|M.Sc Thesis||Department of Electrical Engineering|
|Full Thesis text|
A new approach to image interpolation using spatial relationships between adjacent pixels of an image is introduced. In the first stage, the localized structural relationships are studied based on the sparse version of the image. In the second stage, the relationships and the concluded governing rules of the image are used to build an interpolated image. Our method is compared with the main existing interpolation methods - bilinear, bi-cubic and spline for enlargement of black and white images. The results indicate significant reduction in the blockiness and smoothing effects compared to existing methods. The new method is also applicable to one-dimensional signals such as audio. In many cases, depending on the input, it outperforms the traditional Sinc interpolation.
The proposed method may be also implemented in CCD cameras for Demosaicing purposes using the consistent relationships between adjacent pixels. The correlation rule is studied first for each color component separately, then difference images (modified hues) are built to eliminate the colors correlation. The obtained smoother signal is studied and reconstructed. Since in the case of CCD Bayer pattern not all the color components are equally represented, the algorithm deals with the major green component differently from the red and blue components. The green component is used as a basis for the whole image reconstruction due to its largest amount of pixels. A statistical extension is added to the algorithm to construct a large number of relations' cases and improve the visible results accordingly.
We compared our method to other known techniques and observed a significant improvement in side effects such as ghost colors and unreal edges. Our conclusion is that the proposed method can improve interpolation and demosaicing tasks in image processing, and may be applied to CCD images reconstruction as a software tool or implemented in hardware inside digital cameras and similar equipment.