1. Field of the Invention
This invention relates to motion adaptive image processing.
2. Description of the Prior Art
Video image capture represents a spatial and temporal sampling process. An image is captured as a set of pixels arranged in rows or lines. Successive images are captured at spaced instants in time.
A complication is the common use of interlaced video capture and processing. In an interlaced video signal, each image is handled as two sets or fields of alternate lines of pixels. For example, odd numbered lines might be included in one field, whereas even numbered lines could be included in the next field. An advantage of interlaced techniques is that they give an apparent doubling of the image rate, so reducing flicker effects, for no substantial increase in video signal bandwidth.
All of these aspects of sampling can give rise to alias effects if an attempt is made to capture or process video material having spatial or temporal frequencies which are too high for the respective sampling rate. But a particular alias problem will be described here in the area of interlace to progressive scan video conversion.
If it is desired to convert between interlaced video and progressive scan (non-interlaced) video, then for non-moving images it is merely necessary to interleave two successive fields to recreate a non-interlaced frame having all lines of pixels present. However, if there is any significant inter-field motion, this approach may not work. In such circumstances it can be more appropriate to derive the lines of pixels which are missing in one field from other pixels in that same field. In other words an intra-field interpolation process is used.
In practice, a video source may comprise image sequences in which some regions represent moving images whilst some regions do not. For example, when a newscaster speaks to a fixed camera, the newscaster's mouth, face, and head may move considerably, whilst their torso, the desk and the wall behind them do not.
Therefore the different conversion strategies noted above may be appropriate within different regions of the same image. It is therefore important to determine which strategy to use for a given pixel.
Interpolation will generally give a worse result than interleaving for non-moving portions, whereas interleaving and will generally give a worse result than interpolation for moving portions. So, the choice of the more appropriate technique is very important.
In particular, interleaving between image fields generally gives worse results than interpolation of a single image field when there are moving portions of the image because it results in a striped effect in the resultant image as alternate lines taken from different image fields will have visibly different content.
Moreover, under some circumstances it is also possible for successive fields to be visibly different despite a lack of substantial motion. In particular, when there is a sudden change in lighting (for instance, when a light is switched on in a room, or there is lightning or a strobe effect). In such circumstances it is also desirable to avoid interleaving between image fields within affected portions of an image.
It is an object of the present invention to seek to alleviate or mitigate this problem.
In an aspect of the present invention, a method of image processing comprising the steps of calculating a difference measure of a difference in pixel values between corresponding blocks of a current image field and an adjacent image field, calculating a variability measure of a variability of pixel values in respect of at least one of said corresponding blocks, determining whether the difference measure exceeds the variability measure, and if the difference measure exceeds the variability measure, setting respective inter-image mixing values for pixels in the corresponding block of the current image field to indicate that the current image field should be mixed with the adjacent image field at that pixel position to at most a predetermined extent.
In another aspect of the present invention, an image processing apparatus comprises difference calculation means for calculating a difference measure of a difference in pixel values between corresponding blocks of a current image field and an adjacent image field, variability calculation means for calculating a variability measure of a variability of pixel values in respect of at least one of said corresponding blocks, comparison means for determining whether the difference measure exceeds the variability measure, and data modification means for setting an inter-image mixing value of a pixel, and in which if, in operation, the comparison means determines that the difference measure exceeds the variability measure, then the data modification means sets respective inter-image mixing values for pixels in the corresponding block of the current image field to indicate that the current image field should be mixed with the adjacent image field at that pixel position to at most a predetermined extent.
Advantageously, therefore, the above aspects enable the detection of motion-independent changes to successive image fields, and to set or override existing inter-image mixing parameter values for affected pixels, thereby preventing the use of interleaving during the conversion from interlaced to progressive scan for these pixels and so mitigating the above noted detrimental affects upon image quality.
Further respective aspects and features of the invention are defined in the appended claims.
The above and other objects, features and advantages of the invention will be apparent from the following detailed description of illustrative embodiments which is to be read in connection with the accompanying drawings in which:
a to 4c schematically illustrate gradient detection;
a to 6e schematically illustrate a spatial block matching operation;
a and 7b schematically illustrate an alias situation;
a to 8d schematically illustrate alias detection techniques;
a schematically illustrates a motion adaptive interpolator;
b schematically illustrates motion detection between successive video fields;
It will be appreciated that the source of interlaced material 20 need not be a broadcast receiver, but could be a video replay apparatus such as a DVD player, a network connection such as an internet connection and so on.
