Interpolation method for a motion compensated image and device for the implementation of said method

Abstract

The invention proposes a motion compensated video interpolation that is not sensitive to the errors in the motion estimation vectors. The invention proposes a motion compensated video interpolation method in which the interpolation step varies according to the amplitude and/or a reliability index of the estimated motion vector for the considered pixel. The number of pixels to be taken into account for interpolation and the weighting factors associated with these pixels are made to vary according to the amplitude and/or reliability index of the estimated motion vector.

Description

LIST OF FIGURES

The invention will be better understood upon reading the following description, provided for information only and referring to the annexed drawings wherein:

FIG. 1 is a flowchart representing the steps of the invention method in a first embodiment based on the amplitude of the estimated motion vector,

FIG. 2 is a diagram showing different interpolation patterns to be applied according to the amplitude of the estimated motion vector,

FIG. 3 more specifically illustrates the interpolation in the case wherein the estimated motion vector is less than a threshold SA1,

FIG. 4 more specifically illustrates the interpolation in the case wherein the estimated motion vector is between the threshold SA1 and a threshold SA2,

FIG. 5 more specifically illustrates interpolation in the case wherein the estimated motion vector is greater than the threshold SA2,

FIG. 6 is a device capable of implementing the method of FIG. 1,

FIG. 7 is a flowchart representing the steps of the invention method in a second embodiment based on a reliability index of the estimated movement,

FIG. 8 is a diagram showing different interpolation patterns to be applied according to a reliability index of the estimated motion vector,

FIG. 9 is a device capable of implementing the method of FIG. 7,

Ia FIG. 10 is a flowchart representing the steps of the invention method in a third embodiment based on the amplitude and a reliability index of the estimated motion vector,

FIG. 11 is a diagram showing different interpolation patterns to be applied according to the amplitude and a reliability index of the estimated motion vector,

FIG. 12 is a device capable of implementing the process of FIG. 10.

DETAILED DESCRIPTION OF PREFERRED EMBODIMENTS

As previously indicated, the principle of the invention is to make the interpolation pattern vary, i.e. the number of pixels on which the interpolation and/or the associated weighting factors are based, according to the amplitude of the estimated motion vector for the considered pixel and or/a reliability index associated with the estimated motion vector.

Three embodiments are described hereinafter: a first embodiment based solely on the amplitude of the estimated motion vector, a second embodiment based solely on a reliability index of the estimated motion vector and a third embodiment combining both the above mentioned embodiments.

FIG. 1 is a flowchart representing the steps of the first embodiment of the invention method. It comprises:

• • a motion estimation step 100 for estimating a motion vector (Vx,Vy) for each pixel of the image to interpolate,
  • a step 110 for determining the amplitude A of the estimated motion vector (A=√{square root over (Vx²+Vy²)}),
  • an interpolation pattern selection step 120 for each pixel of the image to interpolate according to the amplitude of the estimated motion vector for this pixel,
  • an interpolation step 130 for generating an interpolated image by using, for each pixel of this image, the interpolation pattern selected for this pixel.

FIG. 2 more specifically illustrates the interpolation step 130 and shows different types of interpolation (bilinear interpolation, interpolation by averaging) and various patterns to be applied according to the amplitude of the estimated motion vector. It can be noted that the direction of the motion vector (positive or negative motion vector) is not involved in the choice of the interpolation. In this embodiment, we use bilinear interpolations or averaging interpolations based on a plurality of neighbouring pixels. For each one of these types of interpolation, the higher the amplitude of the estimated motion vector, the greater is the increase in the number of pixels of the interpolation pattern.

