Research On Key Technology Of Image Super Resolution Reconstruction

Due to the limitation of imaging conditions and imaging modalities,the image resolution obtained cannot satisfy the demand of the practical application.Increasing image resolution beyond hardware limit is of great significance in image processing area.Super resoluiton(SR)reconstruction technology utilizes signal processing methods to obtain a higher resolution(HR)image making use of the complementary information from multiple frames of the same scene.This technology,which break through the limits of the optical sensor manufacturing technology and realizes higher system resolution of observation without changing current hardware conditions,is a cost efficient way to improve the image resolution,and it is of great practical value in various areas such as remote sensing,surveillance,military,infrared,etc.This dissertation researches problems existed in current SR technology,such as subpixel image motion estimation,compressive sensing(CS)based SR reconstruction,visual saliency based SR reconstruction,and infrared image total variation(TV)SR.Main innovations of our research are as follows,(1)By incorporating the idea of coarse-to-fine,a hierarchical subpixel motion estimation algorithm is introduced.Proposed algorithm has a higher estimating accuracy and smaller computation cost by converting traditional motion estimation algorithm which needs to interpolate the LR images in order to obtain a subpixel level motion parameter to only need to interpolate the 3D curved surface.The algorithm not only guarantees the accuracy of the estimation but also reduces the consumption of the calculation.The final experiment shows the effectiveness of the proposed algorithm.(2)Proposed a single image SR algorithm by based on CS theory.The proposed algorithm conducts the contourlet transform as the sparse base without using training dictionary,which is different with the traditional algorithm.In the reconstruction process,the algorithm utilizes modified regularized orthogonal matching pursuit(mROMP)method to solve the optimization problem.The simulation results show that the proposed algorithm is able to reconstruct SR image with a single LR image,and obtain satisfying results.(3)A multi-frame SR algorithm based on multi-scale non-local dictionary is proposed by extending the CS theory into the area of multi-frame SR.The proposed algorithm uses the phenomenon that natural images possess self-similarity characteristic,and trains the dictionary with different scale images and MSE between sample and signal to be reconstructed are used to select the initial dictionary in order to solve the problem that K-SVD training procedure is depending too much on the initial dictionary.The simulation results show that the algorithm is able to exploit useful information effectively and obtained better reconstruction results.(4)Proposed a visual attention based area SR algorithm.By incorporating the characteristic that people only focus the analysis on the area they are interested instead of the whole image when they are staring at an image,the proposed algorithm introducing dynamic pyramid decomposition based on Itti algorithm to analyze the input image in multi-scale which is to make sure the saliency maps of different scale of the same scene are same,and this will increase the stability of the algorithm.In addition,the proposed algorithm is able to obtain the area that people really interested,and this will reduce the calculation tremendously,and improve efficiency.(5)An adaptive regularizer is constructed by using the property that curvature difference is able to distinguish the detailed region and smoothed region to replace the conventional TV regularizer and an adaptive total variation(TV)multi-frame infrared SR algorithm based on curvature difference function is proposed.It is able to solve the over-somooth problem existed in traditional reconstruction algorithm.Majorization-minimization method is used to obtain HR image by solving the optimization problem.Through simulation and comparison with other algorithms,proposed algorithm is able to improve the reconstruction result by incorporating adaptive regularizer and MM theory.