Font Size: a A A

Super-Resolution Image Reconstruction From Low-Resolution Image Sequences

Posted on:2008-07-21Degree:MasterType:Thesis
Country:ChinaCandidate:J YuFull Text:PDF
GTID:2178360215494871Subject:Computer application technology
Abstract/Summary:PDF Full Text Request
Super-resolution image reconstruction technique refers to restoring a high-resolution and high-quality image from multiple low-resolution observations (or a video sequence) degraded by warping, blurring, noise and aliasing. The spatial resolution is the principal factor in determining the quality of an image. A high-resolution image can offer more details that are important for the analysis in various applications. Image interpolation used to increase the size of a single image cannot recover the high-frequency components lost or degraded in the low-resolution sampling process. Super-resolution image reconstruction algorithms investigate the relative motion information between multiple low-resolution images of the same scene and increase the spatial resolution by fusing them into a single frame.Super-resolution image reconstruction has been proved useful in many practical cases, including medical imaging, satellite imaging and video applications. Recently, it has become one of the most active research areas. With the development of super-resolution image reconstruction technique, more and more methods have been proposed to enhance the definition of a single video frame. Existing methods can be classified into two major categories: the frequency domain approach and the spatial domain approach. The spatial domain approach is more flexible and more convenient to apply a priori knowledge for regularization and, therefore, attracts more researchers'attention.In this paper, we use the set theoretic POCS algorithm and the Bayesian MAP estimation algorithm for super-resolution image reconstruction, which are the most popular among spatial domain methods. We design a simulation-correction iteration method to reconstruct high-resolution images according to the theory of POCS, and also propose an improvement to reduce the amount of edge artifacts present in the reconstructed image. The proposed method weights the blur PSF centered at an edge pixel with an exponential function, so that the modified PSF coefficients decrease in the direction orthogonal to the edge. Experimental results show that the modification effectively reduces the visibility of the artifacts and obviously improves the quality of the reconstructed image. Besides, additional computational complexity for edge detection is very small.In this paper, we study and implement the MAP estimation algorithm for super-resolution image reconstruction. Also, we propose two improvements to reduce the computational complexity of the standard MAP estimation algorithm. The first improvement selects an inexact line search to identify the step length, which avoids the computation of the computationally expensive Hessian matrix. The second one computes directly the increment of the MAP objective function as the component of the gradient, which avoids the redundant computation of the objective function. Experimental results show that the improvements not only reduce significantly the computation time, but also guarantee the solution's convergence and maintain the similar image quality.Motion estimation refers to estimating 2-D motion vector field of the scene or object according to temporal information redundancy in a video sequence. When it comes to super-resolution image reconstruction, it is used to map pixels of all low-resolution observations onto corresponding pixels in the reference frame. So its accuracy is highly required. Because of its simplicity and efficiency, block-based motion estimation has recently been widely used in super-resolution image reconstruction. In this paper, we propose a hybrid method combining spatial prediction and the CDS algorithm. If the motion of the current block is similar to that of its neighbor blocks, choose the best candidate block from the neighbor blocks and use its motion vector to form an initial estimate for the current block. The neighbor block whose motion vector yields the minimum block distortion is called the best candidate block. The true motion vector is then obtained by comparing the search points of SDSP centered at the initial estimate. If the current block isn't correlated with its spatial neighbors, search the motion vector from the origin of the search window using the CDS algorithm. Experimental results show that the proposed algorithm achieves a better tradeoff between search speed and matching accuracy for super-resolution image reconstruction, compared with N3SS, DS, HEXBS, CDS, and CDHS.
Keywords/Search Tags:Super-resolution image reconstruction, Motion estimation, POCS (projection onto convex sets), MAP (maximum a posteriori probability) estimation
PDF Full Text Request
Related items