If the amplitude of the motion vector is less than or equal to a first threshold SA1, the motion is considered as being low. The threshold SA1 is for instance fixed at 2 pixels (per frame). The interpolation is then for example a bilinear interpolation based on 4 pixels, these 4 pixels being the neighbouring pixels of the pixel considered in the source images. Neighbouring pixels are understood to mean the source image pixels whose spatial coordinates are close to the coordinates of the considered pixel of the image to interpolate. In the case of a single frame interpolation, only the 4 neighbouring pixels of one of the source images are used. This case is illustrated by FIG. 3. The 4 neighbouring pixels are noted P₁, P₂, P₃ and P₄, the video values associated with these 4 pixels are designated by V₁, V₂, V₃ and V₄ and the associated weighting factors are noted α₁, α₂, α₃ and α₄. The value of the factor α_i is a function of the distance separating the associated pixel P_i and the pixel to interpolate. In the example of FIG. 3, the pixel P1 is the nearest to the pixel to interpolate and its factor, α₁, is therefore the highest. The pixel P4 is the farthest from the pixel to interpolate and its factor, α₄, is therefore the lowest. The value assigned to the pixel to interpolate is then equal to

$\left(\sum _{i=1}^4{\alpha }_i·V_i\right)/\sum _{i=1}^4{\alpha }_i.$

Quite obviously and as indicated previously, if the motion vector points to a specific pixel among the 4 pixels, the weighting factor assigned to the other 3 pixels is nil. In the case of a double frame interpolation, bilinear interpolation is done on the 4 neighbouring pixels of the previous source image and the 4 pixels of the current source image. If V₁, V₂, V₃ and V₄ designate the values of the 4 neighbouring pixels of the current source image and α₁, α₂, α₃ and a designate the associated weighting factors and if V′₁, V′₂, V′₃ and V′₄ designate the values of the 4 neighbouring pixels of the previous source image and α′₁, α′₂, α′₃ and α′₄ designate the associated weighting factors, the value assigned to the pixel to interpolate is then equal to β.

$\left(\left(\sum _{i=1}^4{\alpha }_i·V_i\right)/\left(\sum _{i=1}^4{\alpha }_i\right)\right)+\left(1-\beta \right)·\left(\left(\sum _{i=1}^4{\alpha }_i^{\prime }·V_i^{\prime }\right)/\left(\sum _{i=1}^4{\alpha }_i^{\prime }\right)\right).$

In this formula, β is a factor that is dependent on the temporal position of the image to interpolate with respect to the source images. If the image to interpolate is temporally positioned at mid-distance between the source images, β=½. If the image to interpolate is more of the current source image source than the previous source image, β is greater than ½.

If the amplitude of the estimated motion vector for a given pixel is nil, the motion vector then points to the pixel having the same spatial coordinates in one and the other of the previous and current source images. Bilinear interpolation amounts to recopying the value of the corresponding pixel, i.e. having the same spatial coordinates, from one of the previous or current source images. It may be noted that, if the previous and current source images are progressive images obtained by deinterleaving of an interleaved video sequence, it is possible that the values of the pixels which ought to be equal (since they are linked by a nil motion vector) are not so. In this case, we can still recopy one or the other of the values of the pixels pointed to by the motion vector or possibly work out the average.

If the amplitude of the motion vector is greater than the threshold SA1 and less than or equal to a second threshold SA2 greater than SA1, the motion is considered to be average. The threshold SA2 is for instance fixed at 6 pixels (per frame). Interpolation is then an averaging step based on the 4 neighbouring pixels P₁, P₂, P₃ and P₄ of the pixel considered in the source images. This case is illustrated by FIG. 4. The weighting factors of the 4 neighbouring pixels are equal (α₁=α₂=α₃=α₄=α). In the case of a single frame interpolation, the value assigned to the pixel to interpolate is then equal to

$\left(\sum _{i=1}^4{\alpha }_i·V_i\right)/\sum _{i=1}^4{\alpha }_i=\left(\sum _{i=1}^4{\alpha }_i·V_i\right)/4.$

In the case of a double frame interpolation where (α′₁=α′₂=α′₃=α′₄=α′), the value assigned to the pixel to interpolate is then equal to

$\beta ·\left(\left(\sum _{i=1}^4{\alpha }_i·V_i\right)/\left(\sum _{i=1}^4{\alpha }_i\right)\right)+\left(1-\beta \right)\left(\left(\sum _{i=1}^4{\alpha }_i^{\prime }·V_i^{\prime }\right)/\left(\sum _{i=1}^4{\alpha }_i^{\prime }\right)\right)=\left(\beta ·\left(\sum _{i=1}^4V_i\right)+\left(1-\beta \right)·\left(\sum _{i=1}^4V_i^{\prime }\right)\right)/4.$

