The Dictionary Pair Learning Combined With Graph Embedding For Image Recognition

Since high-dimensional image contain lots of complex information and features,how to effectively dig the internal rules from high-dimensional complex image data and perform efficient analysis and identification has always been a basic problem in the field of computer science.In recent years,sparse representation and dictionary learning have become a research hotspot in the field of image recognition and have been rapidly developed.However,the feature information of high-dimensional image is often non-linear,and this non-linear structure often contains wealth discriminative information which is benefit for enhancing recognition rate.The intrinsic geometry structure of image is often ignored in dictionary learning model,which lead to the performance is not so good.The manifold learning method based on graph embedding framework can effectively mine the low-dimensional sub-manifolds embedded in the high-dimensional data,and reveal the inherent structure and rules hidden in the highdimensional data.In this paper,the image recognition algorithm and its application based on graph embedding theory and dictionary learning theory are studied.The main work is summarized as follows:(1)A dictionary pair learning(DPL)model with graph embedding constraints is proposed to learn a pair of synthesis dictionaries and projection dictionaries simultaneously.In the model,considering that the local geometric structure of the image often contains key discriminant information.The critical step of our method is to construct the graph term,inspired by the relationship between atoms and profiles,the Laplacian matrix is constructed by atoms.This is distinctive from the majority of existing approaches that the Laplacian matrix constructed with training data.By using graph Laplacian as a smooth operator,the local geometric structure of the learned dictionary is preserved.Specifically,it further improves the discriminability of the dictionary.Experimental results on several benchmark data sets have demonstrated the competitive performance of the proposed method for image classification.(2)A cross-view multi-level discriminative dictionary learning model was proposed using the intrinsic correlation of coding coefficients in the feature representations of different views,and applied to pedestrian re-identification task.First,we introduce a feature mapping function in the dictionary learning model of the image horizontal region-level and image-level.This mapping function describes the latent relationship between the coding coefficients of the same person image in different cameras.It can improve the flexibility of feature representation.Then,on the patch-image level,we add a graph embedding constraint on atoms into the dictionary learning objective function by combing the local manifold structure of the image.By learning a graph Laplace matrix adaptively,coding coefficients can preserve the similar local geometry structure with training samples and more discriminant dictionary pairs can be obtained.Experiments on the two challenging person re-identification datasets demonstrate the proposed method can reduce the influence of large variation of resolution in the different cameras and improve the representative and discriminative abilities of learned dictionaries.Compared with the state-of the-art algorithms,the proposed method can improve the performance of person reidentification.