KR100527810B1 - Estimating Method of motion using SAD - Google Patents

Abstract

The present invention relates to a motion estimation method that greatly improves the estimation speed while maintaining a certain level of motion estimation performance by using the principle of a conventional successive elimination algorithm (SEA) among motion estimation algorithms widely used in a video encoder. .

Description

Estimating method of motion using SAD}

The present invention relates to a motion estimation method through SAD estimation. More specifically, the motion estimation performance is fixed by using the principle of the conventional successive elimination algorithm (SEA) among motion estimation algorithms widely used in a video encoder. The present invention relates to a motion estimation method for improving the estimation speed while maintaining a level.

Hereinafter, the concept of a motion estimation method will be described with reference to FIG. 1.

For an NxM block of a given reference picture and a current picture, the goal of motion estimation is to determine the NxM block in the reference picture that is most similar to the block of the current picture. The current picture is defined as an image located at time t, and the reference picture is defined as an image at time t-n located in the past and / or at time t + k located in the future. In this case, motion estimation for reference images located in past and future times is referred to as forward motion estimation and backward motion estimation, respectively.

The position of a block is usually expressed in coordinates (x, y) of the left-top corner of the block. When the position of the best match block of the reference image is (x + u, y + v), the vector from (x, y) to (x + u, y + v) is (x, y) This is defined as the motion vector of the block located at. Normally, a motion vector is expressed in a relative coordinate system, and assuming that (x, y) = (0,0), the motion vector is simply expressed as (u, v).

Although there are various matching criteria, the most popular criterion is SAD (Sum of Absolute Difference), which is defined as follows.

Here, u and v are defined by the motion search region, and f (x + k, y + l, t) denotes a pixel of a block in the current image and f (x-u + k, y-v + l, tn ) Denotes the brightness value of the pixel in the reference image. In the end, u and v at the smallest SAD are called the optimal motion vectors, and the best match occurs.

The simplest way to find the best motion vector is to find all the SADs for all possible u and v and to determine the u and v with the smallest SAD. This is called the full search algorithm (FSA). Global search has the advantage of finding the optimal motion vector, but it has the disadvantage of requiring a large amount of computation by calculating SAD for all possible cases.

Due to these shortcomings, various speeding algorithms exist in the motion estimation method. By reducing the search point, methods for improving the speed include hierarchical motion estimation and three-step motion estimation. These methods cannot guarantee that the optimal motion vector will always be found, which has the advantage of reducing the computational complexity, although the estimation performance is lower than that of the global search method. Unlike these algorithms, there is a continuous elimination algorithm (SEA) that improves speed while searching for the optimal motion vector like global search.

SEA uses the optimal motion vector satisfying the condition shown in the following equation when the coordinate of the current block is (x, y).

here, ,

SAD _min refers to the least SAD of the SADs for the search points to date. R (u, v) and C represent the sum of the brightness values of the blocks in the reference picture and the sum of the brightness values of the blocks in the current picture, respectively. SEA uses a search point that does not satisfy the condition of Equation 2 cannot be an optimal motion vector. That is, if the difference between the sum of the reference block brightness values (R (u, v)) and the current block brightness values (C) is greater than SAD _min , (u, v) cannot be the absolute optimal motion vector. . Therefore, since the SAD needs not to be calculated for the search point (u, v), the calculation amount is reduced by excluding (u, v) from the search point. On the other hand, the calculation of R (u, v), C is additionally required, but this can be ignored because it is a very small amount of calculation compared to the amount of motion estimation.

Hereinafter, a flowchart of a continuous elimination algorithm according to the prior art will be described in detail with reference to FIG. 2.

First, n is an integer indicating the number of search points, and it is determined whether the condition of Equation 2 is satisfied for the search points u and v. LUT (n) is a look-up table that stores the location of the search point. The LUT (n) can be freely designed according to the search strategy. Initialize to n = 0 (step S10), set the search point (u, v) to LUT (n) (step S12), and then check whether Equation 2 is satisfied for the corresponding search point (u, v). Determine (step S14).

If the condition is satisfied, the search point (u, v) may be the optimal motion vector. Therefore, SAD (u, v) is found for the search point (u, v). Otherwise, the search point (u, v) may not be the optimal motion vector. It is determined whether this is the last number, if not the last update to n ← n + 1 and go back to the beginning (step S26).

On the other hand, when the condition of Equation 2 is satisfied, SAD (u, v) is compared with SAD _min (step S18), and when SAD (u, v) is smaller than SAD _min , the current search point (u, v). ) Is more optimal than the search point with SAD _min , so update the best search point (u *, v *) to the current search point (u, v) and update SAD _min to SAD (u, v) do. If SAD (u, v) is greater than or equal to SAD _min , the current search point (u, v) cannot be the optimal motion vector, so it is updated to n ← n + 1 and returned to the beginning.