The interlaced output from the source of interlaced material 50 is supplied to an interlace to progress scan converter 70 to generate a progressive scan signal. This can be processed by the vision mixer 80 along with the progressive scan material from the source 60 to generate a processed progressive scan output. Of course, the progressive scan output of the vision mixer 80 can be converted back to an interlaced format if required, e.g. for subsequent broadcast or recording. It will also be appreciated that the vision mixer 80 is just one example of video processing apparatus; instead, a digital video effects unit, for example, could be used at this position in
The converter of
The motion adaptive interpolator will be described in more detail below. First, the spatial interpolator will be briefly described.
The spatial interpolator comprises a 1:2 horizontal pixel scaler 130, a spatial block matcher 140, a minimum error selector 150, a diagonal interpolator 160, a dot noise reducer 170 and a Kell-factor corrector 180. The operation of each of these is summarised below.
The scaler 130 uses horizontal linear interpolation to generate one additional pixel value between each two pixels of the input interlaced field (i.e. a 1:2 scaling operation). So, the horizontal resolution (at least in terms of number of available pixel values) is doubled, but no difference is made at this stage to the vertical resolution.
The overall operation of the spatial block matcher 140 and the diagonal interpolator 160 is to detect the orientation of an image feature relevant to a pixel position where a new pixel is to be interpolated, and then to apply an interpolation along that image feature direction. So, if a current pixel position to be interpolated lies within a diagonal image feature (a line, an edge etc.) at, say, 45° to the horizontal, interpolation of that new pixel would take place along that 45° direction. This can tend to give a better output result than restricting the interpolation to horizontal or vertical interpolation. A key part of this process, clearly, is therefore to detect the direction of an image feature at each pixel position.
Referring now to
So, referring to
Turning now to
However, in
As noted above, spatial block matching is carried out at sub-pixel accuracy; in this case half-pixel accuracy.
A range of block sizes is used, with corresponding search ranges (maximum displacements relative to the pixel position under test). Taking into account the 1:2 scaling operation, example block sizes and search ranges are given in the following table:
Note that the displacement is indicated as a displacement from the centre. The two blocks are displaced by equal amounts, though in opposite directions. Symmetrical displacements are used because otherwise the block matching could detect lines or edges which are not relevant to the pixel under test.
A sum of absolute differences (SAD) is calculated for each block match. This is defined as:
where x, y represent the current pixel co-ordinate (y being a frame line number), d is the displacement being tested, and n is the “radius” of the block (the block width is n′=2n+1).
In general terms, the SAD values for three colour components (red, green and blue) are combined, and a minimum normalised SAD value determines a gradient for interpolation. Various checks are made to avoid poor interpolation, as described below.
Measures are taken to avoid problems caused by alias situations.
Referring to
However, in
This conflict of block match results is a product of aliasing between the closely spaced diagonal image features in the image portions shown in
The basis of the rule is that the block match process is restricted so that only areas considered to be “line segments” are detected. That is to say, each block in a block match should contain a line segment.
A digitised line segment is considered to have two properties. Firstly, it is monotonic along the central scan line row of the block in question, and secondly there is a vertical transition between scan lines in the block in question. The way in which these properties may be tested will be described with reference to
In
So, turning back to
The tests are performed separately in respect of each of the colour components (e.g. R, G and B). All three tests must be passed separately. Alternatively, for example to save hardware, fewer than three tests could be performed. For example, only the luminance, or only one colour component, might be tested. Of course, a YCbCr or YPbPr representation could be tested instead.
The diagonal interpolator 160 is a simple pixel averager: given a direction it picks the pixel in that direction on the line below and the pixel in that direction on the line above and averages them.
The dot noise reducer 170 involves a process which is applied to the output of the diagonal interpolator 160. A test is applied to detect whether an interpolated pixel lies within the maximum and minimum values of four neighbouring vertical and horizontal pixels, i.e. the pixels immediately above, below, left and right of the interpolated pixel. Note that the pixels above and below the interpolated pixel will be real pixels, whereas those to the left and right will be interpolated themselves.
If the interpolated pixel does not lie within this range, then;
Let v be the original value of the pixel under consideration, and let v′ be v, clipped to lie within the range of the four locally neighbouring pixels.
