Research On The Sparsest Decomposition Of Visual Image Of Intelligent Robot

Robot image has the characteristics of complex background,real-time change,a large amount of data and fast running speed requirement.If the traditional Nyquist sampling theorem is used to obtain the image,a huge amount of sampling data will cause great pressure on the subsequent image processing.The theory of compressed sampling or compression sensing can solve this problem very well.For robot vision,the main purpose of image processing is to obtain the feature information of the image without the need of complete original image.Therefore,it is only necessary to process the low-dimensional data and obtain the characteristic information in the compressed domain so as to improve the response and speed of the robot.Compressed sensing combines the sampling and compression of the data,compresses it while capturing the image signal.Compressed sampling can greatly reduce the number of dimensions of the data,reduce the amount of data,the measurements obtained include the rich information which have the ability to reconstruct the original image signal.The first precondition for compressed sensing is that the image signal has sparsity on the transform base or dictionary.The sparser the image is on the selected sparse dictionary,the less measurements it needs and the faster the post-processing.Therefore,good sparse matrix,in order to be the most effective and concise expression of the original signal,in order to ensure the accuracy of the restored signal.The main work of this paper is as follows:(1)Based on the research of sparse representation method,we think that the redundant dictionary obtained through learning has the better ability to adaptive image reconstruction,effectively expressing the local structural features of the image,and making the signal has sparser expression on the dictionary.However,there are some problems in the dictionary updating phase,such as large computational complexity and time consuming.To solve this problem,we analyzed the principle of various dictionary learning algorithms through further research on various dictionary learning algorithms.In the update phase of the dictionary learning,this paper improves an approximate solution instead of the most time-consuming SVD decomposition in the K-SVD algorithm.It is a simplified method of approximating the K-SVD algorithm.It abandons the process of repeating updating the sparse coefficient matrix in the dictionary update phase of the K-SVD and the approximate K-SVD algorithms.The experimental results show that the simplified approximate K-SVD algorithm can reduce the time spent on dictionary learning to some extent without reducing the Peak signal to noise ratio of the image.(2)The above method has a limited increase in dictionary update speed and a highcomputational complexity,which can not be applied to the sparse representation of robot images.Recently,a dictionary learning algorithm called SGK has been used as an effective alternative to the classical K-SVD algorithm,which has faster execution speed and comparable dictionary training effect than K-SVD.However,if the SGK algorithm only through a single dictionary update,there is still room for improvement in computational complexity.For this reason an improved algorithm proposed based on SGK algorithm.We transform the residual term in the SGK algorithm into the form of simultaneous updating of multiple columns of atoms,and then use the least squares method to update the r atoms in the dictionary.The experimental verification and computational complexity analysis,the algorithm effectively reduces the single iterative computation and the time of sparse representation,improve the efficiency of dictionary learning.(3)Due to the high complexity of the robot image and the randomness of the scene,an online dictionary learning algorithm suitable for robot images is finally proposed in this paper.In the sparse coding phase,the redundant dictionary learned from the previous image dictionary is used to update the sparse coefficient matrix,which makes the dictionary inherit the useful information of the previous frame while learning new information and adapt to the changing environment.The dictionary update phase uses the least square method to update parts of the dictionary in sub batch.Compared with off-line dictionary,the dictionary obtained by online algorithm is used in sparse representation and reconstruction,and its reconstructed image quality is improved.With the gradual change of robot scene image,its advantage will be more obvious.