Finally, if the amplitude of the motion vector is greater than the threshold SA2, the movement is considered to be high. Interpolation is then an averaging step based on 8 neighbouring pixels, for example those shown in FIG. 5. The weighting factors of the 8 neighbouring pixels are equal (=α). In the case of a single frame interpolation, the value assigned to the pixel to interpolate is then equal to

$\left(\sum _{i=1}^8V_i\right)/8.$

In the case of a double frame interpolation, the value assigned to the pixel to interpolate is then equal to

$\left(\beta ·\left(\sum _{i=1}^8V_i\right)+\left(1-\beta \right)·\left(\sum _{i=1}^8V_i^{\prime }\right)\right)/8.$

FIG. 6 represents a device capable of implementing the previously described process. This device comprises at least two image memories 200 each of which is capable of memorizing a source image, namely the current source image and the previous-source image, a motion estimator 210 for determining a motion vector (Vx,Vy) for each one of the pixels of the image to interpolate from the current source image at the input of the device and the previous source image stored in one of the image memories 200, a circuit 220 for determining the amplitude A of the estimated motion vectors coming from the estimator 210 and an interpolator 230 for generating the interpolated image from the current and previous source images stored in the image memories 200 and the amplitudes A determined by the circuit 220. The interpolator 230 compares, for each pixel to interpolate, the amplitude A of the estimated motion vector for this pixel at the thresholds SA1 and SA2 and the nil value and subsequently applies the appropriate interpolation pattern.

FIG. 7 is a flowchart showing the steps of the second embodiment of the invention. It comprises:

• • a motion estimation step 300 for estimating a motion vector (Vx,Vy) for each pixel of the image to interpolate,
  • a step 310 for determining a reliability index C for the estimated motion vectors, and
  • an interpolation pattern selection step 320 for each pixel of the image to interpolate according to the reliability index of the estimated motion vector, and
  • an interpolation step 330 for generating an interpolated image by using, for each pixel of this image, the interpolation pattern selected for this pixel.

In this embodiment, the lower this reliability index, the higher is the number of pixels used by the interpolation pattern. In fact, the interpolated image can be made more blurred when we have less reliability in the value of the estimated motion vector. This reliability index is determined for instance by analyzing the dispersion of the value of the motion vectors in an image area including the considered pixel. The more dispersed the values of the motion vectors in this area are, the lower is the reliability index associated to the considered pixel. The reliability index C is for example inversely proportional to the variance of the motion vectors in a window of 5×5 pixels including the considered pixel:

$C=1-\frac{\sum _{i=1,j=1}^{5,5}\sqrt{{\left(Vx_{3,3}-Vx_{i,j}\right)}^2-{\left(Vy_{3,3}-Vy_{i,j}\right)}^2}}{24\sqrt{{\left(2·Vx_{\max }\right)}^2+{\left(2·Vy_{\max }\right)}^2}}.$

The second term of the equation is standardized in order to be contained between 0 and 1. The coordinate pixel (3,3) designates the current pixel. Vx_i,j and Vy_i,j designate respectively the horizontal and vertical components of the motion vector of the pixel (i,j). Vx_max and Vy_max designate the maximum horizontal and vertical components of the motion vectors. In the case of an 8-bit coding of each of these components, Vx_max=Vy_max=255.

FIG. 8 more specifically illustrates the interpolation step 330 and shows examples of interpolation to apply according to the value of the reliability index associated with the estimated motion vector.

If the value of the reliability index is less than or equal to a first threshold SC1, the reliability index is considered as being low. Interpolation is then for example an averaging interpolation based on 8 pixels, these 8 pixels being the neighbouring pixels of the pixel considered in the source images. In the case of a single frame interpolation, only the 8 neighbouring pixels of one of the source images are used. If we reuse the notations defined for the first embodiment compared with FIGS. 2 to 5, the value assigned to the pixel to interpolate is then equal to

$\left(\sum _{i=1}^8\alpha ·V_i\right)/\sum _{i=1}^8\alpha =\left(\sum _{i=1}^8V_i\right)/8.$