Finally, it is checked whether all search points are considered (step S24), and if all search points are considered (n = last), the whole process is terminated and the optimal motion vector is determined as (u *, v *).

Comparing the SEA as described above with the global search method, the global search method is shown in FIG. It is the same as omitting the process of judging. Therefore, SEA The amount of computation is reduced by calculating only the SAD for the search points satisfying. That is, since the SAD calculation is omitted for the search points that do not satisfy the condition, the larger the number of search points that do not satisfy the condition, the better the throughput reduction performance of the SEA is. However, since the condition of Equation 2 is not a strict condition, the amount of calculation reduction performance is not greatly improved.

Accordingly, an object of the present invention is to provide a method for reducing the amount of calculation by reducing the number of search points satisfying the conditions in the SEA.

As a means for solving the above-described problem, the present invention relates to a method of changing and applying the condition of Equation 2 described above, wherein the value of the left side | R (u, v) -C | of Equation 2 is SAD (u Using the estimates of SAD (u, v) instead of | R (u, v) -C | to reduce the amount of computation by reducing the number of search points that satisfy the condition I was able to do that.

As a means for solving the above-described problems, an aspect of the present invention provides a method for estimating motion through SAD estimation, comprising: (a) a minimum SAD value (SAD _min ) of an SAD estimate at a current search point and a SAD obtained so far; (B) when the SAD estimate is greater than SAD _min , and determines that the current search point cannot be the optimal motion vector, and thus skips SAD calculation, and when the SAD estimate is less than or equal to SAD _min , Calculating an actual SAD (i.e., SAD (u, v)), (c) comparing SAD (u, v) and SAD _min , and (d) SAD (u, v) is greater than SAD _min or is now determined by the search point is not the optimal motion vector, and if the SAD (u, v) is smaller than the SAD _min the SAD _min as SAD (u, v), the best motion vector (u *, v *) if: It provides a motion estimation method using SAD estimation to update to (u, v).

On the other hand, the SAD estimate at the current search point is the difference between the sum of pixel values R (u, v) of the block in the current image and the sum of pixel values C of the reference image (| (R (u, v))-C The value proportional to | can be set to the value obtained by subtracting the value proportional to the SAD standard deviation.

Further, in step (a), the SAD estimate and the | (R (u, v))-C | Compare whether the larger value is less than or equal to the minimum SAD value (SAD _min ).

Preferably, the proportional value is Can be.

The above objects, features and advantages will become more apparent from the following detailed description taken in conjunction with the accompanying drawings. Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings.

Hereinafter, a flowchart of a motion estimation method according to a preferred embodiment of the present invention will be described with reference to FIG. 3.

First, before performing the algorithm, in the initialization of step S101, λ is initialized to an appropriate value (λ ₀ ) in consideration of a target calculation amount and a motion estimation error (step S101), and the search point (u, v) is LUT (n). (Step S103). LUT (n) is a lookup table storing the location of the search point, and LUT (n) can be freely designed according to a search strategy.

(Calculate SAD Estimation EST_SAD (u, v))

In step S105 after initialization, the SAD estimate EST_SAD (u, v) is calculated for the search point (u, v) corresponding to n = 0. Here, the process of obtaining EST_SAD (u, v) will be described in detail.

here,

Of equation (3) Wow Are estimates of the mean and standard deviation for SAD (u, v), respectively. This was obtained by the following mathematical background. Assuming that the difference between the brightness values between the block pixels in the current image and the pixels in the reference image has a Gaussian distribution and there is no correlation between the brightness differences, the average of SAD (u, v) , Standard Deviation Is given by Where E [] is the expected value, α is a positive real number, and the standard deviation is proportional to the mean. the theoretical result of α to be. Based on these mathematical results, the estimate of the mean And the estimate of the standard deviation Decide on

In Equation 3, λ is a coefficient having a non-negative real value and is used for biased estimation. At λ = 0, EST_SAD (u, v) = μ ' _SAD (u, v) and as λ increases, EST_SAD (u, v) decreases with the slope of α, whereby EST_SAD (u, v) is the actual SAD value. This is to ensure that a value greater than (ie, SAD (u, v)) does not occur as much as possible.

For example, when (u, v) is an optimal motion vector, if EST_SAD (u, v) is estimated to be greater than SAD (u, v), (u, v) may be excluded from the search point. To reduce this case, EST_SAD (u, v) should be set as small as possible.

However, too small estimates generate a large number of search points, so they should be larger than the existing SEA criteria (ie, | R (x, y) -C |) to reduce the amount of computation. In this respect, a large value of λ reduces the likelihood of excluding the optimal motion vector from the search point (which leads to an increase in computation), and a small value of λ increases the possibility of excluding an optimal motion vector from the search point (this means Should be chosen). That is, it controls the trade-off relationship between the amount of calculation and the motion estimation error according to the value of λ.