Let the new pixel value be kDNR v′+(1−kDNR)v, where kDNR is a programmable constant.
The operation of the Kell-factor corrector 180 will now be described.
In the present discussion, references to the Kell-factor are simply to help explain the operation of this part of an exemplary system. What the filter is actually exploiting is simply the knowledge that the source image did not use the full bandwidth available to it, whether that is because of scanning artifacts or because of a low pass filtering process.
The Kell-factor is a quantity which represents a property of progressive scan and interlaced images. In order to represent the information being scanned, it is generally considered that only 70% (the Kell-factor) of the possible vertical bandwidth is (or should be) represented. Hence when performing an interlace to progressive scan conversion, it is potentially hazardous to attempt to produce a full vertical bandwidth image. Instead, a compensation to account for a Kell-factor of less than unity may be used.
One method to compensate for the Kell-factor would be to use a 70% bandwidth filter on the frame output of any interlace to progressive scan algorithm. However, one of the fields in the frame is ‘real’ data—i.e. it was sampled correctly, so the content arising from that field must by definition be perfect. Thus a method to filter just the interpolated lines is used.
a schematically illustrates the operation of the motion adaptive interpolator 110. The interpolator 110 comprises and inter-field block matcher 600, an optional high frequency checker 610 and a mixer 620.
The inter-field block matcher 600 uses data from the current input field and the three field stores 120 to carry out inter-field motion comparisons. This involves comparing blocks of pixels the current field (FN in
In particular, weighted sums of absolute differences (SWADs) are generated as follows.
Four block matches are performed to produce two SWADs, SWADAREA and SWADLOCAL. These are:
a 5hx4v weighted block match on fields FN and FN-2.
a 5hx3v weighted block match on fields FN-1 and FN-3.
a 1hx1v weighted block match on fields FN-1 and FN-3.
a 1hx2v weighted block match on fields FN and FN-2.
Weighted block matches sum weighted absolute differences between coincident pixels, SWAD.
where FN-1(dx,dy) is the value at the frame-relative position dx, dy to the current pixel.
Typical values for the weights are:
Summing the first two SWADs gives an area-based block match, SWADAREA
Summing the latter two SWADs gives a localised block match, SWADLOCAL
All three colour components contribute to the SWADs in the same manner. The system need only maintain a SAD of the three components for each pixel, which is then weighted and combined with the values from the other pixels in the block. This means that this aspect of the process requires only 5 line stores of about 10 bpp (bits per pixel).
Optionally, the high frequency checker 610 is arranged to detect high frequencies in the input fields. The algorithm is based on the following principle. If interleaving the two source fields produces a lot of high frequency energy, then it is appropriate to try to make sure that the inputs are reasonably static. Only static video can produce reliable high frequencies; highly aliased motion can produce high frequencies, but this is not a desirable situation for inter-field interpolation. If motion is present, then high frequencies may be produced where the fields are incorrectly interleaved.
Referring to
Let HFCthresh1 and HFCthresh2 be two programmable constants, with the former greater than the latter.
Set a flag: exceededHighEnergy=false
Over each component (or a subset of them) (RGB/YPbPr)—where YPbPr indicates the colour space in a high definition system, in a similar way to YCbCr in a standard definition system:
Set energy=0
For the pixels having a horizontal position x=−2, −1, 0, 1, 2 (relative to the current pixel), let the interleaved (FN-1) field value be v0, and the current field value of the line above and below be v−1 and v1, then:
If (diff>HFCthresh1), set energy=energy+(HFCthresh1−HFCthresh2)*weighting[x]
else if (diff>HFCthresh2), set energy=energy+(diff−HFCthresh2)*weighting[x]
If energy>HFCallowance, set flag exceededHighEnergy=true
This ends the processing carried out over each component.
Subsequently, if exceededHighEnergy=true, increase SWADAREA by a programmable constant value, HFCpenalty.
The increase in SWADAREA will tend to act against the use of the motion adaptive pixel at that output position.
The mixer 620 operates according to the criteria SWADAREA and SWADLOCAL and also various thresholds thresh1,2, etc.
If SWADLOCAL>thresh1, use only spatially interpolated field, FN′.
Else if SWADAREA>thresh2, use only spatially interpolated field, FN′, only
Else if SWADAREA<thresh3, use only field FN-1
Else mix field FN-1 and FN′:
The resulting pixel value=αFN-1+(1−α) FN′. In other words, α represents pixel motion and determines contributions from the intra and inter-field interpolators.
