Inter-Frame Video Filter Having Selective Linear and Non-Linear Transfer Characteristics as Function of Inter-Frame Differences

1. Field of the Invention

The present invention concerns video filters of the nonlinear temporal type, particularly, but not exclusively, for use in video coding systems.

2. Related Art

Nonlinear inter-frame digital filters are utilized to reduce noise while minimizing the blurring of moving objects or scene changes. In effect, a greater degree of filtering is introduced for small inter-frame differences than for larger ones.

This invention provides look-up memory savings by restricting the non-linear filter characteristic to a small fixed portion of the filter transfer characteristic.

Some embodiments of the invention will now be described, by way of example, with reference to the accompanying drawings, in which:

FIG. 1 is a block diagram of a filter;

FIG. 2 illustrates a known characteristic for the nonlinear lookup unit 4 FIG. 1;

FIG. 3 illustrates a modified characteristic in accordance with one form of the present invention;

FIG. 4 is a block diagram of one form of lookup unit 4 in accordance with the invention; and

FIG. 5 is a block diagram of an alternative form of lookup unit 4.

A nonlinear temporal filter for digitally coded video signals has (FIG. 1) a video input 1. Each picture sample S is supplied to a subtractor 2 which takes the difference D between the new sample and the corresponding "previous picture" sample P from a frame store 3. The difference passes through a lookup unit 4 which forms a nonlinear function D'=f(D). The modified difference D' is added in an adder 5 to the previous picture sample P and the resulting filtered new sample S' is forwarded to an output 6 and also written into the frame store 3 to form the new previous picture sample for the next frame.

Hence, if f(D)=D then S'=S and no filtering occurs. If 0<[f(D)/D]<1 where f(D) has the same sign as D, then some filtering takes place.

The object of this temporal filtering is to reduce noise; however it has the disadvantage of causing blurring of moving objects on scene changes. Where the difference between the old and new samples is large, this is likely to be caused by changes in picture content, whilst a small difference is more likely to be caused by noise. Therefore the lookup unit 4 is given a nonlinear characteristic in order to effect a greater degree of filtering for small differences than for large ones. A typical characteristic curve is shown in FIG. 2, with a slope less than unity for small D, and asymptotic to the line D'=D.

Assuming a resolution of 8 bits/sample, the difference D (and modified difference D') has a range of -255 to +255, and therefore a memory of 512.times.9 bit capacity is required for the lookup table. Although this is not a large memory bu current standards, nevertheless for an integrated circuit video coder implementation where silicon area may be at a premium, it is desirable to reduce this. Accordingly, it is arranged that the characteristic is nonlinear for values of D less than 50% of its maximum absolute value, or more preferably less than 10%, and linear with unity slope for greater values.

For example, the characteristic in FIG. 3 has a unity slope for absolute values of D greater than or equal to 16. This can be implemented by the lookup unit shown in FIG. 4, where the four least significant bits of D form the address input to a read-only memory 10 having 16 locations each containing a four-bit word representing half of the characteristic depicted in the central region of FIG. 3. The data output from the memory 10 forms the four least significant bits of an 8-bit word (whose four most significant bits are always zero) forming one input to an electronic change-over switch or data selector 11. The other input to the selector 11 is the input word D (without its sign bit SGN).

The four most significant bits of D are connected via an OR gate 12 to the control input of the selector 11 so that whenever one of those bits is nonzero (i.e. /D/.gtoreq.16) the selector passes the data input D directly to the output D'; otherwise the output is the output of the lookup table 10. The sign bit of the output D' is of course the sign bit of the input D, unmodified. This structure assumes a sign-bit plus magnitude representation of negative values of D. In practice, a twos-complement representation is more convenient for implementing the addition and subtraction, in which case the more complex arrangement shown in FIG. 5 would be appropriate It differs from FIG. 4 in that (1) the read-only memory 10 covers a range from -16 to +15, with its additional address line fed from the sign bit of D, and (2) the range detector comprises an OR gate 14 which produces a logic zero output when the four most significant bits of D, and the sign bit are zero (i.e. D<16) a NAND gate 15 which produces a logic zero when these five bits are one (i.e. D>-17), and an AND gate 16 which produces an active low output when the output of one of the gates 14, 15 is zero.