Television – Image signal processing circuitry specific to television – Noise or undesired signal reduction
Reexamination Certificate
2000-11-10
2003-12-02
Hsia, Sherrie (Department: 2614)
Television
Image signal processing circuitry specific to television
Noise or undesired signal reduction
C348S608000, C348S620000, C348S619000
Reexamination Certificate
active
06657676
ABSTRACT:
FIELD OF THE INVENTION
The present invention generally relates to the field of video image processing and, more particularly, to techniques for filtering noise from digital video pictures.
BACKGROUND OF THE INVENTION
The present invention is useful for filtering digital video sequences corrupted by high noise levels. Because of the particular importance of the Moving Pictures Experts Group (MPEG) standard in treating digitized video sequences, reference will be made to an MPEG2 system to illustrate an implementation of the present invention. Of course, those of skill in the art will appreciate that the invention may also be used in various systems for transferring video sequences based on different standards, as established from time to time. The main application of the invention is picture pre-processing before the MPEG2 or other standard coding. Even so, it is possible to use the invention outside a coding process, such as in a TV set for filtering the pictures to be displayed upon reception, for example.
Pre-processing pictures to be coded according to the MPEG2 standard provides great enhancement of the coding efficiency. For example, see A. van der Werf et al., “I.McIC: a single-chip MPEG-2 video encoder for storage”, IEEE Journal of Solid-State Circuits, Vol.32, n. 11, November 1997; and L. Yan, “Noise reduction for MPEG type of code”, Proc. IEEE Int. Conf Acoust., Speech, Signal Process., 1994. Additionally, many filtering techniques of pictures have been developed. See J. C. Brailean, R. P. Kleihorst, S. Efstratiadis, A. K. Katsaggelos and R. L. Lagendijk, “Noise reduction filters for dynamic image sequences: a review,” Proc. IEEE, vol. 83, pp. 1270-1292, September 1996. In many cases, burdensome aspects of these techniques may have include high calculation power requirements or the need for several iterations on the same picture. As a consequence, these methods are either too expensive or too difficult to implement for real-time video applications.
Prior art real-time video filtering techniques (see, e.g., E.Dubois and S.Sabri, “Noise reduction in image sequences using motion-compensated temporal filtering”, IEEE Trans. on Communications, vol. COM-32, pp. 826-831, July 1984) typically base the filtering operations on the separation distances among the unfiltered grey level of the pixel to be filtered (P
o
) and the grey levels of other pixels of a working window. The effectiveness of the filtering in this case is inversely proportional to the size of these distances. As a consequence, the pixels corresponding to such large distances are not involved in the filtering operations. This is a coarse method of segmenting the picture that avoids the calculation of the average of pixels belonging to different objects. This operation would cause an out of focus of the picture and, in the case of a temporal filter, also the appearance of “tails.”
The document WO 97/30545 discloses a motion-compensated recursive filter based on the above method. Such a filter, depending on the absolute value of the difference between the current pixel and the corresponding pixel in the preceding picture, establishes the value of a coefficient or fraction &bgr;. The filter further combines the fraction &bgr; of the current pixel with a fraction 1-&bgr; of the corresponding pixel in the preceding motion-compensated picture. Such a system is substantially an infinite impulse response (IIR) adaptive filter of the first order. This technique has the drawback of leaving unchanged the current pixel if it is too different from the corresponding pixel of the preceding picture, even if such difference could be due to noise.
A spatial adaptive low pass filter whose adapting mechanism is based on the calculation of the difference among P
o
and the other pixels of the working window is disclosed in EP 878,776 A1. A local evaluation of certain parameters of the picture is carried out (for example, whether the considered pixel belongs to a uniform zone or not). On that basis it may be determined whether the pixel to be filtered is near to or far from the other pixels, by way of a fuzzy logic process. Near pixels are given more weight than far pixels in the calculation of the value of the filtered pixel. Yet, even in this case the effectiveness of such a filter in the presence of high noise levels is not high. This is because the decision that is made to assess whether the considered pixel belongs to a uniform zone (and therefore if it is possible to filter more) or not is based on distances among P and the surrounding pixels.
In the article entitled “A method of noise reduction on image processing”, IEEE Transactions on Consumer Electronics, Vol. 39, N. 4, November 1993, by S. Inamori et al., a temporal filter is described that is based on the above-described filter of Dubois and Sabri but which is more sophisticated. That is, the input-output characteristic is adaptively chosen as a function of the expected power of the noise. The filter is turned off by signals coming from a motion detection section to avoid the generation of trails and by an edge detection block to prevent causing the picture to be out of focus. Even so, this filter suffers similar drawbacks to the previously mentioned prior art filters and must be turned off in the presence of noise peaks.
These known techniques become unsatisfactory in the presence of high noise levels. It becomes necessary to turn off the filter every time a noise peak (“spike”) is superimposed on the pixel P
o
being processed. High noise peaks are noise samples of a value much higher than other samples. Although these noise peaks have a rather small probability because the queues of the stochastic distribution of many kinds of noise (e.g. Gaussian noise) are theoretically infinite, they occur nonetheless. In such cases, the above mentioned prior art filters always interpret higher than normal differences among the current pixel and the neighboring pixels as if the picture is not stationary and not as if the picture is corrupted by noise. Therefore, the performance of noise reduction of the filters is significantly limited.
A well known technique for selecting the pixels of the working window is the so called “Duncan Range Test” (DRT) (see D. B. Duncan, “Multiple range and multiple f-tests”, Biometrics, vol. 11, pp. 1-42, 1955) which has been used with good results. For example, see R. P. Kleihorst, “Noise filtering of image sequences,” Ph.D. Thesis TU-Delft, Information Theory Group, 1994. In order to highlight the characteristics of the DRT, a brief description follows. Suppose a set of data, e.g., the luminance values of the pixels to be filtered, is organized in an increasing order as follows:
g
(l)
≦g
(2)
≦ . . . ≦g
(n)
where the number in parentheses indicates the position assumed in the ordered set. The DRT is based on the definition of a set of “similar” values, meaning that two values are considered “similar” if:
&LeftBracketingBar;
g
(
i
)
-
g
(
j
)
&RightBracketingBar;
σ
n
≤
ρ
&LeftBracketingBar;
i
-
j
&RightBracketingBar;
,
a
where &sgr;
n
is the standard deviation of the superimposed noise that is considered known. &rgr;
|i-j|,a
is a value that depends on the precision level a of the test, by the number |i-j| of data, that in the specific case is a number of pixels between g
(i)
and g
(j)
, and by the stochastic distribution of the noise. &agr; represents the probability of taking a wrong decision. Obviously, the smaller &agr; is, the smaller &rgr; is, and so the more restrictive the test is.
With these premises, the object of the DRT is to find the sub-set including the current pixel P
o
and including the greatest number of “similar” pixels. In other words the following sub-set of pixels must be determined:
g
(i)
, . . . , P
o
, . . . , g
(j)
where i and j are such that:
j
-
i
=
MAX
i
′
,
j
′
(
j
′
-
i
′
)
where
j
′
and
i
′
satisfy
the
&emsp
Borneo Antonio Maria
Salinari Lanfranco
Allen Dyer Doppelt Milbrath & Gilchrist, P.A.
Hsia Sherrie
Jorgenson Lisa K.
STMicroelectronics S.r.l.
LandOfFree
Spatio-temporal filtering method for noise reduction during... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Spatio-temporal filtering method for noise reduction during..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Spatio-temporal filtering method for noise reduction during... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3164454