In the case of a double frame interpolation, interpolation is done on the 8 neighbouring pixels of the previous source image and the 8 pixels of the current source image. The value assigned to the pixel to interpolate is then equal to

$\left(\beta ·\left(\sum _{i=1}^8V_i\right)+\left(1-\beta \right)·\left(\sum _{i=1}^8V_i^{\prime }\right)\right)/8.$

If the value of the reliability index is greater than the threshold SC1 and less than or equal to a second threshold SC2 greater than SC1, the reliability index is considered as being medium. Interpolation is then for example an averaging interpolation based on 4 pixels, these 4 pixels being the neighbouring pixels of the pixel considered in the source images. In the case of a single frame interpolation, only the 4 neighbouring pixels of one of the source images are used. If we reuse the notations defined for the first embodiment in relation to FIGS. 2 to 5, the value assigned to the pixel to interpolate is then equal to

$\left(\sum _{i=1}^4\alpha ·V_i\right)/\sum _{i=1}^4\alpha =\left(\sum _{i=1}^4V_i\right)/4.$

In the case of a double frame interpolation, interpolation is done on the 4 neighbouring pixels of the previous source image and the 4 pixels of the current source image. The value assigned to the pixel to interpolate is then equal to

$\left(\beta ·\left(\sum _{i=1}^4V_i\right)+\left(1-\beta \right)·\left(\sum _{i=1}^4V_i^{\prime }\right)\right)/4.$

Finally, if the value of the reliability index is greater than the threshold SC2, the reliability index is considered as being strong. Interpolation is then for example a bilinear interpolation based on 4 pixels, these 4 pixels being the neighbouring pixels of the pixel considered in the source images. In the case of a single frame interpolation, only the 4 neighbouring pixels of one of the source images are used. If we take up the notations defined for the first embodiment in relation to FIGS. 2 to 5, the

value assigned to the pixel to interpolate is then equal to

$\left(\sum _{i=1}^4{\alpha }_i·V_i\right)/\sum _{i=1}^4{\alpha }_i.$

In the case of a double frame interpolation, bilinear interpolation is done on the 4 neighbouring pixels of the previous source image and the 4 pixels of the current source image. The value assigned to the pixel to interpolate is then equal to β

$\left(\left(\sum _{i=1}^4{\alpha }_i·V_i\right)/\left(\sum _{i=1}^4{\alpha }_i\right)\right)+\left(1-\beta \right)·\left(\left(\sum _{i=1}^4{\alpha }_i^{\prime }·V_i^{\prime }\right)/\left(\sum _{i=1}^4{\alpha }_i^{\prime }\right)\right).$

FIG. 9 represents a device capable of implementing the previously described method. This device comprises at least two image memories 400 each of which is capable of memorizing a source image, namely the current source image and the previous source image, a motion estimator 410 for determining a motion vector for each one of the pixels of the image to interpolate from the current source image at the input of the device and from the previous source image in one of the image memories 400, a circuit 420 for determining a reliability index C for each of the estimated motion vectors coming from the estimator 410 and an interpolator 430 for generating the interpolated image from the current and previous source images stored in the image memories 400 and the reliability indices C determined by the circuit 420. The interpolator 430, compares, for each pixel to interpolate, the reliability index C of the estimated motion vector for this pixel at the thresholds SC1 and SC2 and subsequently applies the appropriate interpolation pattern.

FIG. 10 is a flowchart showing the steps of an embodiment of the invention method in which the interpolation pattern is selected according to the amplitude A and the reliability index C of the estimated motion vector. It comprises:

• • a motion estimation step 500 for estimating a motion vector (Vx,Vy) for each pixel of the image to interpolate,
  • a step 510 for determining, for each one of the estimated motion vectors, its amplitude A (A=√{square root over (V/x²+Vy²)}) and a reliability index C (which is for instance the variance of the motion vectors in a window of pixels including the considered pixel as described for the second embodiment), and
  • an interpolation pattern selection step 520 for each pixel of the image to interpolate according to the amplitude and the reliability index of the estimated motion vector, and
  • an interpolation step 530 for generating an interpolated image by using, for each pixel of this image, the interpolation pattern selected for this pixel.

