Geometric Distortion Correction by Utilizing Image Features and Image Quality Measures

Abstract

Geometric distortion is a very common problem in image capture, remote sensing, image display, and medical imaging. In a complicated imaging system, it can further induce ghosting, blurring, and intensity change. Traditional methods will adapt the approach of pre-calibration. These pre-calibration methods usually require some human intervention, thus their application scope is quite limited. In addition, they are not applicable in variable geometric distortion. The basic geometric distortion correction is to apply appropriate geometric transformations for images. In this thesis, geometric distortion correction means a generalized one that adjusts spatial transformation parameters. By extracting image features and utilizing image quality measures, we propose three novel algorithms for three geometric distortion correction problems, i.e. radial distortion correction, motion compensation in Magnetic Resonance (MR) images, and disparity adjustment. Due to the nonlinearity in radial distortion, it will result in spatially varying distortion for the image. We adapt the framework of utilizing feature transform and image measure to estimate the radial distortion parameters. First, we use edge extraction and transfer features into the feature map; and then assess the quality of feature map to estimate distortion parameters for the image correction. By using this framework, we develop fully automatic calibration and it can be applied in popular fisheye lenses and medical wide-angle endoscopes, which usually require real-time correction. In addition, we also extend the framework to different types of calibration patterns and zoom lenses with varying geometric distortion. Because of the special capturing procedure of MR images in Fourier domain, the motion of subject during the imaging process will result in ghosting and blurring, and thus different motion compensations for different phase encoding lines have to be estimated. Traditional methods use greedy and exhausted search for optimizing an image quality measure, and ignore other important information. We search repeating edge for collecting candidate motion vectors and use graphical models to fuse different information (including symmetry of frequency, smoothness of motion, and an image quality measure) to solve the problem. The disparity adjustment can be regarded as geometric distortion correction. In order to provide better viewing experience of stereoscopic images, we have to adjust the left and right images geometrically. The simplest method is to shift the left and right images to adjust the disparity range within a comfort zone. However, the shifting in stereoscopic images with large disparity range may not work. Hence, we take image feature for image segmentation and utilize a stereoscopic image quality measure to decide different geometric transformations for different segments for the image correction to improve the viewing experience. In this thesis, we focus on a general geometric distortion correction over the domain of spatial parameters. By utilizing image features and image quality measures, we propose novel algorithms to resolve the three image correction problems described above.