UCRMD-1272(.K) UrsulaG&M&n

Transcription

1 UCRMD-1272(.K) nterlace. Restoration UrsulaG&M&n Thi$icmirlfomd repatintcmkdpdmiuily forintan81ar limitcdextcmml. Thcopiniomandoonclusionstatedaetboscoftheauthoraldmayor 7 neyootbethoseof theabowq. Wmk~bti~~oftiU3.~ofW~byti LawltnccLivcnnorcNationalbbwatq undwcomjww-7405-eng-48. b

2 DSCLAMER This document was prepared as an account of work sponsored by an agency of the United States Government. Neither the United States Covernmen t nor the Univemity of California nor any of their fm the accuracy, completene41s#or ua&heaa of any information, apparatus, product, m process disclti or represents that its use would not infringe privately owned rights. Referance herein to any apedfic commercial product, process, or service by trade name, trademark, manufacturer, or otherwise, doea notneceamhly constitute or imply ita end orsement# recommendation, or favoring by the United States Government or the University of California. Theviews and opinions of authors expressed herein do not ntxesadly state or reflect those of the United States Government or the UNversity of California, and shall not be used for advertising or product endorsement ptqoaea. employees, makes any Warran% ~ ~ ~@e4 m SSSUmSSanylegalMilkyorSSpOn.9ibuity Thiareporthasbasn reproduced directlyfromthe best available copy. Available to DOE and DOE amtracbm fromthe Off& of %ientific and Tdmical nformation P.o. Box62,oak m Pxiceavatlable from (61S) v 68401, F Available to the public from the National Technical nfmnah onservice us. -ant ofcommarce sz8s PortRo Rd., SpringfieldPA 22161

3 nterlace Restoration LDRD Final Report for 96-ERD-066 Ursula Goldstein As recorded video becomes more common as a method of capturing information, the ability to extract a high-quality still image from a videotape is becoming more important in areas such as intelligence analysis, law enforcement, surveillance, and monitoring. Unfortunately, a single image frame extracted from a videotape has a lower perceived quality than the full motion video because of the human eye s ability to integrate information over time. The conventional method used to overcome this problem is to digitize 10 to 20 successive frames from the videotape, spatially align as many of these frames as possible, and then average them in order to increase the signal-to-noise ratio. This technique produces excellent results for firstgeneration videotapes. n practice, however, it is often necessary to work with videotapes that have already been copied several times. The recopying process degrades the synchronization signal on the tape so that an image display/frame-grabbing device camot achieve a proper horizontal alignment of the image rows. Since the common video formats (NT!5C, PAL, and SECAM) scan the odd-row field of an image separately from the even-row field and then interlace them, a horizontal shift in a small section of one of these fields produces a row-by-row misalignment that is visually very objectionable in the interlaced image. The phenomenon varies over the vertical extent of the image and can be improved only slightly with a timebase corrector. Because this interlace jitter on the digitized image makes it impossible to produce a good spatial alignment between successive images, the frame-averaging technique for producing high-quality images gives unsatisfactory results. Since most researchers in video processing work with first-generation videotapes or predigitized sequences, there has been very little interest in correcting these subtle row-alignment problems. The goal of this project was to develop a digital-image restoration method that would make it possible to produce a high-quality still image from a videotape with a degraded synchronization signal. The general approach we 1

4 used to align successive rows of an image was to compute cross-correlation functions between adjacent row pairs and then compute the shifts needed to align each row to sub-pixel accuracy. The location of the peak of the crosscorrelation function, determined to subpixel accuracy by means of a parabolic fit to the main lobe, was used as the relative shift between two rows. The absolute shifts for each row were computed by summing the relative shifts from the top of the image to the current row. n order to validate this approach, we first did some experiments to characterize the nature of the degradations present in the original images. A comparison of row shifts computed on different subsections of the image showed that row misalignments were indeed constant over the entire horizontal extent of the image and therefore probably not produced by a tape stretching or playback speed problem. There was some correlation between computed row shifts of one image and a subsequent image, but not enough to allow this phenomenon to be exploited during restoration. f applied directly to the image, the shifts computed by the cross-correlation technique spatially distort the image by forcing near-vertical features in the image to be vertical. n order to prevent this distortion from taking place, the sequence of shifts (one for each row) was formed into a signal and filtered (see Figs. 1 and 2). Since scene-dependent phenomena vary slowly across the image, a zero-phase high-pass filter can suppress these effects while retaining the high-frequency interlace jitter information. With the sampling frequency normalized to 1, a normalized cutoff frequency in the range of 0.15 to 0.25 appeared optimal. The data used to explore this problem consisted of several image sequences digitized from videotape into CCR-601 digital video format. This format stores intensity information separately from the color information and at twice the data rate. Since the intensity of a video image has more than twice the spatial resolution of the color information, we used the intensity values to compute the required shifts and then applied these shifts to both the intensity and the color information. ndividual rows were shifted to subpixel resolution using a bandlimited interpolator. The resulting image (Fig. 4) shows considerably less interlace jitter than the original (Fig. 3). 2

5 1 1 5 o -5 1 i 1 o x 10+2 Figure 1. Absolute row shifts before filtering # b -0.4/ o x 10+2 Figure 2. Absolute row shifts filtered with a high-pass filter with cutoff at

6 Figure 3. Magnified of-interlace Figure j;tter 4. Magnified subsection of original that are most easily subsection image. Arrows seen on th e vertical of image processed 4 point to regions p ipes. to remove interlace jitter.

7 Technical nformation Department Lawrence Livermore National Laboratory University of California Livermore, California 94551