权利要求

The present invention relates to a method and to an apparatus for motion compensated upconversion.

Background

The existing TV standards have a field frequency of 50 or 59.94 or 60Hz. When such pictures are displayed line flicker can arise in case of special picture content or large area flicker can arise in case of bright pictures, in particular on large screens. Such artefacts can be reduced when a display upconversion is performed in the receiver whereby the field frequency is doubled. However, simple upconversion algorithms like AABB can introduce new motion artefacts. To avoid such artefacts a motion compensation is carried out in the receiver which requires a lot of memory.

Invention

It is one object of the invention to disclose a method of motion compensated upconversion which requires less memory capacity, in particular less line memory capacity. This object is reached by the method disclosed in claim 1.

It is a further object of the invention to disclose an apparatus which utilises the inventive method. This object is reached by the apparatus types disclosed in any of claims 24 to 27.

Motion compensated upconversion requires motion estimation and motion compensated interpolation. A lot of line memories are needed in the motion estimator for motion vector processing. Since such line memories require a significant amount of chip area straight forward circuitry will lead to the need of two integrated circuits for motion compensated upconversion. According to the invention the memory capacity, in particular the number of line delays, can be reduced which allows to build a motion compensated upconverter in one integrated circuit only (apart from the field memories). This is reached by using a special line memory configuration in the motion estimator, taking into account the line memories in noise filter means, and by the use of main block motion vectors and refined subblock motion vectors being finally applied in the motion compensated interpolator.

One field memory is used for doubling the field frequency. This may happen prior to motion estimation or afterwards. Since prior to field frequency doubling the available field processing time is double one would normally try to perform motion estimation in this domain, resulting in cheaper low-speed circuitry.

A second field memory is used in connection with the motion estimator. A third field memory may be necessary for the motion compensated interpolation, which can be accompanied by a predictor, recursive filter means and appropriate processing time delays to perform on the high field frequency signals motion compensated interpolation together with a noise reduction.

The invention can be used for e.g. 50/59.94/60/50Hz interlaced to 100/119.88/120/75Hz interlaced upconversion, 50/59.94/60/50Hz interlaced to 100/119.88/120/75Hz proscan upconversion and 50/59.94/60/50Hz proscan to 100/119.88/120/75Hz proscan upconversion.

One advantage of the invention is that even when spatial and recursive temporal noise reduction are added to the motion compensated upconversion (MCU), in a further developed embodiment two field memories only are sufficient to perform all these functions. This can be achieved with special circuitry under the condition that the motion compensated interpolator receives partly the same video input signal as the motion estimator and that, once a wrong motion vector is calculated, the temporal filter will not propagate this error.

Tests have shown that the latter can be achieved due to the kind of control used in the temporal noise reduction and that the error will decay from field to field.

A further advantage of the invention is that motion vectors for large main blocks need to be stored only so that the vector memory can have a low capacity. This can be achieved by calculating from said main block motion vectors refined subblock motion vectors just before they are needed in the processing so that the subblock motion vectors need not to be stored.

This means that the main block motion vectors can be calculated during one output field period and the subblock motion vectors can be calculated, just in time for use in a next output field, during the repetition of the first output field.

Advantageously, the calculation of the motion vectors and/or the temporal filter can be controlled by a computed noise level.

In principle, the inventive method is suited for motion compensated upconversion of interlaced video signals, in which fields of source field frequency are converted to form fields of doubled field frequency and in which motion estimation is performed and motion compensated interpolation on said fields of doubled field frequency is performed, using motion information from said motion estimation, whereby in said motion estimation motion vectors for main blocks and motion vectors for subblocks being smaller than said main blocks are calculated for one input field and the pixels required for this calculation are stored in line memory means for one input field, whereas said motion compensated interpolation is carried out on said subblock motion vectors in subsequent output fields.

Advantageous additional embodiments of the inventive method are resulting from the respective dependent claims.

In principle the inventive apparatus for motion compensated upconversion of interlaced video signals includes:

• second memory means receiving input fields, the input and output of which is connected to a motion estimator which performs field-to-field motion estimation and contains line memory means for storing pixels which are used to calculate motion vectors for main blocks and motion vectors for subblocks being smaller than said main blocks for one input field;
• succeeding first memory means for converting the output fields of said second memory means into fields of doubled field frequency;
• third memory means the input and output of which is connected to a motion compensated interpolator which performs in subsequent output fields field-to-field motion compensation on output signals being derived from the output of said first memory means and which receives said subblock motion vectors from said motion estimator and which delivers the motion compensated upconverted output signals,

or includes:

• first memory means being used for converting source fields into fields of doubled field frequency;
• succeeding second memory means the input and output of which is connected to a motion estimator which performs field-to-field motion estimation and contains line memory means for storing pixels which are used to calculate motion vectors for main blocks and motion vectors for subblocks being smaller than said main blocks for one input field;
• third memory means the input and output of which is connected to a motion compensated interpolator which performs in subsequent output fields field-to-field motion compensation on output signals being derived from the output of said second memory means and which receives said subblock motion vectors from said motion estimator and which delivers the motion compensated upconverted output signals,

or includes:

• first memory means being used for converting source fields into fields of doubled field frequency;
• succeeding second memory means the input and output of which is connected to a motion estimator which performs field-to-field motion estimation and calculates and stores in memory means motion vectors for main blocks for one output field and thereafter calculates for a subsequent output field from said main block motion vectors motion vectors for subblocks being smaller than said main blocks;
• third memory means the input and output of which is connected to a motion compensated interpolator which performs in subsequent output fields field-to-field motion compensation on output signals being derived from the output of said second memory means and which receives said subblock motion vectors from said motion estimator and which delivers the motion compensated upconverted output signals,

or includes:

