This presentation discusses the various techniques used for video de-interlacing.

1. De-Interlacing Techniques Ramesh Prasad

2. Introduction De-interlacing is the process of converting interlaced video, such as common analog television signals or 1080i format HDTV signals, into a non-interlaced form. 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 matched the properties of CRT screens. However, most current displays are inherently progressive, so the two fields need to be combined to a single frame, which leads to various visual defects which the de-interlacing process should try to avoid.

3. Introduction Progressive Frame Odd Field Even Field

4. Introduction Interlacing is a form of “Spatio-temporal” sub sampling De-interlacing is the reverse operation, aiming at removal of sub-sampling artifacts For stationary pictures (no camera or object motion or intensity changes) the de-interlacing is trivial as alternating even and odd fields completely describe the captured scene However, for pictures with camera/object motion and intensity changes (which always is the case with real life scenario); de-interlacing becomes non-trivial and entails processing, as it requires interpolation of picture data that was never transmitted or even captured. A good de-interlacing algorithm should try to avoid interlacing artifacts as much as possible and not sacrifice image quality in the process.

5. Introduction

6. Techniques for De-Interlacing Intra field de-interlacing Inter field de-interlacing Spatio-Temporal Interpolation Motion adaptive de-interlacing Motion compensated de-interlacing

7. Intra field de-interlacing - Overview Based on Spatial filtering Exploits the correlation between vertically neighboring samples in a field when interpolating intermediate pixels They have all pass temporal frequency response which guarantees the absence of motion artifacts Low computational cost Loss of vertical details and artifacts when the object exists in only one parity field

8. Intra field de-interlacing – Scan line duplication (Bob) The missing line is generated by duplicating the line directly above it Cons- Jagged effect will occur in the oblique line flicker Blur

9. Intra field de-interlacing – Scan line duplication (Bob) Jagged Effect Flicker Flicker

10. Intra field de-interlacing – Scan line interpolation The missing line is generated by taking the average of the lines vertically above and below it Less blur, flicker and jagging than scan line duplication method

11. Intra field de-interlacing – Edge Line Average (ELA) Interpolation is done in the direction of edge Three pixels in the previous scan line and the next scan line are referenced to determine the obvious edge in the image This method can eliminate the blurring effect of the bilinear interpolation and gives sharp/straight edges.

12. Intra field de-interlacing – Edge Line Average (ELA)

13. Inter field de-interlacing - Overview Based on Temporal filtering Exploits the correlation between temporally neighboring samples between fields when interpolating intermediate pixels They have all pass spatial frequency response which guarantees the absence of artifacts for stationary pictures Low computational cost but require additional field memory Artifacts in areas where motion occurs

15. It is best solution in case of still images as all vertical frequencies are preserved

16. If there is motion between the even and odd fields being woven; artifact called serration or mouse teeth occurs because the moving objects are not shown at the same position for even and odd lines of the single output frameOdd Field Even Field

17. Inter field de-interlacing – Field Insertion (Weave) Difference between Bob and Weave Bob Weave

18. Inter field de-interlacing – Field Insertion (Weave) Mouse Teeth Artifact

19. Inter field de-interlacing – Temporal Averaging Types - Mean – Display a half-picture that is created as the average of the two original half-pictures The Mean algorithm simply pairs the original lines, and averages each pair into one output line. Line 1 of output is the mean of lines 1 and 2 in input, line 2 of output is the mean of lines 3 and 4 in input, and so on. Blend – Each line of the picture is created as the average of a line from the odd and a line from the even half-pictures. This ignores the fact that they are supposed to be displayed at different times The first line of output is copied from the first line of input. For any other output line N, the line is the mean of input lines N and N-1. That is, the second line of output is the mean of lines 1 and 2 in input, the third line is the mean of lines 2 and 3, ... and finally, the last line of output is the mean of the last two lines in input. The sliding averaging procedure preserves the original vertical resolution. Pros – Mouse Teeth artifact is avoided Cons – Artifact called ghosting is introduced

20. Inter field de-interlacing – Temporal Averaging Ghosting Artifact

21. Spatio-Temporal Interpolation - Overview Based on Spatio-Temporal filtering Exploits the correlation between both spatially and temporally neighboring samples when interpolating intermediate pixels The filter is usually designed such that the contribution from the neighboring fields is limited to the higher vertical frequencies Hence, motion artifacts are absent for objects without vertical details that move horizontally The vertical detail from the previous field is being combined with the temporally shifted current field, so some motion artifacts may occur.

