Image Classification Based On Multi-layer Laplacian Sparse Coding

Image classification automatically classifies images according to their features.Affected by illumination,scaling and some other factors,how to extract image features and represent images effectively has become an attractive issue of image classification.Sparse coding is an efficient feature representation method,it finds a set of over-complete dictionary bases and then reconstracts images based on these dictionary bases,which has achieved good performance in image classification.However,there are still some drawbacks in the research of sparse coding,like the model ignores the group effect between dictionary bases,the Euclidean distance is not good enough to measure the distance between features and dictionary bases,and the Laplacian regularization ignores the spatial topology information of the features,which results in lacking of extrapolation,and so on.On another aspect,compared with the methods based on deep learning,a single-layer model has some limitations in representing the features,and it is difficult to find out discriminative features of data.Therefore,this dissertation mainly focuses on sparse coding and deep learning,and proposes corresponding methods to solve the above-mentioned problems.A series of comprehensive experiments on the standard datasets have been performed.The main work and innovations of this dissertation are as follows:1.Dealing with problem that ₁l-norm only considers sparsity and ignores group effect of the coding,this dissertation proposes a sparse coding method based on elastic net and histogram intersection.l₂-norm is introduced in the optimization to obtain the sparsity which is working similarly to ₁l-norm,and consideres group effect as well.By introducing the histogram intersection,the distance between features and dictionary bases is redefined to obtain more efficient image coding.2.In order to solve the problem that Laplacian regularization neglects the spatial topological structure information of features,which results in weak generalization of the model,this dissertation proposes a non-negative local sparse coding method with Hessian regularization and ₂l-norm.Hessian regularization makes better use of the topological structure information,and describes the intrinsic geometric characteristics of data more accurately,thus Hessian regularization can better predict the data that are outside the sample area.In addition,the method is extended with a non-negative local constraint,which further improves the stability and consistency of the coding.The algorithm is applied to image classification of standard datasets,and obtains excellent classification results,which verifies the superiority of the algorithm.3.Focusing on the issue of limitation of feature learning ability from single-layer structure,and combining with the superior feature learning ability of deep learning,this dissertation proposes a two-layer Laplacian sparse coding algorithm.In the first and second levels,non-negative dictionary learning and sparse coding are performed respectively to obtain sparse representation at image block level and image level.This algorithm is capable of capturing the deeper feature information of images,thus has efficient ability in feature representation.Validated on four standard datasets,and compared with the representation ability of single-layer sparse coding,the classification performance of the two-layer coding model has been further improved.