• first memory means being used for converting source fields into fields of doubled field frequency;
• succeeding second memory means the input signal of which is derived from the output signal of said first memory means and is fed to the first input of a motion compensated interpolator which performs field-to-field motion compensation on output signals being derived from the output of said second memory means and which receives motion information from a motion estimator and which delivers the motion compensated upconverted output signals, whereby the signal at the second input of said motion compensated interpolator is derived from the output signal of said second memory means;
• said motion estimator which receives the output signal of said second memory means at its second input and which performs field-to-field motion estimation, whereby the signal at the first input of this motion estimator is derived from the output signal of said first memory means and the motion estimator calculates and stores in memory means motion vectors for main blocks for one output field and thereafter calculates for a subsequent output field from said main block motion vectors motion vectors for subblocks being smaller than said main blocks.

Advantageous additional embodiments of the inventive apparatus are resulting from the respective dependent claims.

Drawings

Preferred embodiments of the invention are described with reference to the accompanying drawings, which show in:

Fig. 1

block diagram of a first embodiment of a motion compensated upconverter;

Fig. 2

block diagram of a second embodiment of a motion compensated upconverter;

Fig. 3

block diagram of a third embodiment of a motion compensated upconverter;

Fig. 4

block diagram of a fourth embodiment of a motion compensated upconverter;

Fig. 5

block diagram of a fifth embodiment of a motion compensated upconverter;

Fig. 6

basic block diagram of a motion compensated upconverter with two field memories;

Fig. 7

block diagram of a motion compensated upconverter with two field memories and noise reduction;

Fig. 8

main stage time diagram for the upconversion performed in Fig. 2 to 7;

Fig. 9

proscan grid produced in the motion compensated interpolator and estimator;

Fig. 10

horizontal prefilter and subsampling;

Fig. 11

main block and subblock;

Fig. 12

principle of two-sided block matching;

Fig. 13

small positional error for some motion vectors resulting from the subgrid used in the two-sided block matching;

Fig. 14

noise measurement zones;

Fig. 15

possible vectors in a periodic structure;

Fig. 16

current vector block and adjacent vector blocks in more detail;

Fig. 17

vertical motion vector component and vertical transition;

Fig. 18

subblock areas and related mainblocks, the motion vectors of which are used to generate the subblock motion vector;

Fig. 19

three-dimensional grid with subblock errors;

Fig. 20

simplified block diagram for noise reduction;

Fig. 21

first output field for motion compensated interpolation;

Fig. 22

third output field for motion compensated interpolation;

Fig. 23

second output field for motion compensated interpolation in a first mode;

Fig. 24

fourth output field for motion compensated interpolation in the first mode;

Fig. 25

second output field for motion compensated interpolation in a second mode;

Fig. 26

fourth output field for motion compensated interpolation in the second mode;

Fig. 27

soft switch and odd/even select;

Fig. 28

global timing diagram.

Preferred embodiments

The motion compensated upconversion process can be split into the following main stages:

• Motion estimation, horizontally and vertically, luminance only;
• Optionally, a noise reduction, spatially and temporally for luminance and temporally only for chrominance;
• Motion compensated interpolation, luminance and chrominance.

The motion estimator calculates the motion vectors denoted as MVs, the confidence denoted CL belonging to these MVs and one noise level per field denoted N. These results are then used to produce the motion compensated output pixels.

In the following it is considered that in the 50/60Hz domain:

• odd source fields = source fields 1, 5, ...,
• even source fields = source fields 3, 7, ...,

and that in the 100/120Hz domain:

• output field 1 ← from source field 1,
• output field 2 ← combination of source field 1 and source field 3,
• output field 3 ← combination of source field 3 and source field 5,
• output field 4 ← combination of source field 3 and source field 5.

The input and output samples are represented by 8 bit words, 0=LSB to 7=MSB.

Video input format (4:1:1):

• Y = 16 (black)...240 (white), blank value = 16
• U, V = -112...0...+112, blank value = 0 (2's complement)

Video output format (before DAC):

• Y = 0 (ultra black)...255 (ultra white), blank value = 16
• U, V = 0...128...255, blank value = 128.

In Fig. 1 a field memory or a FIFO FM1 is connected to two respective input of motion estimator ME and receives the interlaced source signal I. ME operates in the 50Hz domain and contains two delay arrays LD11 and LD12 each of which has 12 line delays. ME calculates motion vectors MV for 16 pixels * 8 lines main blocks MB in a horizontally subsampled proscan pixel grid as explained below. ME may further calculate subblock motion vectors SBMV for 2 pixels * 2 lines subblocks SB from the main block vectors MBMV. The required number of line delays depends on the line number of the main blocks, on the line number of the subblocks and on the possible range of the vertical motion vector components for both. In this example,

the MB line number is 8, the SB line number is 2 and the vertical component of both respective motion vectors is 1 which totals in a number of 12 line delays. Since these line delays store horizontally subsampled lines only they may have a length of 704/2 = 352 pixels.

Because MBMV and SBMV are calculated in the 50Hz domain but used in motion compensated interpolator MCI in the 100Hz domain they need to be stored in a vector memory RAM in ME. Since a source frame has e.g. 704 active pixels and 576 active lines and the original main blocks have a size of 32*16 there are (704/32=22)*(576/16=36) = 792 MBMVs and 792*(16/2=8)*(8/2=4) = 25344 SBMVs in the two source fields. As explained later a MBMV and a SBMV can be coded with 6 bits each. For each SBMV a confidence value CL is coded with e.g. 8 bits. Therefore, RAM must be able to store 792*6 + 25344*(6+8) = 359568 bits or 44946 bytes. This corresponds to a memory capacity of 44946/704 ≈ 64 full line delays.

