Research On Video Classification And Image Colorization Based On Subspace Learning

Image and video classification has become a hot and difficult research topic in computer vision due to its cross application in machine learning,pattern recognition and image processing.In addition,how to perform automatic color reconstruction for gray image is also a challenging research topic in computer vision.For the above two problems,this paper presents the corresponding algorithms.In chapter 3,this paper proposes a robust video classification algorithm based on spatial-temporal low-rank representation subspace learning,which utilizes both the spatial manifold structure and temporal consistency of video data: On one hand,this algorithm uses the improved low-rank representation model to extract the global subspace structure;on the other hand,this algorithm enhances the temporal consistency of the video effectively by adding the sparse constraint between adjacent frames.Compared with traditional methods,the proposed method is applicable to exploit the overall structure of data space while maintaining the temporal consistency of video data,and it is more robust to outliers and noises.An efficient numerical solution based on adaptive penalty linear alternating direction method(LADMAP)is established for the proposed model.Finally,we evaluate the performance of the proposed algorithm on two standard databases(HONDA and YaleB).The experimental results show that the algorithm is more robust.In chapter 4,this paper proposes an example-based image color reconstruction method based on weighted sparse representation learning.Firstly,each image is segmented into superpixels,and then the intensity and texture features are respectively extracted for each superpixel,which are grouped as the final feature descriptor.The feature descriptors from the reference image are defined as the representation dictionary.The corresponding relationship between the target gray image and the color reference image is established by solving the weighted sparse representation matching problem.The color information of target superpixels is reconstructed based on the color from the corresponding reference image.Finally,a bilateral filter with the guidance of intensity information is used to improve the consistency of final color reconstruction results.The experimental results show that the algorithm has achieved satisfactory results on the task of color reconstruction of natural images.