problems.The classification is to divide similar objects into same group,and divide dissimilar objects into different groups.The classification problem is usually faced with high-dimensional data,which will lead to the problems of high memory consumption and long computation time,and the algorithm effectiveness will be reduced due to the "dimension disaster".To solve this problem,sparse representation can be utilized.The sparse representation is to find a suitable dictionary for ordinary densely expressed samples,and convert the samples into a suitable sparse expression form,thereby simplifying the learning task and reducing the complexity of the model.This thesis proposes an image clustering algorithm based on sparse representation,the sparse convolution subspace clustering,and the geometrical sparse representation on single image which can be applied to image classification.The details are as follows.Firstly,this thesis propose an image clustering algorithm based on spectral clustering,the sparse convolution subspace clustering.This method is inspired by sparse subspace clustering,that obtains the sparse representation of each image by using all the images to be clustered to compose a dictionary.Considering that sparse subspace clustering will expand the image into column vectors when solving the sparse representation of the image,which destroys the internal structure of the image.In order to preserve the internal structure of the image,the convolution combination is proposed to replace the linear combination in the sparse subspace clustering,because the 2D convolution has local weighting effect.The convolution combination is the sum of a series of convolution of each specified image with a convolution kernel in specified size.Taking the group sparse as the optimization goal,a sparse convolution self-representation problem is constructed,which can be solved by the alternating direction multiplier method,and obtain the partition of data set by applying the spectral clustering to the obtained result.In fact,the sparse subspace clustering is a special case of sparse convolution subspace clustering,because when only the center position of the convolution kernel in the sparse convolution subspace clustering can take value,the sparse convolution subspace clustering is equivalent to sparse subspace clustering.Experimental results show that sparse convolution subspace clustering can be effectively applied to face clustering,and achieves good clustering results on different scale data sets.Then this thesis propose the geometrical sparse representation on single image.This method is based on the local linear model between the input image I and the dictionary image D,and introduces regular terms to ensure the sparsity of the coefficients.Theoretically,this method is effective in preserving the high-frequency information in the image.In order to ensure the robustness of the method,the LBP method is used to encode the coefficients obtained from the local linear model to obtain the final sparse representation binary code of the image.Experimental results show that this method can be effectively used in face classification and shows strong robustness on different occlusion degrees.