The delay arrays LD11 and LD12 correspond to 12 full line delays. Thus, ME may need a memory capacity of at least 64+12 = 76 full line delays. In 0.5µm chip technology one full line delay requires ≈2mm² of chip area which would lead in this example to a total memory chip area in ME of at least 76*2 = 152mm². However, to have a good yield in the chip production (and thereby cheap chips) the maximum chip area should not exceed 85 to 90mm² . Thus, at least two integrated circuits (apart from field memories) are required to implement the circuitry of Fig. 1.

The output signal of FM1 passes through a 50Hz to 100Hz speed-up field memory SPUFM performing AABB upconversion directly to the first input of MCI and to the input of a further field memory FM2 which is required for the operation of MCI and the output of which is fed to a second input of MCI. The output signal of SPUFM may also first pass through spatial SF and/or temporal TF noise reduction filter means before entering MCI and FM2.

The output of ME sends motion vectors MV, confidence values CL and/or a noise measure N to the motion compensated interpolator MCI.

Advantageously, the temporal noise reduction filter means TF can be recursive filter means which combine the spatially noise filtered signal with the motion compensated output signal, namely the previous field, of FM2. This motion compensation is performed by a predictor PRED.

Advantageously, TF and MCI can be controlled by the noise level N and the confidence CL as explained below.

The output signal O of motion compensated interpolator MCI is the final motion compensated, upconverted and eventually noise reduced signal to be stored or displayed on a screen.

The processing times within the MCU circuit can easily adapted using appropriate delays, e.g. PTD1 and PTD2 in further figures.

The function of the various blocks will be explained later in connection with the other drawings.

In Fig. 1 to 5 and 7A the parallel bold dashes represent a line delay of full length (e.g. for 704 pixels) and the other parallel dashes depict a half line delay for storing subsampled picture data. The divided parallel bold dashes can be common line delays used in both, MCI and PRED or ME and SF.

These line delays are used to make available in the depicted diagram blocks such prior calculation results which are required when the pixels of the current line are output.

In Fig. 1 SF and MCI need each two full line delays LD13 and LD15. MCI also and PRED need each three full line delays LD16 and LD14 which can be common line delays. Thus, the MCU circuit of Fig. 1 may need apart from field delays a memory capacity of at least 76+2+2+3 = 83 full line delays.

Basically, the embodiment of Fig. 2 basically differs from that of Fig. 1 in that the 50Hz to 100Hz speed-up field memory SPUFM performing AABB upconversion is arranged prior to FM1 and ME and that ME does not store MBMV, SBMV and CL. MBMV and SBMV are calculated in the same field period wherein they are used in MCI. Because it takes some time until the subblock motion vectors SBMV are available a ten-line delay LD 28 is arranged between the output of SPUFM and the filter means SF or FM2, respectively.

The number of line delays in LD28 depends on the block sizes and on the maximum vertical SBMV component size.

A further one-line delay LD 27 is arranged in the MV / CL connection between ME and MCI. In case of even output lines EOL the MV information of the input signal of LD27 is passed via switch SW1 to PRED. In case of odd output lines OOL the MV information of the output signal of LD27 is passed via switch SW1 to PRED. LD27 is written in for one line and read out for two succeeding lines. The reason is that the same SBMV vectors are applied for two lines.

The number of line delays LD21 to LD26 corresponds to that of Fig. 1. This means that about 12+10+2+2+3+1 = 30 full line delays are required for the MCU circuitry of Fig. 2 which corresponds to a chip area of ≈60mm².

Basically, the embodiment of Fig. 3 differs from that of Fig. 2 in that the first ten line delays of array LD31 perform in addition the function of LD28. This means that these ten line delays need to be full line delays. The number of line delays LD32 to LD37 corresponds to that of Fig. 2. However, the total line delay capacity decreases by five full line delays.

Basically, the embodiment of Fig. 4 differs from that of Fig. 3 in that the two further line delays of LD41 are commonly used for the SF line delays LD43. This means that these two further line delays of LD 41 need to be full line delays, too. The number of line delays LD42 and LD44 to LD47 corresponds to that of Fig. 3. However, the total line delay capacity decreases by one full line delay.

In Fig. 1 to Fig. 4 all MBMV and SBMV vectors are calculated for each source field.

Basically, the embodiment of Fig. 5 differs from that of Fig. 4 in that the MBMV vectors are calculated in a first field whereas the SBMV vectors are calculated in a subsequent field directly before being used so that they need not to be stored. This necessitates on one side that the 792 MBMV vectors are stored with 6 bits each in a memory RAM in ME, but on the other side ME needs only two full line delays LD51 (commonly used for LD53 in SF) and two half line delays LD52. The number of line delays LD44 to LD47 corresponds to that of Fig. 4. However, the total line delay capacity decreases by 10*704 + 10*704/2 - 792*6/8 = 9966 bytes corresponding to ≈14 full line delays.

In Fig. 5 a field memory or a FIFO SPUFM is used to double the field frequency of input signal I. The output signal is fed to a motion estimator ME containing a low-capacity vector memory, to a second field memory or FIFO FM1 and to a spatial noise filter SF. A second input of ME is supplied by the output of FM1. The output of ME sends subblock motion vectors MV and some other information to a motion compensated interpolator MCI and to a first input of a predictor PRED. The spatially noise filtered output of SF is fed via a temporal noise filter TF to a first input of MCI and to a third field memory or FIFO FM2. A second input of MCI is supplied by the output of FM2. The output of FM2 is also fed to a second input of predictor PRED. The output of PRED is supplied to a second input of temporal noise filter TF. The output signal O of motion compensated interpolator MCI is the final motion compensated, upconverted and noise filtered signal to be displayed on a screen.

