Research On Sparse Representation Algorithm And Its Applications In Character Video

Sparse representation is an algorithm proposed to solve a large amount of redundancy in natural signals.It is an effective method for processing high-dimensional data and is applied to image classification,image reconstruction,and other aspects.Dictionary learning is an effective method to achieve sparse representation.The dictionary learning is applied to human motion recognition by learning the dictionary that represents the entire training set by using as little data as possible.In this paper,by L2,1 regularization,the sparse representation is strengthened at the atomic level;on the basis of this,a class-specific dictionary and a shared dictionary are combined to improve the sparseness and robustness of the dictionary level;for semantic information in video,Convolutional neural network(CNN)achieves the extraction of semantic features and implements the semantic classification of human action semantics.The experiment verifies the validity of the method.(1)An atomic selection method based on the L2,1 norm is proposed and applied to image compression sampling.For the problem that the selection rules of atoms are difficult to determine,an atomic selection method based on the L2,1 norm is proposed.Using image blocks as processing units,L2,1 norm selection effectively eliminates useless atoms in the image block and enhances the discriminability of the block features.It also takes into account the correlation between image blocks,that is,an atom is useless in determining an image block.The atom is useless on all image blocks,avoiding the instability of feature selection among image blocks,improving the accuracy and efficiency of reconstruction.(2)Based on the tensor RPCA preprocessed video,a combination of class-specific dictionary and shared dictionary is proposed and used for human motion recognition.First of all,for the recognition of moving objects,the background of the human body motion recognition is complex and affects the recognition effect.Therefore,the tensor RPCA method is used to remove the background in the video and extract the video foreground,thereby eliminating the interference of the complex background.Second,the video foreground is trained using a class-specific dictionary and a shared dictionaryto form a dictionary that represents the action class.Finally,experiments were conducted in the UCF sports,UCF50,and HMDB51 databases.Experimental results confirm that the classification effect is improved and the dictionary method is effective in this paper.(3)A human action semantic recognition based on convolution neural network(CNN)and kernel dictionary learning is proposed.Firstly,convolution neural network(CNN)is applied to convolution processing of key frames,highlighting the semantic features of images,and reducing dimensions through pool layer.Secondly,a semantic kernel dictionary is constructed to classify human action semantics without occlusion or occlusion.Finally,through experiments in UCF sports,UCF50 and HMDB51 database,the experimental results confirm that the classification effect is improved and the classification time is reduced.