Fuzzy Motion-Adaptive De-Interlacing with Edge-Adaptive Spatial Interpolation

Abstract

The quality of the motion-adaptive de-interlacing algorithm developed in previous Chapters can be increased by improving the performance of the spatial interpolator. Until now, line average has been used but its results are poor in the reconstruction of areas with edges. Since the human visual system is specially sensitive to the edges of an image, a correct interpolation of them has a vital importance to achieve a positive visual inspection. Hence, although an interpolated image obtained by line average could feature an overall PSNR value superior to the PSNR result obtained by an edge-dependent algorithm, human appreciation would usually prefer de-interlaced pictures obtained with edge-dependent algorithms. Let us illustrate the advantages of using an edge-adaptive algorithm by examining an image with a high number of clear defined edges as shown in Figures 4.1 and 4.2. The reconstruction of clear, sharp, and high-contrast edges are significantly improved by using a conventional edge-dependent algorithm such as ELA (Edge-based Line Average) because line average introduces a blurred effect. Edge-adaptive de-interlacing algorithms explore a neighborhood of the current pixel to extract information about the edge orientation [1], [2]-[9]. Once determined the direction with the major probability of an edge, the luminance is interpolated in that direction. The first proposal following this approach, which was called ELA [1], works with 3+3 pixels in the upper and lower lines. It performs well when the edge direction agrees with the maximum correlation, but otherwise introduces mistakes and degrades the image quality. These mistakes usually appear when the edges are not clear, the image is corrupted by noise, there is a high number of details, etc. In addition, ELA lacks the ability to detect horizontal edges. Several proposals have been presented recently in the literature to avoid the above shortcomings of the ELA algorithm [2]-[9]. The approaches reported in [2] and [3] focus on enhancing the reconstruction of horizontal edges. In [2], line doubling and an edge-adaptive de-interlacing algorithm based on ELA are combined to interpolate horizontal edges. In [3], an edge-based line average is proposed, which uses the 3+3 pixels of ELA and the interpolated pixel value that was calculated previously. Other algorithms use a larger neighborhood to get more information about the possible edge direction. For instance, the algorithm presented in [4] consists of a modified ELA module and a contrast enhancement module. The ELA module increases the processing window up to 5+5 pixels whereas the second module reduces detail losses due to the interpolation. The neighborhood is enlarged up to 6+6 taps in [5], 7+7 taps in [6], [7], 11+11 taps in [8], and 34+34 taps in [9]. A higher number of pixels in the aperture provides much information about edges but at expense of increasing the algorithm complexity. The algorithms presented in this Chapter are inspired by the ELA scheme and include a simple fuzzy system that models heuristic knowledge to ensure the presence of edges. Thus, they have been named Fuzzy-ELA algorithms. This Chapter is organized as follows: Section 4.1 describes the ‘Basic Fuzzy-ELA’ algorithm and its performance when de-interlacing the battery of sequences presented in the previous Chapters. Further modifications of the ‘Basic Fuzzy-ELA’ algorithm are presented and analyzed in Section 4.2. After comparing these approaches, one of them is selected as spatial interpolator to be used in our fuzzy motion-adaptive algorithm in Section 4.3. Finally, some conclusions are expounded in Section 4.4.