In Fig. 1 to 5 and 7A the parallel bold dashes represent a line delay of full length (e.g. for 704 pixels) and the other parallel dashes depict a half line delay for storing subsampled picture data. The divided parallel bold dashes can be common line delays used in ME, SF, MCI and PRED, respectively.

These line delays are used to make available in the depicted diagram blocks such prior calculation results which are required when the pixels of the current line are output.

Motion estimation chips like STi3220 of SGS-Thomson and two-dimensional interpolation filters like IMSA110 of SGS-Thomson and MPEG2 decoder chips containing a two-dimensional motion compensated interpolator like STi3500 of SGS-Thomson are already on the market.

The different MCU arrangement depicted in the basic block diagram of Fig. 6 needs two field memories only and shows a first field memory or a FIFO SPUFM, a second field memory or FIFO FM, a motion estimator ME and a motion compensated interpolator MCI. Contrary to Fig. 5, the output of SPUFM is connected via a first processing time delay PTD1 to the first input of MCI and to FM. PTD1 represents the processing time needed in ME. The output of FM is connected to the second input of ME and via a second processing time delay PTD2 to the second input of MCI.

The delay times of PTD1 and PTD2 are equal. The delay time of FM is one field minus the delay time of PTD1. This means that between the first and second input of ME and between the first and second input of MCI, respectively, a total delay time of one field (e.g. 10ms) is achieved.

The block denoted PTD1 may also contain a noise reduction circuit which receives an input signal from the output of PTD2 and motion information from ME.

Fig. 7A shows the basic circuitry of Fig. 6 in more detail. In Fig. 7A, the block PTD1+NR of Fig. 6 includes in series a first line delay LD1, a processing time delay PTD1, a spatial noise filter SF and a temporal noise filter TF. The output of FM is passed through a second line delay LD2, the first input of a second switch SW2 and PTD2 to the second input of MCI and to the second input of a predictor PRED. The second input of switch SW2 is connected to the output of FM.

Motion vectors MV and e.g. information on the noise level N and on the confidence CL of the vectors is passed through a third line delay LD77 from ME to MCI. The input of LD77 is connected to the first input of a first switch SW1 and the output of LD77 is connected to the second input of a switch SW1. The output of SW1 supplies motion vectors MV to the first input of PRED. The output of PRED is connected to the second input of temporal noise filter TF. In case of even output lines EOL the first input of SW1 is active whereas in case of odd output lines OOL the second input of SW1 is active. In case of fields 1 and 2 the first input of SW2 is active whereas in case of field 3 and 4 the second input of SW2 is active, i.e. connected to the output of SW2.

Like in Fig. 5 the motion estimator ME includes a vector memory RAM which stores main block motion vectors and can store in addition, like in Fig. 5, main block zero vectors the function of which will be explained below. Line delays LD71 to LD77 have the same number and function as in Fig. 5.

In Fig. 6 and 7A, ME receives a non-noise reduced input signal and a noise reduced input signal on which previously calculated motion vectors have already been applied, whereas MCI receives two noise reduced input signals. FM has a delay of one field minus the delay of TF, e.g. two lines.

Fig. 7B shows such output fields (with line numbers) in signal O which are present in the input signal I. However, fields F1 and F5 are both first fields in the O output frames.

Luminance processing

The source signal I, 50/60Hz/2:1/625, is sped up by means of a field memory to a 100/120Hz/2:1/625 AABB field repetition format. The time diagram of Fig. 8 shows the odd (o) / even (e) fields F1, F3, F5, ... of signal I, the sped-up (SPU) fields F1, F1, F3, F3, F5 ... and the fields F1, F2, F3, F4, F5, ... of output signal O.

An intermediate step in the processing is a proscan conversion the grid of which is depicted in Fig. 9. In the motion estimator ME (Fig. 9A) only the luminance odd source fields are vertically interpolated in order to place the lines of the odd and the even field onto a common vertical grid.

In the motion compensated interpolator a proscan grid (Fig. 9B) for both source fields and for luminance as well as for chrominance is produced by means of vertical filters.

In Fig. 9 the following symbols are used:

X

source grid;

○

lines introduced by means of vertical proscan interpolation;

●

spatial position of the output lines.

F1 is a odd (o) source field whereas F3 is an even (e) source field.

As depicted in Fig. 10A for the odd field F1 and in Fig. 10B for the even field F3, in the motion estimator both fields are now horizontally pre-filtered and subsampled. The grid so produced corresponds to the vertically V and horizontally H subsampled proscan grid shown in Fig. 9A. The horizontal pre-filter may have the coefficients (1/16, 2/16, 3/16, 4/16, 3/16, 2/16, 1/16) and is carried out on luminance only. Filtering can be done by first adding, followed by truncation.

In Fig. 10 the following symbols are used:

X

source grid;

X or ○

proscan grid. The ○ values can be calculated by vertically averaging of X values;

□

vertically and horizontally subsampled proscan grid (= 1st, 3rd, 5th, ... active pixels). This is the grid used for motion estimation, confidence and noise level calculation;

⊂⊃

subsampling filter. The center pixel denoted by a larger circle is the location for the output pixel.

Fig. 11 depicts within the vertically and horizontally subsampled proscan grid a main block MB and a subblock SB. The latter may have a size of 2*2 pixels. A non-subsampled main block has the size of 32 pixels * 16 lines in a frame and the size of 32 pixels * 8 lines in a field. A main block has the size of 16 pixels * 8 lines in the horizontally and vertically subsampled proscan grid.

