FIELD OF THE INVENTION The invention relates to a method and arrangement for adding field images of an interlaced video frame image received in the form of frame-encoded transform blocks obtained by an image transform.

The invention also relates to a method and arrangement for detecting a watermark embedded in the fields of a plurality of interlaced video frames.

BACKGROUND OF THE INVENTION Some video processing applications require the two fields of an interlaced video frame to be added or averaged. An example of such an application is watermark detection. International Patent Application WO-A-99/45705 discloses a video watermarking system in which the same watermark is embedded in successive fields of a video signal. The watermark detector of this system accumulates the fields over a number of frames so that the video signal averages to zero, whereas the watermark adds constructively.

Adding (or averaging) the two fields of an interlaced video frame is trivial in the pixel domain, but far from trivial in the digital (MPEG) domain. The reason is that MPEG encoders may have joined the fields of a frame prior to encoding, and produce so-called frame_pictures with DCT blocks containing information from both odd and even fields.

Figs. 1 and 2 show diagrams to illustrate the problem underlying this invention. Fig. 1 shows one of MPEG's modes of encoding pictures. In this encoding mode, known as frame-pictures with frame encoded macroblocks, the two fields 11 and 12 are joined together to a frame 13 by interleaving the lines from the two fields. Both fields have the same watermark W embedded. The frame is then subjected to a discrete cosine transform (DCT), which transforms blocks of 8x8 pixels into blocks of 8x8 coefficients. Four DCT blocks collectively constitute a macroblock 14. Each DCT block represents half of the pixels from the first field 11 and half of the pixels from the second field 12.

Because the DCT is a linear transform, the effect of adding DCT-coefficients is the same as adding the corresponding pixels. Accordingly, the accumulation of frames

carried out by the watermark detector may be performed in the DCT domain. The inverse DCT may be postponed until after the accumulation is completed. However, the accumulation requires a frame-based memory. Accordingly, the watermark detector of the system disclosed in WO-A-99/45705, in which 128x128 watermark patterns are tiled over each field, requires a 256x128 buffer.

Another trivial way to add the two fields is to perform the inverse DCT on every block as it arrives, then add the odd lines to the even lines, and store the result in a memory. This option only needs a field-based memory (i. e. a 128x128 buffer in the watermark detector), but requires an on-the-fly inverse DCT transform on every DCT block, which is neither attractive from an implementation point of view.

Fig. 2 shows another one of MPEG's modes of encoding pictures. In this encoding mode, known as framewictures withfield_encoded macroblocks, the two fields 11 and 12 are joined together to a frame 15 by taking 8 consecutive lines from the first field 11, followed by the same 8 lines from the second field 12. The frame is DCT-transformed. Four DCT blocks collectively constitute a macroblock 16. In this encoding mode, a macroblock contains blocks from one field and blocks from the other field, but all of the pixels represented by one DCT block are from the same field. Because the DCT is a linear transform, the effect of adding DCT-coefficients is the same as adding the corresponding pixels. Accordingly, the two vertically adjacent DCT blocks of each macroblock 16 may be added together in the DCT domain. This operation requires a field-based memory. The inverse DCT may be postponed until after the accumulation of all frames is completed.

However, this straightforward solution cannot easily be combined with the above-mentioned solutions for adding the fields of frame-encoded macroblocks.

In practice, MPEG frame pictures contain a mix of frame-encoded macroblocks and field-encoded macroblocks. The majority (70.. 85%) of the macroblocks is frame-encoded. The technically most awkward situation is thus also the most common.

OBJECT AND SUMMARY OF THE INVENTION It is an object of the invention to provide a method and arrangement for adding fields of an interlaced video image frame, with which the above-mentioned problems are alleviated.

To achieve these and other objects, the method in accordance with the invention comprises the steps of multiplying the frame-encoded blocks with a sparse matrix

which is representative of the inverse image transform, field addition, and a predetermined basis transform ; and subjecting the result of said multiplication to the inverse of said predetermined basis transform.

The invention exploits the mathematical insight that the operations of inverse DCT and field addition may be followed by a basis transform which is subsequently undone by the inverse of said basis transform. The composed operations (inverse DCT, field addition, and basis transform) are now physically replaced by a single matrix multiplication. The basis transform is chosen to be such that said matrix multiplication is a multiplication with a sparse matrix, i. e. a matrix with few non-zero elements. The sparse matrix multiplication is carried out on-the-fly, but is much easier to implement than an on-the-fly inverse DCT. The method requires a field-based memory only. A further significant advantage of the invention is that execution of the inverse basis transform can be postponed until after all frames (20 or so for watermark detection) have been accumulated in the field memory.

BRIEF DESCRIPTION OF THE DRAWINGS Figs. 1 and 2 show diagrams to illustrate the problem underlying the invention.

Fig. 3 shows a schematic diagram of an arrangement for accumulating frame- encoded macroblocks in accordance with the invention.

Fig. 4 shows a schematic diagram of an embodiment of an arrangement for accumulating a mix of frame-encoded blocks and field-encoded blocks in accordance with the invention.

DESCRIPTION OF PREFERRED EMBODIMENTS Fig. 3 shows a schematic diagram of an arrangement for accumulating a plurality of frame-encoded macroblocks according to the method of the invention. The arrangement receives an 8x8 frame-encoded DCT block 31 (i. e. one of the blocks from macroblock denoted 14 in Fig. 1). The DCT block is multiplied 32 with an 8x4 sparse matrix V2. This multiplication yields an 8x4 block of intermediate values, which is indicative of the sum of the two fields. An adder 33 and a memory 34 accumulate as many intermediate blocks as necessary for the application in question (e. g. watermark detection). Upon completion of the accumulation, the accumulated blocks are subsequently subjected to an

inverse basis transform by multiplication 35 with a matrix Uo. Finally, the actual application (here watermark detection 40) is carried out.

