JP2003143541A - Method and device for editing image - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve the image quality of an intermediate frame in an application for generating the intermediate frame from a key frame with interpolation. SOLUTION: An image input part 12 inputs the data of the key frame, and a matching processor 16 calculates corresponding point information between the data. An intermediate image generating part 18 generates an intermediate frame from the key frame and the corresponding point information. When a user edits the intermediate frame with an image editor 24, a registering part 26 registers the frame in a key frame storing part 14 as a new key frame. The matching processor 16 regenerates corresponding point information by adding the new key frame.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an image editing technique, and more particularly to a technique for generating and editing an intermediate frame from a key frame.

[0002]

2. Description of the Related Art MP is the de facto world standard for video compression.
The target area of EG (Motion Picture Expert Group) has expanded from storage media such as CDs to transmission media such as networks and broadcasting. Digitization of broadcasting cannot be considered without a compression coding technique centered on MPEG. The barrier between broadcasting and communication has collapsed, and the diversification of service providers is essential, and it is difficult to predict how the digital culture will progress in the broadband era.

[0003]

Even in such chaos, there are certain things regarding the direction of moving image compression technology. That is, it is compatible with a higher compression rate and image quality. As is well known in MPEG, block distortion may prevent high compression. Moreover, as the distribution of moving images becomes more practical, it is necessary to give more consideration to real-time processing.

The present invention has been made in view of the above circumstances, and an object thereof is to provide an editing technique for realizing efficient compression of image data. Another object is to provide an editing technique that meets the contradictory demands of maintaining the image quality and improving the compression rate. Still another object is to improve the image quality of an intermediate frame in an application that generates an intermediate frame by interpolation from a key frame. Still another object is to realize editing work for improving image quality through high interaction with the user.

[0005]

The present invention relates to an image editing technique. This technology was previously filed by the applicant in Japanese Patent No. 292735.
The image matching technology proposed in No. 0 (hereinafter referred to as "premise technology") can be used.

One aspect of the present invention relates to an image editing method. In this method, when an intermediate frame generated by, for example, interpolation calculation based on a key frame is edited, the edited intermediate frame is registered as a new key frame.

When the intermediate frame is generated based on the key frame, the image quality of the intermediate frame is not always sufficient depending on the situation. In that case, an image editor (hereinafter, also simply referred to as a user) may correct the intermediate frame by himself. For example, when a subject moves greatly between two keyframes, the subject may remain like a ghost in the intermediate frame at the position where it existed in both keyframes. At that time, the user erases the ghost with an image editor or other image editing tool and draws a subject at a desired position. The subject can be taken out from one of the key frames, for example, by cutting and pasting.

However, if the coded data of the image is formed by only the keyframes and the corresponding point information between the keyframes, the intermediate frame whose corners have been corrected is not reflected in the coded data. Therefore, the present inventor considered a technique of upgrading the thus modified intermediate frame to a new key frame and adding it to the first key frame group.

According to this technique, first, key frames are taken out at a considerable distance, an intermediate frame is generated to check the image quality, and the key frame can be added only to a problematic part. Can be maintained with a small number of key frames. In addition, since a series of processes can be performed by a very simple editing work, its high usefulness can be easily understood.

Another aspect of the present invention is also an image editing method, a step of accepting a user's edit for an initial intermediate frame generated based on a key frame, and registering the edited intermediate frame as a new key frame. And inquiring of the user whether or not to do so.

Instead of the inquiring step, a step of forcibly registering the edited intermediate frame as a new key frame may be included.

The method may further include a step of acquiring corresponding point information between the key frames, which includes a case of acquiring from the outside and a case of calculating internally. The method may further include the step of generating the initial intermediate frame by interpolation calculation based on the corresponding point information. Further, the method may further include the step of regenerating the intermediate frame including the new registered key frame. Further, the method may further include the step of displaying the regenerated intermediate frame and the step of accepting again the editing of the displayed intermediate frame.

Among the above-mentioned aspects, the base technology can be used for the part for generating the corresponding point information between the key frames and the part for generating the intermediate image using the information.
However, the present invention does not require the prerequisite technology. Further, the present invention is not limited to a moving image, and may be a still image such as generation of a smooth intermediate image between a plurality of viewpoint images.

The respective configurations and steps described above may be arbitrarily exchanged, or the expressions may be partially or wholly exchanged between the method and the device.
Alternatively, additions or changes in expressions to a computer program, a recording medium, etc. are also effective as the present invention.