Advantageously, the motion estimator can perform a two sided block matching on large luminance main blocks denoted MB. A main block may have a size of 16 pixels * 8 lines. Two-sided block matching is described in more detail in PCT/EP94/02870 of the applicant.

In the motion estimator each possible MV is applied to each MB. For each motion vector an error E is calculated by accumulating the absolute differences (two's complement) of the corresponding pixels within the main block of the previous field and current field along the motion direction. Then the vector giving the best match, i.e. the smallest error, is chosen as the main block motion vector, denoted MBMV.

If there are several equal smallest errors, i.e. several equally good motion vectors, the vector is chosen in the following order:

1st

MVs with the horizontal component closest to zero;

2nd

MVs with the horizontal component positive;

3rd

MVs with the vertical component closest to zero;

4th

MVs with the vertical component positive.

MV consists of the horizontal component MVH and the vertical component MVV. The horizontal component is positive when the motion on the screen is from left to right. The vertical component is positive when the motion on the screen is from top to bottom. Fig. 12 depicts the two sided block matching within the specified search windows SWW in the previous field PF (z-1) and in the current field CF (z) whereby for each subsequent main block MB in the field to be interpolated a motion vector MV is generated.

Due to the subgrid used a small positional error of one half subgrid pixel distance can arise for some motion vectors (-14, -10, -6, -2, +2, +6, +10, +14) as can be seen in Fig. 13B. This small error is acceptable.

Fig. 13A shows the vertical motion vector components MVV: +2, 0, -2. In principle, more values for MVV are possible when more line delays are spent. Since line delays are expensive (require large chip area) tests have been carried out which show that for upconversion purposes these three values are sufficient.

Fig. 13B shows the possible horizontal motion vector components MVH. The symbols used have the following meaning:

□

vertically and horizontally subsampled proscan grid;

○

correct position to calculate motion vectors;

●

slightly wrong position to calculate motion vectors.

One feature of the invention is to use a noise measurement for controlling the motion compensated upconversion. One example of noise measurement which could be used in the invention is described in EP-A-0562407 of the applicant.

A further noise measurement method will be explained in connection with Fig. 14. In TV standards B and G the active picture ACP has a height of 36 MBs, in TV standard M a height of 32 MBs. One central zone PNL0 or e.g. three (due to a possible picture-in-picture feature) zones PNL1, PNL2 and PNL3 may be used to measure the noise. First, the minimum main block error for each zone is calculated and, second, the three minima are median filtered to give the preliminary noise level PNL. In the current source field z is

PNL = MEDIAN [PNL1, PNL2, PNL3]

PNL is divided by 16:

PN_(z) = PNL/16

The final noise level N_(z) is formed by a further median filtering:

N_(z) = MEDIAN [N_(z-1)+2, [PN_(z)+PN_(z-1)]/2, N_(z-1)-2]   {truncation}

Hence, the noise estimate is updated once per input field and may change from field to field by a maximum of ±2.

Since N_(z) is not known until the end of the current active field, the N_(z) can be updated during the blanking period following the calculation of the MBMVs. The N_(z) is used by the adaptive temporal noise reduction. In case of 11 bit word length N_(z-1) is limited to between 2 and 2045 to avoid underflow and overflow when calculating above median value. N_(z) or part of its bits and can be readable via I2C bus.

Power-up initialisation: PNL, PN_(z), PN_(z-1) = 0, N_(z-1) = 2.

A further feature of the invention is to detect and to take into account periodic structures for controlling the motion compensated upconversion, i.e. for correcting the motion vectors having been found by the two-sided block matching. One example of a periodic structure detection which could be used in the invention is described in EP-A-94115732 of the applicant.

In periodic structures the motion estimation is very hazardous. There are several vectors giving a very good match, i.e. a very small error. Therefore a periodic structure correction, preferably in horizontal direction, is carried out. From the block matching process a corresponding error for every possible main block motion vector is known. Theoretically, a periodic structure in horizontal direction results in several error minima within a row, as can be seen from Fig. 15A to D, but also in several error maxima within the same row. In Fig. 15 the upper row of pixels depicts line n of source field (z) and the lower row line n of source field (z-1). Fig. 15A, B, C, D, respectively, show a first, second, third and fourth possible motion vector for the periodic picture brightness signal (e.g. a fence) depicted in lower Fig. 15A. Hence, if there is a big difference between the maximum and the minimum error and if there is a true second minimum error the current main block is deemed to contain a periodic structure. In case it is confirmed that this block contains a periodic structure then the motion vector belonging to either the upper or the left neighbour main block, that one giving the smaller error, replaces the motion vector found for the current main block. In case of equality preference is given to the left main block vector.

For the first, i.e. upper left, main block in the field the periodic structure correction is switched off. For the remaining part of the first main block row only the left candidate is available. For the remaining part of the first main block column only the upper candidate is available. This leads to the following algorithm:

The 2nd Min Row Error must be a true minimum with higher errors at each horizontal side. Therefore it is considered that the borders of a row represent a very small error, i.e. error = zero.

The 2nd Min Row Error is only used to decide whether the current block is periodic or not and hence, if there are several equal 2nd Min Row Errors, it does not matter which one is finally used.

Fig. 16 gives an example showing a current vector range block CUB, a corresponding left block LEB and upper block UPB. MVH is the horizontal motion vector component, MVV is the vertical motion vector component, MVUP is the location of a chosen motion vector from the upper block and MVLE is the location of a chosen motion vector from the left block.

MBE is the location of the maximum block error, MRE is the location of the maximum row error, 2MRE is the second minimum row error, MINRE is the minimum row error which is in this example equal to the minimum block error MINBE, CHR is the chosen row which corresponds to the row with the minimum block error. The arrows starting with locations MVLE and MVUP, respectively, represent such replacement vectors from the adjacent blocks which are chosen according to the corresponding minimum error CERR.

