An Optimized Method Of Image Multi-feature Sparse Representation Model

The essential problems in the field of computer vision is considered to be a"where and what",so an efficient representation of an image is one of the most basic research,also has important academic value and practical significance.Different types of images have different characteristics,so that there are many unsolved problems in application.Image representation is to analyze images effectively and obtain the key information,of which the performance influence image classification and recognition definitively.It is of great significance to the real work and social development.Sparse representation of images is a good image representation method with good robustness and anti-interference.It has been widely used in scene reconstruction,image perception,spatial-temporal technology and other computer vision applications.With the aid of multi-task learning which study features of multiple tasks simultaneously to improve the generalization ability of the algorithm,the sparse representation is done in-depth research in this dissertation.In this dissertation,a decorrelating multi-task learning framework joint sparse representation(DMTJSR)is proposed,which is used to learn a representation images with deeper image information.The main steps are as follows:First,because of the limitations of single task feature representation and to make more effective use of multi-features,a decorrelating multi-tasking multi-feature learning model(L2,1,1)is proposed,which was an improvement of sparse coding and multi-task learning basic models.Under the premise of few constraints,multiple features were combined in this model with the aid of multi-task learning theory to improve sparse perception model.Fixed dictionary was adopted while learning the coefficients of image sparse representation.On the one hand,the mixed norm constrains guaranteed the same characters share the same sparse mode,so that the similar features could be coded to similar coefficient,which solve the problem of locality.On the other hand,it removed correlation and redundancy among different features.Secondly,dictionary is an important factor affecting the sparse representation,the dictionary learning method was combined to optimize L2,1,1 model.A decorrelating multi-task learning framework joint sparse representation was proposed,which increase the discrimination of dictionary and performance of features representation.Finally,owing to the linear coding on the dictionary influencing the final representation,DMTJSR is extended to reproducing kernel Hilbert space.Using kernel technology to solve the shortage of linear combination in the coding process,it retains the features which were the nature of the image but has a nonlinear relationship with the dictionary.The experimental results show that L2,1,1 model was superior to basic multi-task models,and the performance of feature representation was superior to state-of-art results.Compared with L2,i,1,the performance of DMTJSR is still improved,which not only proved the necessity of dictionary learning,but also the effective ne ss of DMTJSR proposed in this dissertation.DMTJSR expanded to kernel space made up for the problem of linear representation,so that the final image presentation perform better than state-of-art algorithm.