Note that the accumulation memory 34 in Fig. 3 is an 8x4 memory (in the Figure, each element has been drawn in proportion with the actual matrix size, e. g. 8x8,8x4, or 4x4). In practice, two memories 34 are required, one for accumulating the 4 lines derived from the upper DCT block of a macroblock, and one for accumulating the 4 lines derived from the corresponding lower DCT block. Collectively, they constitute a field-based memory (i. e. a 128x128 buffer in the case of watermark detection).

As has been attempted to show with dashed lines in Fig. 3, the matrix V2 represents a combination of three mathematical operations. The first and second operations are an inverse DCT 321 and a summation 322 of the two fields, respectively. As described in the introductory paragraphs, it is not attractive to physically carry out the inverse DCT on each received DCT block. The arrangement avoids this by mathematically subjecting the result of the two operations 321 and 322 to a basis transform 323. This basis transform is denoted Uol in Fig. 3. The unconventional notation Ut'ils used in this patent application for the basis transform itself, whereas the notation Uo is used for the inverse basis transform. The reason is that the basis transform 323 is only a mathematical notion, whereas the inverse basis transform 35 is physically executed by the arrangement.

The invention also advantageously exploits the insight that the inverse DCT 321 needs to be carried out in the vertical direction only. The inverse horizontal DCT (necessary if the application 40 needs to process the accumulated fields in the pixel domain), can be postponed until after completion of the inverse basis transform 35. Accordingly, the inverse DCT operation 321 is performed by multiplying the DCT block 31 with a matrix D81. The latter matrix is the inverse of the well-known 8-point DCT: The matrix S in Fig. 3 is the matrix representation of the summation of odd and even lines: The matrix V2, which physically replaces the three matrices D8', S, and Uo'can be mathematically expressed as:

V2 =Uol SD8 The basis transform Uol can be chosen arbitrarily, provided that an inverse transform Uo exists. The invention resides in selecting a basis transform such that the matrix V2 is sparse, i. e. has many zeroes. The inventors have found that the basis transform Uol=D4 is a clever choice, where D4 is the 4-point DCT (see equation 1). This choice results in the following matrix:

The matrix Va is extremely sparse. Multiplication with DCT block 31 requires only one multiplication per DCT coefficient, which can easily be done on-the-fly. It can be shown that a matrix having fewer non-zeroes than this one does not exist.

As already mentioned in the introductory paragraphs, MPEG's frame-encoded pictures are generally composed of a mix of frame-encoded macroblocks (Fig. 1) and field- encoded macroblocks (Fig. 2). It is not possible to directly add field-encoded blocks to the V2-transformed frame-encoded blocks accumulated in memory 34, because they live in different bases. This causes a problem when different blocks of the same frame need to be added together, or when a plurality of frames need to be added together (both situations occur in the watermark detector mentioned in the introductory paragraphs).

Fig. 4 shows a schematic diagram of an embodiment of the arrangement which is arranged to accumulate a mix of frame-encoded blocks and field-encoded blocks in accordance with the invention. In this embodiment, field-encoded DCT blocks 36 are transformed to the same basis as the Va-transformed frame-encoded DCT blocks 31 by multiplying 37 them with a further matrix Vl. For consistency, the matrix Vl must be:

where D4 is the 4-point DCT transform (see equation (1) ) and 04 is the 4x4 0-matrix.

Unfortunately, the matrix Vl is not considerably sparse. However, it can be approximated by:

which is sparse in the sense that it requires one multiplication per DCT coefficient (all multiplications are identical up to a sign).

Using the Vi-matrix of equation (4) instead of the proper definition in equation (3) captures 85% of the energy in field-encoded macroblocks, assuming a uniform distribution of DCT coefficients.

The inventors have found that it is possible to do better than this by slightly modifying the basis transform Uo such that more field-encoded energy is captured, although as a consequence thereof some frame-encoded energy is lost. Assuming that there is significantly more energy in frame-encoded macroblocks than in field-encoded macroblocks, the corresponding matrices V2 and V1 are:

where c4, is the same as in equation (2), and ai and bi are chosen in accordance with the video image statistics. With the following formulas, ai and b ; can be calculated for video sequences having a variance #i2 of the ith DCT coefficient of the columns of frame-encoded DCT blocks and a variance #i2 of the ith DCT coefficient of the columns of field-encoded DCT blocks: 1 bi = b8=i = a2i = ; i = 0,1,2,3<BR> 1 + (U(i+2)mod4,i)2<BR> <BR> <BR> <BR> <BR> <BR> <BR> <BR> <BR> <BR> <BR> + ( (i+2) mod 4, i)<BR> Gi i + Gi (i+2) mod4U (i+2) mod4,i : i = 0,1,2,3,<BR> <BR> <BR> <BR> <BR> <BR> 2i+1 = ; i = 0,1,<BR> 1 + (U(i+2)mod4,i)2 where:

Generally, ai and b ; are close to one, so that the matrices substantially resemble the ones defined in equations (2) and (4).

The invention can be summarized as follows. For some video processing applications, most notably watermark detection (40), it is necessary to add or average (parts of) the two interlaced fields which make up a frame. This operation is not trivial in the MPEG domain due to the existence of frame-encoded DCT blocks. The invention provides a method and arrangement for adding the fields without requiring a frame memory or an on- the-fly inverse DCT. To this end, the mathematically required operations of inverse vertical DCT (321) and addition (322) are combined with a basis transform (323). The basis transform is chosen to be such that the combined operation is physically replaced by multiplication with a sparse matrix (32). Said sparse matrix multiplication can easily be executed on-the-fly. The inverse basis transform (35) is postponed until after the desired addition (33,34) has been completed.