Image reconstruction in the contour domain

Abstract

Digital image enhancement systems are essentially pixel-based. In this paper, a novel method of multi-frame image enhancement employs contours rather than pixels as the primitive working unit. The feasibility of this proposal is based on the fact that gray-scale images can be sufficiently and accurately represented by their contour maps if a suitable contour model and scale selection method is used [5]. In this process several low-resolution shifted frames of the same scene are (a) registered (b) transformed into their contour domain, and (c) grouped or mapped globally within a singular framework to create an enhanced contour composite. Inverting this improved contour representation yields a high-fidelity reconstruction of an image in pixel format with more details than any of the original input frames. The mapping of the low-resolution contour images within a singular framework takes advantage of the differences (or shifts) existing between their contour representations. These shifts are computed using an innovative image registration technique that can achieve accuracies of up to 0.01 of a pixel. The technique is based on DCT (Discrete Fourier Transforms) and normalized cross-correlation. Controlled tests show that this global contour mapping approach, with subsequent converting to a compact and improved raster image, provides for an efficient and simplified method for multi-frame image enhancement. As compared to pixel-based multi-frame image enhancement processes, the proposed method may be computationally more efficient as it allows the detection, filtering and elimination of redundant contour data which may have little or no effect in the final enhancement. Although still in the experimental stage and restricted to gray-scale images, the proposed method is particularly viable in applications involving facial enhancements of compressed video frames.