A Multiple Kernel Learning K-SVD Algorithm And Its Application

With the rapid development of computer hardware,a large amount of data input into the computer processor.Most of the existing dictionary learning algorithms adopt the linear dictionary learning method,which do not need pre-training,so it has the advantages of fast learning and high sparsity and is suitable for large quantities of input data.However,the linear dictionary learning method is not effective in image recognition when the input sample is missing or the has denoise problem.In order to solve these problems,this paper improves the classical dictionary learning algorithm under the framework of collaborative representation and applies it to face recognitionBased on the classical K-SVD dictionary learning algorithm,this paper studies the input data preprocessing and improved dictionary learning methods.The main research contents are as follows:(1)Introduces the research background and significance,summarizes the relevant theoretical support,and analyzes the main work.(2)Describes the signal sparse technique and cooperative representation.In particular,introducing the classical K-SVD dictionary learning algorithm.(3)An improved nonlinear K-SVD dictionary learning algorithm based on label preprocessing and multiple kernel learning is proposed.Firstly,by setting up label regularization terms and computing the threshold,using three methods:directly deleting data,using small fault-tolerant data and mean completion data,preprocessing the labels of the same class,which can obtain a complete and expressive dictionary.Then,the K-SVD dictionary learning algorithm is introduced into the multi-kernel function composed of linear kernel function,polynomial kernel function and gaussian kernel function to learn the dictionary after preprocessing.In the update stage,the update signal is projected into the high-dimensional space for learning,so as to expand the dictionary learning space and train an over-complete dictionary,so as to obtain better classification performance.Finally,the cooperative representation is used to classify the test samples,and the test samples are put into a class with small residual error by calculating the residual value of L2 norm.The experiment shows that the nonlinear dictionary learning method has better classification ability when the input data is missing or has denoising,which indicates that the multiple learning K-SVD method has excellent performance in the field of image recognition in pattern recognition.