Research On Spatio-temporal Data Denosing And Detection Technology Based On Tensor

With the increasing demand for visual data display and information mining in electronic products,the structural integrity and information richness of the obtained visual data are becoming more and more perfect.However,in the actual operation process,due to issues with the shooting equipment,shooting angle,and shooting method,it is easy to obtain visual data with strong noise,which makes the display effect easily unsatisfactory.At the practical application level of infrared imaging weak small target detection and video background subtraction,due to the complex and varied background,diverse target morphology,and complex noise,the detection effect is not perfect even with high equipment accuracy.For an image data,most of the information is sparse(meaning that the modeled matrix of the image is low rank),so it is particularly important to separate the sparse part(i.e.sparsity measure)from the key part to establish a model.Decomposition and segmentation in the modeling of spatiotemporal data such as videos,traffic flow,and other scenes containing temporal information dimensions are also key technologies for exploring optimization breakthroughs.This article focuses on the model of visual data,and the main research content includes the following aspects:(1)A multi-channel weighted truncated Schatten-p norm minimization model for RGB color image denoising is proposed to address the lack of flexibility of the nuclear norm minimization method due to excessive shrinkage of rank components and equal processing of different rank components.Nuclear norm minimization is a special non convex rank minimization convex relaxation scheme,which is mainly applied to denoising methods and background subtraction in image denoising and object detection,which still pose research challenges.In this paper,the nuclear norm regularization is extended to minimize the Schatten-p norm with different weighted singular values.The model is based on the characteristics of color image data: different noise intensities of each channel and non local self-similarity patches,and the model is solved by alternating direction multiplier.(2)Meanwhile,for the first time,the multi-channel weighted truncated Schatten-p norm minimization model was extended to tensor space,and an improved truncated Schatten-p norm weighted tensor infrared patch model was proposed.The alternating direction multiplier method framework was used to solve the problem,and some non convex penalties were used instead of L1 norm penalties.Due to the fact that the Schatten-p norm is more flexible and efficient in tensor decomposition and completion compared to the nuclear norm,and also adds the advantage of truncating the Schatten-p norm,the fitness of this method can be seen through existing research on tensor nuclear norms.This method is characterized by grouping non local self-similarity distributed in multidimensional imaging data and similar patches in sparse linear approximation,and it is applied to infrared imaging target detection based on spatiotemporal tensor model in this paper.The results of qualitative and quantitative analysis experiments in infrared target detection show that this method is superior,and the visual results also show that the improved method performs better than other methods,and the obtained target structure information has higher integrity.(3)A tensor non convex sparse metric robust principal component analysis model is proposed for most grayscale video data with high sparsity.Sparse modeling,as a promising method for analyzing data,holds a pivotal position in the entire field of computer science and engineering.Background subtraction is a key technology in computer vision,object detection and tracking,and other fields.For most natural data,it is also necessary to recover a low rank tensor from a small number of observation tensors,and denoising is particularly important in sparse data.In processing these data,denoising algorithms are used as a prior to model based inversion.The combination of global truncated singular value decomposition and local robust principal component analysis can make use of more spatial information.In video algorithms modeled as low rank tensors,it is shown that models using non convex penalty functions have good foreground separation and background subtraction performance.