In even output fields the horizontal motion vector component MVH has a slightly bias towards zero to avoid some ambiguity due to noise. The vertical motion vector component MVV suffers a stronger bias towards zero owing to the vertical interpolation of one of the two fields used to calculate the motion vectors. It is important to avoid problems on vertical transitions as depicted in Fig. 17 with a vertical main block border VMBB. Between an odd source field OSF and an even source field ESF an odd interpolated field OIF is formed.

For example, B is a black pixel of value 16, G is a grey pixel of value 128 and W is a white pixel of value 240. Then MBMV(0,0) gives an error of 16 pixels * value 128 = 2048 from the lowest row in the main block above VMBB. MBMV(0,+2) gives an error of zero.

Zero forcing vertical:

If

Error(h, 0) ≦ [Error(h, v) + 2048]

then MBMV = h, 0

else MBMV = h, v

Zero forcing horizontal (use result from above):

If

Error(0, v) ≦ [Error(h, v) + 64 + ZFE * N_(z)/4]

then MBMV = 0, v

else MBMV = h, v

ZFE = Zero Forcing Even fields. ZFE = 0, 1, 2 or 4, can be programmable via I2C-Bus. Default: ZFE = 1. N_(z) is the noise level mentioned above.

Zero forcing in output field 3:

Field OF3 is very critical, leading to line flicker, if existing zero motion is not detected. But since the horizontal and the vertical zero forcing to calculate the MBMVs is separated the MBZVs can directly be deduced from the MBMVs.

If

(MBMV = 0) then MBZV = 0

else MBZV = 1

The algorithm so far described provides one motion vector per main block. MVH has 16 possible values and can be coded with 4 bits. MVV has 3 possible values and can be coded with 2 bits. A complete active picture (see Fig. 14) has 36*22 main blocks which leads to 792 main blocks per current picture (z) which are stored in the vector memory RAM in ME with a word length of 6 bits each. This vector memory stores also the values of the main block zero vectors MBZV for the calculation of output field OF3.

The next step is to use this information to calculate motion vectors specific to smaller areas known as subblocks which have a size of e.g. 2*2 pixels on the subsampled proscan grid or 4*4 pixels on the proscan grid. One example of subblock motion vector calculation is described in EP-A-94113494 in more detail.

The advantage therefrom is that the subblock motion vectors are calculated directly before they are required and therefore need not to be stored, resulting in a low-capacity RAM in ME. The main block vectors are calculated in a first field, the subblock motion vectors are calculated in the next field.

As motion vector for a particular subblock one of the three spatially nearest neighbour main block motion vectors will be chosen by applying them to the specific subblock and calculating the three corresponding subblock errors SE.

In Fig. 18 the main block C is divided into four areas. Candidate motion vectors for subblocks in the upper left area are chosen from the motion vectors of the nearest main blocks A, B and C. Candidate motion vectors for subblocks in the upper right area are chosen from the motion vectors of the nearest main blocks A, C and D. Candidate motion vectors for subblocks in the lower left area are chosen from the motion vectors of the nearest main blocks B, C and E. Candidate motion vectors for subblocks in the lower right area are chosen from the motion vectors of the nearest main blocks C, D and E.

The subblock errors are calculated by accumulation of the four pixel differences d_i within the area of subblock SB between field (z) and field (z-1), see Fig. 19:

SE(x,y) = |d₁| + |d₂| + |d₃| + |d₄|,

wherein x,y are the components of one of the three main block motion vectors.

The mainblock motion vector MBMV(x,y) leading to the minimum error SE(x,y) is preliminary chosen for the current block. Subsequently, a 3-tap horizontal median filtering operation is performed for each component SBMVH and SBMVV of the subblock motion vectors to erase single erroneous vectors. The resulting motion vector is then used to calculate the even output field pixels belonging to this subblock. x and y is also the horizontal and vertical subblock position , respectively.

Advantageously, the above calculated subblock error, SE, is directly representative for the confidence of the motion vectors. If SE has a small numerical value the confidence CL is high and vice versa. EP-A-94115733 of the applicant describes a confidence measurement in more detail.

CL can also be calculated in ME in the following way:

CL'' = median[SE_(x-1)/2, SE_(x)/2, SE_(x+1)/2]

CL = CL''/CLdiv   {truncation}

CL is e.g. limited to 8 bit. CLdiv is a preselected factor 1, 2, 4, 8, which is programmable e.g. via I2C bus.

Default: CLdiv = 2.

The horizontal median filter is applied in order to ensure that, if the confidence is good on both sides of a block boundary, it remains good across a boundary, where the unfiltered confidence is poor.

For a soft switch used in the chrominance processing a chroma confidence CC can also be calculated in ME: CC = CL/2.

Further, a temporal noise reduction on/off signal TNRS depending from the confidence CL and noise level N is created in ME as follows:

If

(CL*CLdiv < 2*N)   then TNRS = 1   else TNRS = 0

A further feature of the invention is noise reduction. There are two stages of noise reduction, namely the spatial noise reduction SNR and the temporal noise reduction TNR. The spatial noise reduction is preferably performed on luminance only but can be applied to chrominance also, whereas the temporal noise reduction is preferably performed on both, luminance and chrominance. Fig. 20 shows the common principle for noise reduction in the embodiments according to Fig. 1 to 7A. Additional processing time or line delays are not shown. The sped-up input signal ISPU is first processed in SNR means. However, the order of SNR and TNR could also be vice versa.