Whilst only FN-1 and FN′ are mixed in the above equation, it will be appreciated that additional image fields or portions thereof may be mixed with FN-1 and FN′ at a given pixel position, for example the unfiltered lines of FN for alternate lines of the image, or earlier image field FN-3if there is substantially no motion at all.
As noted previously, it will be understood that areas of image fields where there is rapid motion should not be interleaved, as this will result in a striping or striation of the resultant image as alternate lines taken from different image fields will have visibly different content.
However, under some circumstances it is also possible for successive image fields to be dissimilar despite a lack of substantial motion. In particular, when there is a sudden change in lighting (for instance, when a light is switched on in a room, or there is lightning or a strobe effect), this may not be detected by the SWAD analysis described above.
Consequently in an embodiment of the present invention, a means to detect such circumstances is provided that overrides the value of a calculated above, setting it to indicate motion and thus prompting a spatial interpolation of the relevant area of the image rather than a use of image interleaving.
In an embodiment of the present invention two corresponding blocks, respectively comprising all or part of successive image fields, are determined to have a large field difference (irrespective of motion) if they satisfy the inequality
|DCN−DCN-1|>(SDN+SDN-1)/2,
where DCN,N-1 are the mean values of pixel data of the corresponding blocks in the current and preceding image fields, and SDN,N-1 are the standard deviations of the pixel data from those blocks.
Thus, when the difference between the mean pixel values of corresponding blocks from image fields N and N-1 exceeds the average standard deviation of pixel values from these blocks, then the blocks are categorised as having a large field difference.
The relevant pixel values typically relate to luminance and/or chrominance. For example, the pixel value may be a respective sum or average of red, green and blue channels, indicating luminance. Alternatively, any or all separate colour channels may be used directly or as difference signals to indicate chrominance. In the latter case, an inequality as disclosed above may be used for each selected colour channel, and a large field difference is determined if any one of the inequalities is satisfied.
When a block in the current image field (field N in the above expression) is categorised as having a large field difference, the a values of pixels in that block are set to zero. In addition, optionally the α values of pixels in all blocks immediately neighbouring that block are also set to zero.
When α=0, the resulting pixel value αFN-1+(1−α)FN′ reduces to FN′, i.e. solely the spatially interpolated version of the image at that pixel.
Thus by forcing α=0 in this manner, only spatial interpolation will be apparent in those parts of the converted image corresponding to blocks categorised as having a large field difference and so modified to have α values of zero.
In an embodiment of the present invention, blocks 32 pixels horizontal and 16 pixels vertical are used in the above test. It will be appreciated however that blocks of any suitable size can be used, although if the blocks are too small then the mean and standard deviation values used to determine a large field difference may become unreliable. Blocks near one or more edges of the image may be expanded to include remaining pixels up to the edge.
Modifying α values over a block in the manner described above can however result in visible discontinuities between modified blocks and neighbouring unaffected areas that may instead, for example, be utilising a degree of interleaving to generate their contribution to the output image.
In the above table, pixel 1 is innermost within the modified block, and pixel 4 is on the edge of the block, as illustrated.
It will be appreciated however that the transition region may be of any suitable size, may not be the same size horizontally as it is vertically, and may be wholly within the modified block, wholly outside but adjacent to the modified block, or may span the edge of the modified block.
Where a pixel is near a corner of a block where two transition regions converge, its a value is determined according to which edge it is closest to.
Where a modified block is the immediate neighbour of another modified block, such a transition region is not necessary where their edges join.
In an alternative embodiment, the categorisation of two blocks of successive image fields as having a large filed difference is determined by satisfying the inequality
|DCN−DCN-1|>Max(SDN,SDN-1),
where all the terms are as defined previously, and the function Max returns the largest of it input values.
Optionally, the squares of |DCN−DCN-1| and (SDN+SDN-1) or Max (SDN,SDN-1) may be used in the above calculations.
More generally, therefore, the determination of significant dissimilarity between corresponding blocks in successive image fields is based upon a comparison of a measure of the difference between these blocks with a measure of the variability of at least one of those blocks.
