Edge Preserving Interpolation of Digital Images Using Fuzzy Inference

Hou-Chun Ting, Hsueh-Ming Hang
1997 Journal of Visual Communication and Image Representation  
Blurring and jaggedness are the two most obvious arti-This paper presents a novel edge preserving interpolation facts in the interpolated images using the preceding intermethod for digital images. This new method reduces drastically polation methods. These artifacts are produced by the lowthe blurring and jaggy artifacts at the high-contrast edges, pass filter used to remove the unwanted highpass replica which are generally found in the interpolated images using of the interpolated images in
more » ... frequency domain. Beconventional methods. This high performance is achieved by cause it is not possible to implement the ideal low pass two proposed operations: a fuzzy-inference based edge preservfilter in practice, nonideal filters such as the zero-order ing interpolator and a highly oblique edge compensation scheme hold (nearest neighbor) and the first-order hold (bilinear) developed based on an edge orientation detector. The former operators are often employed to filter out the high fresynthesizes the interpolated pixels to match the image local quency replica. These nonideal lowpass filters suppress characteristics. Hence, edge sharpness can be retained. Howlow frequency components and bring in high frequency ever, due to the small footage of the fuzzy interpolation method, it cannot avoid edge jaggedness along the highly oblique edges component aliasing. The former, low frequency suppresthat have very sharp angles against one of the coordinates. sion, reduces the spatial resolution of the interpolated im-Therefore, a segment matching technique is developed to idenages (blurring) and the latter, undesired high frequency tify precisely the orientation of the highly oblique edges. Comaliasing, produces broken edges (jaggedness). bining these two techniques, we improve significantly the visual Recently, several adaptive nonlinear methods have been quality of the interpolated images, particularly at the highsuggested to tackle the aforementioned problems. As it contrast edges. Both the synthesized images (such as letters) has been pointed out to various levels of extent [8-13] that and the natural scenes (captured by camera) have been tested in order to achieve a better visual quality, the lost high and the results are very promising. © 1997 Academic Press frequency components need to be reconstructed based on certain assumptions (models) of image characteristics. The basic concept is as follows. In the down-sampling process, INTRODUCTION the high frequency components in the original high resolu-Interpolation is an important technique in multi-rate tion pictures are removed by lowpass filtering before resoimage signal processing such as pyramid coding [1], multilution reduction. Now, the interpolated image has a higher resolution television broadcasting [2, 3], and image zoomspatial resolution and thus our eyes expect to see more ing for viewing comfort. Linear interpolation methods are details in the picture. For most typical pictures, the informacommonly used [4] [5] [6] [7] . According to sampling theory, an tion contained in the subsampled (down-sampled) image ideal lowpass filter is needed to remove the replicated usually can provide clues about the lost components. The copies of the down-sampled (low resolution) signal spectra most obvious example is the high-contrast edges, which located at high frequencies. Therefore, the conventional are modeled as step signals. Thus, if a segment of image linear spatial invariant interpolator is designed to be a is recognized as an edge, proper high frequency compolowpass filter. An extension of the above approach is patchnents can be added into the interpolated (lowpass) signals ing pieces of continuous (spline) functions that match the to enhance the edge shape. Therefore, non-linear and/ given (known) pixels, and then, the pixels in-between can or spatially variant interpolators designed based on this be synthesized from the continuous functions [5] . concept could potentially produce images with better perceptual quality than the conventional filtering methods. Our approach in this paper is another attempt to regener-
doi:10.1006/jvci.1997.0364 fatcat:e54yduxyjrghxpydmpyf3y36tm