The spatial noise reduction SNR can be controlled by the noise level N and is bypassed when N is low, e.g. N≦4, in order to avoid possible resolution degradation. N comes from motion estimator ME. The spatial noise reduction may be enabled/disabled via I2C bus. Default = enabled.

The spatial noise reduction filter may comprise a horizontal and/or vertical and/or diagonal lowpass filter and/or may comprise a two-dimensional or a directional median filter. In case of directional filtering the filter direction is a function of the correlation of the central pixel with the corresponding neighbouring pixels. The filter operates in a window of e.g. 3*3 pixels and reduces impulse noise but preserves texture, details and edges. The filtering result corresponding to the direction of the minimum correlation is selected as final output pixel value.

Referring again to Fig. 20, the output of SNR passes a temporal noise filter TNR the output of which is input to motion compensated interpolator MCI. Basically, the temporal filter combines data from different fields including the current field. Advantageously, the temporal filter TF (luminance and chrominance) can be a recursive filter with a non-linear coefficient. It is accompanied by the predictor PRED, the interpolator MCI and the field delay FM. TF is controlled by TNRS and N. PRED receives the output signal of field memory FM and subblock motion vectors SBMV from motion estimator ME. For luminance, the predicted subblocks from the spatially and temporally noise reduced source field (z-1) are considered in TF together with the spatially noise reduced source field subblocks of field (z) coming from SNR. The TNR output signal comprises spatially and temporally noise reduced subblocks for source field (z).

If due to noise a wrong motion vector is calculated the recursive temporal filter could propagate this error in terms of partly wrong video information at the wrong spatial position, because ME makes use of motion compensated signals which have been constructed using previously calculated motion vectors. Hence, the calculation of temporally following motion vectors could be wrong. Advantageously, due to the coefficient applied in the temporal noise reduction to the predicted video information, this disturbing effect is decaying from field to field. Careful tests have shown that in fact such errors do not propagate and that therefore the inventive cheaper MCU circuitry can be used.

The coefficient mentioned above allows to add a share of only up to about 3/4 of the predicted video signal to the video signal of the current field. If, e.g. in case of noisy source signals the confidence CL for the motion vectors is lower this share will be reduced and thereby the possible error propagation is decreased even more.

MCI receives from ME the signals SBMV, N, CL, CC and MBZV, the latter of which is the main block zero motion vector from the last motion estimation for source field (z-1).

Luminance processing for the temporal noise reduction on/off signal TNRS:

if

(TNRSL enabled [via I2C-Bus]) and (TNRS = 1)

then temporal noise reduction luma is switched on

else it is bypassed in order to avoid blurring due to estimation failure

Default = TNRSL enabled

Chrominance processing for signal TNRS:

if

(TNRSC = forced [via I2C-Bus]) or if (TNRS = 1)

then temporal noise reduction chroma is switched on

else it is bypassed.

Default = TNRSC forced

Predictor PRED makes use of the subblock motion vectors SBMV to project the previous field in motion direction and, due to interlace, performs an interpolation, e.g. a vertical average, to generate the missing pixels for a non-interlaced picture.

The temporal luminance filter receives the spatially noise reduced current field (z) and the spatially and temporally noise reduced last source field (z-1) coming out of the predictor. The combination, preferably non-linear combination, of both gives the spatially and temporally noise reduced luminance input fields (z) for MCI.

The temporal chrominance filter receives the current field (z) and the temporally noise reduced last source field (z-1) coming out of the predictor. The combination, preferably non-linear combination, of both gives the temporally noise reduced chrominance input fields for interpolator MCI.

Motion compensated interpolator MCI produces four output fields for every two input fields. Odd first output field OF1 is the noise reduced source field F1 as shown in Fig. 26. F1' is the first field from the next input source frame.

In figures 21 to 26, X denotes input pixels, □ denotes output pixels, CH denotes chroma source pixels, CI denotes horizontally interpolated chroma pixels and CO denotes chroma output pixels.

Note that in Fig. 23B to 26B due to the double sided block matching CI pixels are required for positions of which lie not within the multiple-of-four grid used in the chrominance predictor PRED.

The processing for the third output field OF3 is very special, see Fig. 22. The grid of the noise reduced source field F3 does not match with the grid of the output field OF3, owing to interlace and upconversion. Therefore, depending on the value of the main block zero vector MBZV, one of two different filter types is applied to construct the pixel values for OF3, e.g. a vertical-temporal median filter and a vertical filter with coefficients (1/2, 1/2):

If

MBZV(z-1) = 0 then vertical-temporal median filter

else vertical filter.

The area with dashed boundaries depicts the pixels involved in the vertical-temporal median filter. Area shows the pixels involved in the vertical filter.

For each pixel in the even output fields OF2 and OF4 a motion compensated pixel and a fall back pixel is calculated. The final luminance output is then, depending on the confidence CL, a combination of the motion compensated pixel value and the fall back pixel value by means of a soft switch which is depicted in Fig. 27.

Advantageously, different processing modes can be used for calculating OF2 and OF4 pixels and these modes can be made selectable via I2C bus.

Mode 1 (default):

A vertical temporal median filter is used like that described with Fig. 22. The pixels of output field OF2 are constructed by taking the vertical temporal median (applied in the direction of motion) of the noise reduced field F1 and field F3. In other words, the output pixel is the median value of the pixels f, g and h as shown in Fig. 23A for the vertical vector component MVV and in Fig. 23B for the horizontal vector component MVH. Fig. 23B shows (in the horizontal-temporal domain) the output line n (OLn) of field F2 between lines n and n+2 (Lnn+2) of field F3 and line n-1 (Ln-1) of field F1.

