Interlaced video frame consists of two sub fields taken in sequence, each sequentially scanned at odd and even lines of the image sensor. Analog television employed this technique because it allowed for less transmission bandwidth and further eliminated the perceived flicker that a similar frame rate would give using progressive scan. The cathode-ray tube (CRT), based displays were able to display interlaced video correctly due to its complete analog nature. All of the newer displays are inherently digital, in that the display comprises discrete pixels. Therefore, the two fields need to be combined into a single frame, which leads to various visual defects, which the deinterlacing process should try to minimize. Deinterlacing has been researched for decades and employs complex processing algorithms. However, consistent results have been very hard to achieve.1 There are various deinterlacing methods with different levels of quality and complexities, which are: (1) basic fixed methods such as the line average and field average; (2) adaptive methods that select one of the basic methods according to the motion information extracted from the image sequence; and (3) methods that utilize the motion estimation and compensation. In this letter, we focus on intra-field deinterlacing method, which belongs to category (1) and gives good performance with low complexity.
Most edge-based deinterlacing methods utilize the uniform weighted sum of two-tap filter [TTF: (Ref. 2)] to estimate the current pixel to be interpolated in the determined edge direction.3 The line average (LA) method4 is commonly used in the software approach, which adopts a single field to reconstruct a progressive frame in 90 deg edge direction. However, the frequency response of two-tap average filter looks like a bell shape as shown in Fig. 1(a). It is obviously known that LA shows high peak signal-to-noise ratio (PSNR) performance. However, it provides apparent jaggedness at edge area.
Meanwhile, the sinc function is known as an ideal filter, and its frequency response is rectangular-shape. By the sinc function, we can design the filter to have a more rectangular shaped frequency response than the TTF, so interlaced signals can be rebuilt more precisely. In Refs. 3 and 4, the authors adopted one of the sinc interpolation filters, which is a one-dimensional six-tap filter (), and coefficients of STF are determined by approximating the sinc function. This is same to the one in the video codec, such as MPEG-4, H0.264/AVC, and high efficiency video coding (HEVC), to decrease residual errors.5 However, this method only regards the similarity of sinc function, and consciousness of the closeness or spatial locality is not taken into account.
In this letter, we proposed filter design algorithm, which minimizes the mean-squared deinterlacing error over a training set (TS). In Sec. 2, we propose filter design algorithm in detail. The simulation results are shown in Sec. 3 and the conclusions are made in Sec. 4.
Proposed Filter Design
Least mean squares method is a class of adaptive filter utilized to simulate a desired filter by obtaining the filter coefficients that relate to presenting the least mean squares of the error signal, e.g., difference between the desired and the real signal. A general approach is introduced in Refs. 67.–8. Let us assume is original image, that we reconstruct by filtering interlaced image with the linear, shift-invariant filter having unit-sample response , where is desired filter obtained by TS. If we suppose the remainder between and can be designed as a stationary random field, then we may select . Because a good random field model is hard to be modeled, we can minimize the substantial error on TS of representative images. Let us suppose that we have TS of original images. For every image in this TS, we can create sized interlaced image from the sized desired image . Assume we segment the TS into sub-images, where ’th sub-image is specified on the partial block . Assume that the desired filter is a two-dimensional (2-D) finite impulse response filter with area of support , where rectangular lattice . Then the desired filter can be acquired as solution to
We assume and be the number of filter coefficients to be decided and the number of samples in the sub-images. We create an column vector from the filter coefficients by examining the area in some fixed command. Correspondingly, we create an column vector from by examining the pixels of in a fixed manner. Eventually, we create an matrix from the components of as each column of matches an element examined in the same way as used to create ; this column is obtained by scanning the elements of for in the identical manner exploited to create . Finally, Eq. (1) can be described in matrix form as2) is a standard least-squares (LS) issue with solution9
Figure 1 shows the perspective view of frequency response of the (a) ; (b) ; (c) (LS filter obtained by image #1); (d) (LS filter obtained by other 23 images to deinterlace a image); (e) (LS filter obtained by whole 24 images); and (f) (LS filter obtained by image #2). In this letter, filters were designed.
Table 1 shows the PSNR results of 24 Kodak images. The gave the highest PSNR (however, this is infeasible method), and and gave the same result. This implies that the proposed method is not sensitive to TS. In Fig. 2, reconstructed results of the image #5 are shown. It can be noticed that the proposed method with various filters present less artifacts with respect to the other filter based methods.
PSNR (DB) results of different filter-based methods for various kodak images.
Table 2 shows PSNR performance of the proposed and other conventional deinterlacing algorithm systems from the literature, they are LCID,11 FDD,2 EMD,12 FEPD,13 MCAD,14 CASA,15 and DCAD.16 At first observation, it is noticed that our proposed method outperforms all other methods over the entire images in terms of PSNR.
PSNR (DB) results of conventional deinterlacing methods for various kodak images.
The proposed filter based deinterlacing has shown the lowest reconstruction error among published filter-based techniques on the reference Kodak dataset. This letter has demonstrated that the feasible filter is sufficient to achieve the full benefits of the technique, improving over the method by 3.46 dB. Moreover, the proposed filter is superior to CASA by 0.21 dB, which is the best edge-based method in literature.
This work was supported in part by the National Science Foundation of China (NSFC) under Grant 61001100, 61077009, and the Provincial Science Foundation of Shaanxi, China, under Grant 2010K06-15.