Диссертация, Ghent University, 2009, -325 pp.In this dissertation, we study and develop several image and video resolution enhancement techniques. As the first main contribution, we have developed a novel non-linear image interpolation technique that eliminates unwanted artefacts produced by linear interpolation. The proposed algorithm sharpens edges by mapping the image level curves using constrained adaptive contrast enhancement techniques. Level curves are defined as spatial curves with a constant intensity level. To avoid jagged edges, the level curves are preprocessed by a constrained isophote smoothing scheme. Experiments show improvements in both numerical PSNR results as well as in visual quality compared to other state-of-the-art interpolation techniques. A second novelty in this dissertation is the introduction of two new image colour priors in the Bayesian restoration framework. On the one hand, the adaptive bimodal colour prior assumes that the value of an edge pixel is a combination of the colours of two connected regions, each having a dominant colour distribution. On the other hand, the multimodal colour prior is proposed for images that normally just have a few dominant colours. Restoration results show the effectiveness and the visual superiority to other interpolation/restoration schemes for images with a strong colour modality. Both colour priors are found very suitable to the restoration of drawings and cartoons, logos, maps, etc. Common image restoration techniques only exploit the spatial redundancy in a local neighbourhood. In this work, we have demonstrated that the estimation of the restored pixel intensity can be based on information retrieved from the whole image, thereby exploiting the presence of similar patterns and features in the image, which we call repetitive structures. The new approach is referred to as the non-local strategy, which is also related to the exemplar- and fractalbased approaches. As the third important contribution, we have developed a novel resolution enhancement scheme that exploits these repetitive structures. We also have extended and optimized this algorithm for document image processing applications. An improved character segmentation scheme is introduced to reduce the computational complexity and an additional text specific image prior is included in the Bayesian restoration framework. Experiments show that characters are reconstructed very well. In addition, OCR accuracy results show significant improvements in comparison with other existing resolution enhancement methods. The proposed algorithm is not restricted to font type or alphabet, therefore, it is also suitable to generic symbols such as musical notes, hieroglyphics or mathematical symbols. The same strategy can also be applied in an exemplar-based search engine and in an efficient document compression scheme, which opens up new possibilities in future applications. Multi-frame super-resolution is quite a complex problem, which spans over several fields of image processing, such as motion estimation or image registration, image reconstruction from irregularly spaced samples (also called fusion), image deconvolution and denoising. Due to its relatively low computational load and low memory requirements, the standard three-step paradigm of the superresolution approach is recommended in most practical applications. These three successive steps are subpixel image alignment, image fusion and image restoration. As the fourth main contribution in this dissertation, we have developed very accurate registration algorithms, both in photometric and geometric domain. The proposed low-resolution-to-high-resolution gradient-based registration method with steering kernel regression fusion currently produces the most accurate subpixel information among several state-of-the-art methods. For the photometric and joint geometric/photometric registration problem, we have proposed the use of the total least square framework. The total least square solution produces in both cases more accurate and consistent registration parameters compared to the ordinary least square approach, which is commonly employed in the literature. In the same spirit, we also have proposed and developed the kernel regression algorithm, which is a state-of-the-art fusion technique, in the total least square sense to handle positional or registration errors. Numerical experiments show that the proposed method is more accurate and robust compared to the standard kernel regression algorithms. In an extensive study in close collaboration with MEDISIP-IBBT-IBITECH and the department of radiology (Ghent university hospital), we have pointed out some limitations in the recent developments in super-resolution magnetic resonance imaging (MRI) reconstruction and we also have argued that classical super-resolution cannot be applied in the Fourier encoded plane because of the complete absence of frequency aliasing during MRI acquisition. As the fifth main novelty in this work, we have introduced an elegant way to enhance the image resolution by multiple rotated MRI acquisitions. We have proposed a novel hybrid reconstruction algorithm that performs resampling in the image domain followed by fusion of multiple aligned k-space data. Simulations demonstrate the superiority of our method, both quantitatively and qualitatively. The results also demonstrate improvements on real MRI data of a resolution phantom and an onion. Analyzing the Fourier data reveals that we really have gained true spatial resolution. Practical implementations require non-squared voxel sizes as in PROPELLER MRI schemes.General introduction. Linear interpolation theory. Level curve mapping interpolation. High-resolution image restoration using colour priors. Non-local reconstruction methods. Application to document image processing. Multi-frame super-resolution restoration. MRI resolution enhancement. Conclusion.
Чтобы скачать этот файл зарегистрируйтесь и/или войдите на сайт используя форму сверху.