The pixels of output field OF4 are constructed by taking the vertical temporal median (applied in the direction of motion) of the noise reduced field F3 and field F5. In other words, an output pixel is the median value of the pixels i, j and k as shown in Fig. 24A for the vertical vector component MVV and in Fig. 24B for the horizontal vector component MVH. Fig. 24B shows (in the horizontal-temporal domain) the output line n (OLn) of field F4 between line n (Ln) of field F5 and line n (Ln) of field F3.

The fall back pixels for OF2 and OF4 are calculated like the motion compensated pixels above but zero motion is assumed, i.e. SBMV = (0,0). Fallback pixels can also be calculated according to EP-A-94115634 of the applicant.

Mode 2:

A linear average filter is used like that described with Fig. 22. Output field OF2 is the linear average of the noise reduced field F1 and field F3. In other words, an output pixel is 1/2 of pixel f and 1/2 of pixel gh as shown in Fig. 25A for the vertical vector component MVV and in Fig. 25B for the horizontal vector component MVH. Fig. 25B shows (in the horizontal-temporal domain) the output line n (OLn) of field OF2 between lines n+1 (Ln+1) of field F3 and line n-1 (Ln-1) of field F1.

Output field OF4 is the linear average of the noise reduced field F3 and field F5. In other words, an output pixel is 1/2 of pixel i and 1/2 of pixel jk as shown in Fig. 26A for the vertical vector component MVV and in Fig. 26B for the horizontal vector component MVH. Fig. 26B shows (in the horizontal-temporal domain) the output line n (OLn) of field OF4 between lines n (Ln) of field F5 and line n (Ln) of field F3.

As for mode 1, in mode 2 the fall back pixels for OF 2 and OF4 are calculated like the motion compensated pixels as described above, but zero motion is assumed, i.e. SBMV = (0,0).

The circuitry of Fig. 27 is arranged at the output of MCI. Odd output field pixels OOFP are fed to the first input of a switch SWF. The motion compensated pixels MCP are fed to a first input of a soft switch SSW. Fall back pixels are fed to a second input of SSW and the confidence value CL to a third input of SSW. The output signal of SSW passes through limiting means LIM3 (limiting to 0...255) to the second input of switch SWF. In odd output fields the first input of SWF is connected using control signal OF to output OUP. In even output fields the second input of SWF is connected to output OUP.

SSW can be constructed in the following way:

The signal from the first input is subtracted from the signal of the second input in subtracting means SUB. The output is fed to the first input of minimum calculation means MIN. The confidence signal CL from the third input is passed to the second input of MIN and via two's complement means 2CO to the second input of maximum calculation means MAX. The first input of MAX receives the output signal of MIN. The output signal of MAX is added to the input signal of the first SSW input in combining means ADD, the result of which is fed to the output of SSW.

Chrominance processing

Basically, the chrominance (U and V) processing can be performed according to the luminance processing. Insofar the block diagrams are used in a similar way for processing the chrominance where not stated oppositely. Preferably, there are one or more of the following differences:

1) The horizontal component MVH of the motion vector is rounded to the nearest multiple of 4 before being applied to the chrominance to match to the chrominance sampling grid. The following table shows how the components of the motion vectors coming out of ME are applied in PRED and MCI for luminance and chrominance:

PRED luma

MCI luma

MCI luma

PRED chroma

MCI chroma

MCI chroma

MVV

Fb

Ff

MVV

Fb

Ff

+2

-1

+1

+2

-1

+1

0

0

0

0

0

0

-2

+1

-1

-2

+1

-1

MVH

Fb

Ff

MVH

Fb

Ff

+14

-7

+7

+12

-6

+6

+12

-6

+6

+12

-6

+6

+10

-5

+5

+8

-4

+4

+8

-4

+4

+8

-4

+4

+6

-3

+3

+4

-2

+2

+4

-2

+2

+4

-2

+2

+2

-1

+1

0

0

0

0

0

0

0

0

0

-2

+1

+1

0

0

0

-4

+2

-2

-4

+2

-2

-6

+3

-3

-4

+2

-2

-8

+4

-4

-8

+4

-4

-10

+5

-5

-8

+4

-4

-12

+6

-6

-12

+6

-6

-14

+7

-7

-12

+6

-6

-16

+8

-8

-16

+6

-6

Fb means field backward, Ff means field forward.

2) To allow point 1), additional horizontal linear average filters are needed within the interpolator MCI. Hence, a 4:2:2 format is produced from the 4:1:1 format before the motion vectors are applied to the chrominance pixels.

3) No spatial noise reduction SF is carried out.

4) The interpolation for the even output fields is done like in MODE 2 for luminance (linear average, no programmability).

5) Chrominance confidence CC (used for SSW chroma only), is a half of that of the luminance confidence CL: CC = CL/2 .

6) 4:2:2 → 4:4:4 conversion using linear average (1/2, 1/2).

Fig. 28 shows a global timing diagram for the input and output signal of SPUFM, for the input and output signal of FM and for the output signals of ME and MCI. F1NR ... F5NR means noise reduced fields.

There is a special calculation for F3NR''. As described above, in case of a main block zero motion vector MBZV value 0 a vertical temporal median filter is applied which has access to the not yet noise reduced field F5. Since SBMV and CL are unknown, MBZV(z-1) is applied. This causes a small temporal error. However, tests have shown that this error is negligible.

The circuitry shown in this invention can be integrated on one or several chips and can be controlled by an I2C bus. Part of the circuitry can be replaced by appropriate software and processor means.

The invention can be used in 4:3 or 16:9 standard definition or high definition units like TV receivers, satellite receivers, VCRs and video disc players.

In case of digital video signals these may be accompanied by motion information. In such case the invention can make use of this already existing motion information, whereby some processing steps concerning the generation of motion information can be omitted or adapted accordingly.