(Comparison using SAD estimate EST_SAD (u, v))

Next, it is determined whether Equation 4 is satisfied (step S107).

Where max {A, B} is a function that takes the larger of A and B. Using the function max {,} means that the left side of Equation 4 is always | R (u, v) -C | To be greater than or equal to The left side of Equation 4 is always | R (u, v) -C | If it is greater than or equal to, it is more likely to be larger than SAD _min , so it is more likely to omit the actual SAD (ie, SAD (u, v)) from the search point (u, v). As a result, as the number of search points that do not calculate the actual SAD increases, the amount of calculation can be reduced. The comparison criterion proposed by the present invention is always | R (u, v) -C | The reason for setting to be greater than or equal to is because Equation 5 is always satisfied.

Equation 5 is a search that does not calculate SAD by setting an estimate to satisfy Equation 6 below since | R (u, v) -C |, which is a comparison criterion proposed by the conventional SEA, is always less than or equal to the actual SAD. The number of points can be increased and the amount of calculation can be reduced. ＿

That is, to obtain EST_SAD (u, v) that always satisfies Equation 6, information about all pixels of the block in the current image and the block in the reference image is required. This means that there is no improvement in terms of computation. Therefore, in order to obtain an improved result in the calculation amount, EST_SAD (u, v) should be determined using only a part of all pixels or only limited information such as the average and the variance of the pixels.

Here, using only limited information is equivalent to using information that can be obtained with a small amount of computation. As a result, when only limited information is used, a problem occurs in that Equation 6 cannot always be satisfied. Therefore, EST_SAD (u, v) is calculated as in Equation 3 as a probabilistic method that can satisfy Equation 6 while using only limited information.

Next, it is determined whether the condition of Equation 4 is satisfied, and if the condition is satisfied, since (u, v) may be an optimal motion vector, SAD (u, v) is calculated for the search point (u, v) ( Otherwise, since it cannot be an optimal motion vector, it is updated to n ← n + 1 (step S117) and the process returns to step S103.

In more detail, when SAD (u, v) is compared with SAD _min (step S111), when SAD (u, v) is smaller than SAD _min , the current search point (u, v) has SAD _min . Since the search point is more optimal than the search point, the optimal search point (u *, v *) is updated to the current search point (u, v), and SAD _min is updated to SAD (u, v). If SAD (u, v) is greater than or equal to SAD _min , the current search point (u, v) cannot be the optimal motion vector, so it is updated to n ← n + 1 and returned to the beginning.

Finally, it is determined whether all search points are considered (step S115). If n = last, the process is terminated and the optimal motion vector is determined as (u *, v *).

As described above, the present invention is fixed by changing the comparison criteria used in the conventional SEA in order to improve an increase in the amount of calculation that occurs due to the small number of search points excluded from the SAD calculation, which is a problem in the conventional SEA. It has the effect of greatly reducing the amount of calculation while satisfying the level estimation error.

In order to improve the increase in the calculation amount caused by the small number of search points excluded from the calculation of the sum of absolute difference (SAD), which is a problem in the conventional SEA, the object is achieved by changing the comparison criteria used in the existing SEA. How to do it.

In addition, since the motion estimation error and the amount of calculation can be easily adjusted according to the purpose of adjusting the coefficients included in the comparison criteria proposed in the present invention, the present invention is more flexible than the conventional method. Such a method according to the present invention can be applied to a video encoder using a motion estimation method.

1 is a conceptual diagram illustrating a concept of a motion estimation method.

2 is a flowchart of a continuous elimination algorithm according to the prior art.

3 is a flowchart of a motion estimation method according to a preferred embodiment of the present invention.

Claims (4)

In the motion estimation method through SAD estimation, (a) The difference between the sum of pixel values R (u, v) of the block in the current image and the sum of pixel values C in the reference image C (| (R (u, v))-C |) at the current search point. The SAD estimate is set to a value obtained by subtracting a value proportional to the SAD standard deviation, and the set SAD estimate and the | (R (u, v))-C | Comparing whether a greater value is less than or equal to a minimum SAD value SAD _min of SADs obtained so far; (b) If the SAD estimate is greater than SAD _min , it is determined that the current search point cannot be the best motion vector, and SAD calculation is omitted. If the SAD estimate is less than or equal to SAD _min , the actual SAD (u , v); (c) comparing the SAD (u, v) and the SAD _min ; And (d) the SAD (u, v) is the case wherein the SAD _min is greater than or equal to the currently determined by the search point is not the optimal motion vector, the SAD (u, v) is less than SAD _min the SAD _min SAD and (u, v), updating the optimal motion vector (u *, v *) to (u, v). delete delete The method of claim 1, The proportional value is Motion estimation method through SAD estimation, characterized in that.