For the interpolation step 530, we calculate for instance a value, called for example modified amplitude and noted A_mod, according to which an interpolation pattern is selected. A_mod is for example equal to A_mod=A×(1−C). The interpolation to be applied according to A_mod is for example illustrated by the FIG. 10. If A_mod is less than or equal to a first threshold SA′1, the interpolation applied is a bilinear interpolation based on the 4 neighbouring pixels of the considered pixel in one or both the source images. If A_mod is greater than the threshold SA′1 but less than or equal to a threshold SA′2 greater than SA1, interpolation is an averaging step based on the 4 neighbouring pixels of the considered pixel in one or both the source images. Finally, if A_mod is greater than the threshold SA2, interpolation is an averaging step based on a greater number of neighbouring pixels, for example 8 pixels.

FIG. 11 represents a device capable of implementing the process illustrated by FIG. 9. This device comprises at least two image memories 600 each of which is capable of memorizing a source image, namely the current source image and the previous source image, a motion estimator 610 for determining a motion vector (Vx,Vy) for each one of the pixels of the image to interpolate from the current source image at the input of the device and from the previous source image stored in one of the image memories 600, a circuit 620 for determining the amplitude A and the reliability index C of the estimated motion vectors coming from the estimator 610 and an interpolator 630 for generating the interpolated image from the current and previous source images stored in the image memories 600 and the amplitudes A and the reliability indices C determined by the circuit 620.

Naturally, the invention is not restricted to the previously described embodiments. Those skilled in the art will be able to foresee the use of interpolation types and interpolation patterns other than the ones presented here, by using a different number of pixels or different weighting factors.

Claims

1) Method for interpolating at least one pixel of a motion compensated image from at least two source images, comprising the steps of: for said pixel of the image to interpolate, estimating a motion vector from the source images,determining at least one information representative of the estimation of said vector,selecting an interpolation pattern for said pixel of the image to interpolate in accordance with said information representative of the estimation of said vector, said interpolation pattern identifying pixels in at least one of said source images and associating to each one of said identified pixels a weighting factor, andinterpolating said pixel of the image from the pixels and the factors of the selected interpolation pattern.

2) Method according to claim 1, wherein said information representative of the estimation of said vector is the amplitude of the estimated vector.

3) Method according to claim 2, wherein, if the amplitude of the estimated motion vector of the pixel of the image to interpolate is less than or equal to a first non zero threshold, the interpolation pattern comprises the n pixels of one of the source images nearest to the spatial position of said pixel to interpolate, n being greater than or equal to 2, and the weighting factor assigned to each of said source image pixels is dependent on their proximity to said pixel to interpolate.

4) Method according to claim 3, wherein, if the estimated motion vector of a pixel of the image to interpolate is greater than said first threshold, the interpolation pattern comprises the n pixels of one of the source images nearest to the spatial position of said pixel to interpolate, n being greater than or equal to 2, and the weighting factors assigned to said source image pixels are equal.

5) Method according to claim 3, wherein, if the estimated motion vector of a pixel of the image to interpolate is greater than said first threshold, the interpolation pattern comprises the n pixels nearest to the spatial position of said pixel to interpolate in each one of the source images, n being greater than or equal to 2, and the weighting factor assigned to each of said source image pixels is dependent on their proximity to said pixel to interpolate and the temporal position of said image to interpolate with respect to both the source images.

6) Method according to claim 3, wherein the number n of pixels increases with the amplitude of the estimated vector.

7) Method according to claim 2, wherein, if the estimated motion vector of said pixel of the image to interpolate is less than or equal to a first non zero threshold, the interpolation pattern comprises the n pixels of each one of the source images nearest to the spatial position of said pixel to interpolate, n being greater than or equal to 2, and the weighting factor assigned to each of the source image pixels is dependent on their proximity to said pixel to interpolate and the temporal position of said image to interpolate with respect to both the source images.

8) Method according to claim 7, wherein, if the estimated motion vector of a pixel of the image to interpolate is greater than said first threshold, the interpolation pattern comprises the n pixels of one of the source images nearest to the spatial position of said pixel to interpolate, n being greater than or equal to 2, and the weighting factors assigned to said source image pixels are equal.