22. Spatio-Temporal Interpolation

23. Motion adaptive de-interlacing - Overview It combines the advantages of both intra-field and inter-field interlacing It detects the motion first, and then adopts intra field de-interlacing in motion areas and inter-field de-interlacing in static areas High resolution and flicker free picture can be realized in motion and static area It relies on accurate motion detection; erroneous detection cause artifacts

24. Motion adaptive de-interlacing

25. Motion adaptive de-interlacing The motion detection block detects the motion between the two fields The output of the motion detection block controls the selector switch for the intra or inter interpolated values For moving image parts, intra interpolated values are passed and for static image parts inter interpolated values are passed Finally an interleaving switch interleaves the current field and the interpolated sample to form the progressive frame The intra and inter interpolation mechanism could be any of the techniques described earlier

27. Fmot is the result of interpolation for moving image parts

29. Motion adaptive de-interlacing – Motion Detection

30. Motion adaptive de-interlacing – Motion Detection 4-Field motion detection To improve the accuracy of motion detection, 4 fields rather than 2 fields are used The fields are labeled backward, current, forward and forward-forward The differences are computed for fields of same parity; backward-forward and current-forward-forward

31. Motion adaptive de-interlacing – Motion Detection

32. Motion adaptive de-interlacing – Motion Detection

33. Motion adaptive de-interlacing – Implicit Adaptation (Median Filtering) Median Filtering adapts implicitly to motion The simplest version is a 3-tap spatio-temporal filter which uses the immediate vertical up, down and temporal neighbor in the previous field The underlying assumption is that in case of stationarity, the temporal neighbor is likely to have a value between that of the vertical neighbors of the current field In case of motion, intra-field interpolation is likely to result, since the correlation between the samples in the current field is likely to be highest Hence Median filtering automatically realizes the intra/inter switch Pros: Superior properties at vertical edges and low hardware cost Cons: Distortion of vertical details and introduction of alias

34. Motion adaptive de-interlacing – Implicit Adaptation (Median Filtering)

35. Motion compensated de-interlacing - Overview It’s the most advanced technique for interpolation Motion compensated method try to interpolate in the direction with highest correlation; with motion vector available, this is the interpolation along the motion trajectory It allows the moving sequences to be converted virtually into stationary ones and use methods which work well on stationary sequences like line insertion etc. Motion compensated methods needs to be combined with scene change detection, otherwise it will attempt to find motion between two completely different scenes Usually a combination of Edge directed spatial interpolation and motion compensated temporal interpolation is used in practice It outperforms other techniques in terms of quality of the de-interlaced frame It’s the most complex and processing intensive technique as it involves motion estimation

36. Motion compensated de-interlacing

37. Motion compensated de-interlacing Motion estimation is performed on the fields of same parity (backward and forward fields) as they have the same sampling grid Half pel/quarter pel accuracy can be used for motion estimation As the backward and forward fields are of different parity than the current field, the blocks need to be shifted to the sampling grid of the current field To obtain the position of the block in the current interpolated field from the previous field, the motion vectors are halved and the blocks of backward, current and forward field moved simultaneously to the sampling grid of the current interpolated field

38. Motion compensated de-interlacing Block shifting to match the sampling grid Pixel value in the current interpolated field

39. Motion compensated de-interlacing To handle scene change detection, the pixel values obtained by motion compensated interpolation are compared with the spatial edge directed interpolated pixel values If the difference is beyond the threshold value, its assumed that interpolation based on motion estimation is incorrect In such case, directional interpolation is used instead of motion compensated interpolation The motion compensated/edge directed interpolated pixels are finally combined to get the progressive frame

40. References Altera White Paper - High-Definition Video Deinterlacing Using FPGAs Deinterlacing - An Overview GERARD DE HAAN, ERWIN B. BELLERS Deinterlacing using directional interpolation and motion compensation Ohjae Kwon, KwanghoonSohn, Chulhee Lee Motion Adaptive Interpolation with Horizontal Motion Detection for Deinterlacing Shyh-Feng Lin, Yu-Ling Chang, and Liang-Gee Chen

41. Thank You