[0015]

BEST MODE FOR CARRYING OUT THE INVENTION First, a multiresolutional singularity filter technique used in the embodiments and an image matching process using the technique will be described in detail as a "base technique". These techniques are the ones that the applicant has already obtained in Japanese Patent No. 2927350, and are optimal for combination with the present invention. However, the image matching technique that can be adopted in the embodiment is not limited to this. From FIG. 18 onward, the image processing technology using the base technology will be specifically described.

[Background of Prerequisite Technology] First, the elemental technology of the prerequisite technology will be described in detail in [1], and the processing procedure will be specifically described in [2]. In addition, we will report the results of the experiment in [3].

[1] Details of Elemental Technology [1.1] Introduction A new multi-resolution filter called a singularity filter is introduced to accurately calculate matching between images. No prior knowledge of objects is required. The calculation of matching between images is calculated at each resolution while going through the hierarchy of resolutions. In that case, the hierarchy of resolution is traced in order from a coarse level to a fine level. The parameters required for the calculation are fully automatic set by a dynamic calculation similar to the human visual system. It is not necessary to manually specify the corresponding points between images.

The base technology can be applied to, for example, completely automatic morphing, object recognition, stereoscopic photogrammetry, volume rendering, and generation of smooth moving images from a small number of frames. When used for morphing, a given image can be transformed automatically. When used for volume rendering, it is possible to accurately reconstruct an intermediate image between cross sections. The same applies when the distance between the cross sections is large and the shape of the cross section changes significantly.

[1.2] Hierarchical Singularity Filter Hierarchy The multiresolutional singularity filter according to the precondition technique can reduce the resolution of an image while preserving the brightness and position of each singularity included in the image. Here, the width of the image is N and the height is M. For simplicity, N = M = 2 ⁿ
(N is a natural number). Also, the interval [0, N] ⊂R
Is described as I. Let p be the pixel of the image at (i, j)
_It is described as _{(i, j)} (i, jεI).

Here, a multi-resolution hierarchy is introduced. hierarchy
The digitized image group is generated by a multi-resolution filter. Many
The double resolution filter is a two-dimensional search for the original image.
Search for singularities and extract the detected singularities
Generate another image with a lower resolution than the original image.
Here, the size of each image at the m-th level is 2^m× 2 ^m
(0 ≦ m ≦ n). There are four types of singularity filters:
Construct a new hierarchical image recursively downward from n
It

[0021]

[Equation 1] However, here

[Equation 2] And Hereinafter, these four images will be referred to as sub-images. min _{x ≦ t ≦ x + 1} , max
_{When x ≦ t ≦ x + 1} is described as α and β, the sub-images can be described as follows.

[0022] P^{(M, 0)}= Α (x) α (y) p^{(M + 1,0)} P^{(M, 1)}= Α (x) β (y) p^{(M + 1,1)} P^{(M, 2)}= Β (x) α (y) p^{(M + 1, 2}) P^{(M, 3)}= Β (x) β (y) p^{(M + 1,3)} That is, they are like the tensor product of α and β
Conceivable. Each sub-image corresponds to a singular point.
As is clear from these equations, the singularity filter
For each block of 2 × 2 pixels
Detect singularity. At that time, two people of each block
The maximum pixel value or the minimum pixel in the direction, that is, in the vertical and horizontal directions
Search for points with values. As a pixel value, it is bright in the base technology.
Degree, but various numerical values related to images are used.
You can Maximum pixels in both directions
The value pixel is the maximum for both the maximum point and the two directions.
Pixels with a small pixel value are the minimum point and one of the two directions.
Is the maximum pixel value, and the other is the minimum pixel value.
Pixels having a value are detected as saddle points.

The singularity filter represents the image of the block (here, 4 pixels) by the image of the singularity (here, 1 pixel) detected inside each block.
Reduce the image resolution. From the theoretical point of view of the singularity, α (x) α (y) preserves the minimum point and β (x) β
(Y) preserves the maximum points, α (x) β (y) and β
(X) α (y) stores the saddle point.

First, a singular point filter process is separately performed on a start point (source) image and an end point (destination) image to be matched, and a series of image groups, that is, a start point hierarchical image and an end point hierarchical image are generated. Keep it. The start point hierarchical image and the end point hierarchical image are generated in four types corresponding to the types of singular points.

After that, matching of the starting point hierarchical image and the ending point hierarchical image is achieved in a series of resolution levels. First, the minimum point is matched using p ^{(m, 0)} . Then, based on the result, saddle points are matched using p ^{(m, 1)} and other saddle points are matched using p ^{( m, 2)} . And finally p ^{(m, 3)}
The maximum points are matched using.