In an alternative embodiment, a tuning parameter TuningParam is also provided within the inequality, giving, for example,
|DCN−DCN-1|*TuningParam>Max (SDN,SDN-1) or
|DCN−DCN-1|*TuningParam>(SDN+SDN-1)/2
as applicable.
This tuning parameter enables an adjustment or weighting of when the inequality is satisfied, thereby providing greater or lesser sensitivity to changes between image fields. It will be appreciated that the tuning parameter could be additive instead of or in addition to being multiplicative.
It will be appreciated that potentially more than one tuning, or weighting, parameter may be used. For example,
if |DCN−DCN-1|*small>Max(SDN,SDN-1) then limit α to 25% of value;
if |DCN−DCN-1|*medium>Max(SDN,SDN-1) then limit α to 50% of value;
if |DCN−DCN-1|*large>Max(SDN,SDN-1) then limit α to 75% of value;
else α remains unchanged
where small, medium and large are different tuning parameter values, each applicable in turn. The modification to a could be any appropriate modification, and optionally a function of some or all of DCN, DCN-1, SDN, or SDN-1 values.
In an embodiment of the present invention, the determination of block dissimilarity and any consequent modification of α value data is made by the inter-field block matcher 600. However it will be appreciated by a person skilled in the art that this function may in principle be implemented by any element of the converter 70 that is located in a suitable part of the image processing chain to perform such a function.
A corresponding method of image processing for converting interlaced images to progressive scan images is now described with reference to
Calculating (s1) a difference measure of the difference in pixel values between corresponding blocks of a current image field and a preceding image field;
Calculating (s2) a variability measure of the variability of pixel values from at least one of said corresponding blocks;
Determining (s3) whether the difference measure exceeds the variability measure; and
if the difference measure exceeds the variability measure,
setting (s4) an inter-image mixing value for each pixel in the corresponding block of the current image field to indicate that the current image field should not be mixed with the preceding image field at that pixel position or should be mixed to at most a predetermined extent (e.g. by setting the α value to 0, thereby indicating motion and in turn prompting only the use of the spatially interpolated version of the pixel, or by setting the α value to a maximum proportion of its initial value, the extent of mixing is limited).
If the difference measure does not exceed the variability measure, then the values of α are unaffected.
It will be appreciated by a person skilled in the art that variations in the above method corresponding to the various operations of the apparatus disclosed herein are considered to be within the scope of the present invention, including:
It will further be appreciated by a person skilled in the art that in the above embodiments, the comparison between the current and previous image fields may alternatively be between the current and future image fields, where a delay of one or more image fields has been applied during processing, thereby enabling acquisition of the adjacent future image field. Thus the comparison may be performed between the notional current frame and an adjacent field being the preceding or succeeding image field.
It will also be appreciated that rather than using an α value of zero to denote no motion and hence that only spatial interpolation should be used, alternatively an α value close to zero may be used where the resultant degree of mixing is below a threshold level (the threshold level being an empirically determined level acceptable—either due to being visually inoffensive or visually undetectable—to a typical or average viewer of the resulting image).
Thus more generally, in the above embodiments optionally the inter image mixing value α is set to indicate that the current image field is mixed to only to a limited extent with the adjacent image field, or is not mixed at all.
It will be appreciated by a person skilled in the art that the reference to the mixing of fields FN′ and FN-1 is one of several potential mixing options available for interlacing images in an image sequence. Generalising FN′ to FSx and FN-1 to FM, a selection of interpolation modes may be defined as:
Mode 1: Zero field system delay (as in the description)—
Mode 2: One field system delay, backward mixing—
Mode 3: One field system delay, forward mixing—
Consequently, for example, the mixed pixel value would equal α FM+(1−α)FSx.
It will be appreciated that the invention can be implemented in programmable or semi-programmable hardware operating under the control of appropriate software. This could be a general purpose computer or arrangements such as an ASIC (application specific integrated circuit) or an FPGA (field programmable gate array). The software could be supplied on a data carrier or storage medium such as a disk or solid state memory, or via a transmission medium such as a network or internet connection as a signal, or via combinations of these.
Although illustrative embodiments of the invention have been described in detail herein with reference to the accompanying drawings, it is to be understood that the invention is not limited to those precise embodiments, and that various changes and modifications can be effected therein by one skilled in the art without departing from the scope and spirit of the invention as defined by the appended claims.
Number | Date | Country | Kind |
---|---|---|---|
0624405.7 | Dec 2006 | GB | national |