9) Method according to claim 7, wherein, if the estimated motion vector of a pixel of the image to interpolate is greater than said first threshold, the interpolation pattern comprises the n pixels nearest to the spatial position of said pixel to interpolate in each one of the source images, n being greater than or equal to 2, and the weighting factor assigned to each of said source image pixels is dependent on their proximity to said pixel to interpolate and the temporal position of said image to interpolate with respect to both the source images.

10) Method according to claim 7, wherein the number n of pixels increases with the amplitude of the estimated vector.

11) Method according to claim 1, wherein said information representative of the estimation of said vector is a reliability index relating to the estimation of said vector.

12) Method according to claim 11, wherein, if the reliability index of the motion vector of said pixel of the image to interpolate is greater than a first reliability threshold, the interpolation pattern comprises the n pixels of one of the source images nearest to the spatial position of said pixel to interpolate, n being greater than or equal to 2, and the weighting factors assigned to each of said source image pixels is dependent on their proximity to said pixel to interpolate.

13) Method according to claim 12, wherein, if the reliability index of the motion vector of said pixel of the image to interpolate is less than or equal to said first reliability threshold, the interpolation pattern comprises the n pixels of one of the source images nearest to the spatial position of said pixel to interpolate, n being greater than or equal to 2, and the weighting factors assigned to said source image pixels are equal.

14) Method according to claim 12, wherein, if the reliability index of the motion vector of said pixel of the image to interpolate is less than or equal to said first reliability threshold, the interpolation pattern comprises the n pixels nearest to the spatial position of said pixel to interpolate in each one of the source images, n being greater than or equal to 2, and the weighting factor assigned to each of said source image pixels is dependent on their proximity to said pixel to interpolate and the temporal position of said image to interpolate with respect to both the source images.

15) Method according to claim 12, wherein the number n of pixels decreases with the value of the reliability index of the estimated vector.

16) Method according to claim 11, wherein, if the reliability index of the motion vector of said pixel of the image to interpolate is greater than a first reliability threshold, the interpolation pattern comprises the n pixels of each one of the source images nearest to the spatial position of said pixel to interpolate, n being greater than or equal to 2, and the weighting factor assigned to each of said source image pixels is dependent on their proximity to said pixel to interpolate and the temporal position of said image to interpolate with respect to both the source images.

17) Method according to claim 16, wherein, if the reliability index of the motion vector of said pixel of the image to interpolate is less than or equal to said first reliability threshold, the interpolation pattern comprises the n pixels of one of the source images nearest to the spatial position of said pixel to interpolate, n being greater than or equal to 2, and the weighting factors assigned to said source image pixels are equal.

18) Method according to claim 16, wherein, if the reliability index of the motion vector of said pixel of the image to interpolate is less than or equal to said first reliability threshold, the interpolation pattern comprises the n pixels nearest to the spatial position of said pixel to interpolate in each one of the source images, n being greater than or equal to 2, and the weighting factor assigned to each of said source image pixels is dependent on their proximity to said pixel to interpolate and the temporal position of said image to interpolate with respect to both the source images.

19) Method according to claim 16, wherein the number n of pixels decreases with the value of the reliability index of the estimated vector.

20) Method according to claim 11, wherein the reliability index relating to the estimation of a motion vector for a pixel to interpolate is dependent on the variance of the motion vectors in a window of m×m pixels including said pixel.

21) Method according to claim 1, wherein said at least one information representative of the estimation of said vector is the amplitude of said estimated vector and a reliability index relating to the estimation of said vector.

22) Interpolation device of at least one pixel of a motion compensated image from at least two source images, comprising: a memory for storing said source images,a motion estimator for estimating, for said pixel of the image to interpolate, a motion vector from said source images,a determination circuit for determining at least one information representative of the estimation of said vector, andan interpolation circuit for selecting an interpolation pattern for said pixel of the image to interpolate in accordance with said information representative of the estimation of said vector, said interpolation pattern identifying pixels in at least one of said source images and associating to each one of said identified pixels a weighting factor and interpolating said pixel from the pixels and the factors of the selected interpolation pattern.