Font Size: a A A

Image Super Resolution Reconstruction

Posted on:2014-02-27Degree:DoctorType:Dissertation
Country:ChinaCandidate:Y X YangFull Text:PDF
GTID:1228330398472840Subject:Pattern Recognition and Intelligent Systems
Abstract/Summary:PDF Full Text Request
Image super resolution refers to the reconstruction of a high resolution image from one or a set of blurred low resolution images. Considering that there are many cases only one low resolution image is available, so in this paper, we mainly focus on super resolution from single low resolution input image. The goal of single image super resolution is to estimate a high resolution image from a low resolution input. There are mainly three categories of approach for this problem:interpolation based methods, learning based methods, and statistical reconstruction based methods.The interpolation based methods are simple but tend to blur the high frequency details. In this paper, we propose a new interpolation based method. Our approach uses the quad tree segmentation to partition the low resolution image, and takes the edge-directed interpolation to each segmented band of the low resolution image, and then applys a wavelet projection to optimize the high resolution image got from the local interpolation. The experimental results show our interpolation method greatly improves the image edge ringing effects of traditional interpolation algorithm.The statistical reconstruction based methods require a probability density function of the data known as a prior image model. Maximum a Posteriori (MAP) is one of the most popular statistical methods, so MAP statistical reconstruction based methods are sparked within this research community. In this paper, we propose a novel image super resolution method based on MAP statistical reconstruction. Our approach takes the wavelet domain Implicit Markov Random Field (IMRF) model as the prior constraint and utilizes the MAP theory to construct the objective function with this model. Furthermore, we employ steepest descent method to optimize this objective function. The experimental results demonstrate that our method obtains the superior performance in comparison with traditional single image super-resolution approaches.The learning based methods "hallucinate" high frequency details from a training set of high-resolution/low-resolution image pairs, and this kind of methods highly relies on the similarity between the training set and the test set. It is still unclear how many training examples are sufficient for the generic images. In this paper we propose a new learning based super-resolution method which base on the haar wavelet transform and back-propagation network. Considering the self-similarity between the detail subbands of Haar wavelet decomposition, firstly our approach trains the Back-propagation network to approximate the self-similarity relationship and then uses the trained network to predict the detail subbands of Haar wavelet decomposition.Depth image of the real three-dimensional scenes get more and more of our attention. Laser range scanners can provide extremely accurate and dense three-dimensional measurement over a large working volume. But these high quality scanners measure a single point at a time and it limits their applications to static environments only. Recently new time-of-flight sensors have been developed to overcome this limitation. These sensors measure time delay between transmission of a light pulse and detection of the reflected signal on an entire frame once by using extremely faster shutter. Though this technology is promising, in the current generation, these time-of-flight sensors are expensive and very limited in terms of resolution. How to improve the resolution of the depth image is an interesting topic.In this paper, given a low resolution depth image as input, we recover a high resolution depth image using a registered and potentially high resolution camera image of the same scene. Based on the fact that discontinuities in range and color tend to co-align, we assume that color image and depth image of the same scene have an approximately linear relationship in the locally within the small window. And we apply Matting Laplacian Matrix to exploit this linear relationship and construct an optimization problem. Furthermore, we employ steepest descent method to optimize this objective function. But, calculating the Matting Laplacian matrix is very time-consuming, so we apply the guided image filter to exploit this linear relationship between the color image and depth image of the same scene. Using guided image filter, we integrate the registered high resolution camera image into the range data and generate an initial high resolution depth image. Moreover, a reconstruction constraint is also used to further improve the quality of the initial high resolution depth image iteratively.Taking into account that the assumption of the approximate linear relationship between color image and depth image is too simple, some practical complex scenes, color and depth at the edge do not satisfy this assumption, next we use local structural features of high resolution camera image to construct the regularization term and construct a new optimization problem. Experiments demonstrate that our approach can get excellent high resolution range image in terms of both its spatial resolution and depth precision.
Keywords/Search Tags:Image super-resolution, Maximum a posteriori, Wavelet domain ImplicitMarkov random field model, Wavelet domain Hidden Markov tree model, Haarwavelet transform, Back propagation network, Depth image, Laser range scanner, Time-of-filght scanner
PDF Full Text Request
Related items