1 (c) and 1 (d) are respectively shown in FIG.
The sub-image p ^{(5, 0)} of (a) and FIG. 1 (b ⁾ is shown. Similarly, in FIG. 1 (e) and FIG. 1 (f), p ^(5,1) ,
1 (g) and 1 (h) show p ^(5,2) , and FIGS. 1 (i) and 1 (j) show p ^(5,3) , respectively. As can be seen from these figures, the sub-image facilitates matching of characteristic portions of the image. First, p ^(5,0) makes the eyes clear. This is because the eyes are the minimum points of brightness in the face. The mouth becomes clear according to p ^(5,1) . This is because the mouth has low brightness in the lateral direction. According to p ^(5,2) , the vertical lines on both sides of the neck become clear. Finally, p ^{(5,3) defines} the brightest points on the ears and cheeks. This is because these are maximum points of brightness.

Since the feature of the image can be extracted by the singularity filter, for example, the feature of the image taken by the camera,
By comparing the characteristics of several objects recorded in advance, it is possible to identify the subject reflected in the camera.

[1.3] Calculation of mapping between images The pixel at the position (i, j) of the starting point image is defined as p ⁽ⁿ⁾ _{(i, j).}
Similarly, the pixel at the position (k, l) of the end point image is changed to q
^(N) _{Described in (k, l)} . Let i, j, k, lεI. The energy of mapping between images (described later) is defined. This energy is determined by the difference between the brightness of the pixel of the start point image and the brightness of the corresponding pixel of the end point image, and the smoothness of the mapping. First, p ^{(m, 0)} and q with the smallest energy
Mapping between ^{(m, 0)} f ^{(m, 0)} : p ^{(m, 0)} → q
^{(M, 0)} is calculated. Based on f ^{(m, 0 )} , the mapping f ^{(m , 1)} between p ^{(m, 1} ) and q ^{(m, 1)} having the minimum energy is calculated. This procedure is p ^{(m, 3)}
And q ^{(m, 3)} map f ^{( m, 3)} until the calculation is completed. Each map f ^{(m, i)} (i = 0, 1, 2,
...) is called a submapping. For the convenience of calculating f ^{(m, i)} , the order of i can be rearranged as follows. The reason why the rearrangement is necessary will be described later.

[0029]

[Equation 3] Here, σ (i) ε {0, 1, 2, 3}.

[1.3.1] When the matching between the bijective start point image and the end point image is represented by a map, the map should satisfy the bijective condition between both images. This is because there is no conceptual superiority or inferiority between the two images, and pixels of each other should be connected in a bijective and injective manner. However, unlike the usual case, the mapping to be constructed here is a bijection digital version. In the base technology, pixels are specified by grid points.

Start point sub-image (provided for the start point image
From sub-image to end-point sub-image (provided for end-point image
The mapping to the sub-image is f^{(M, s)}: I / 2^nm× I
/ 2 ^nm→ I / 2^nm× I / 2^nm(S = 0,
1, ...). Where f
^{(M, s)}(I, j) = (k, l) is p of the starting point image
^{(M, s)} _{(I, j)}Is the end image q^{(M, s)}
_{(K, l)}Means to be mapped to. For simplicity
And when f (i, j) = (k, l) holds, the pixel q
_{(K, l)}Q_{f (i , J)}Write.

The definition of bijection is important when the data is discrete like the pixels (lattice points) handled in the base technology. Here, it is defined as follows (i, i ', j, j', k,
l is an integer). First, each square region represented by R in the plane of the starting point image,

[Equation 4] (I = 0, ..., 2 ^m −1, j = 0, ..., 2 ^m −
1). Here, the direction of each side (edge) of R is determined as follows.

[0033]

[Equation 5] This square must be mapped by the map f to a quadrilateral in the end image plane. a quadrilateral represented by f ^{(m, s)} (R),

[Equation 6] Must satisfy the following bijection conditions.

1. The edges of the quadrilateral f ^{(m, s)} (R) do not intersect each other. 2. The directions of the edges of f ^{(m, s)} (R) are equal to those of R (clockwise in the case of FIG. 2). 3. Contraction map (retraction: retrac) as relaxation condition
tions).

This is because the mapping that completely satisfies the bijection condition is only the unit mapping unless some relaxation condition is set. Here ^{f (m, s)} the length of one edge of the (R) is 0, i.e. ^{f (m, s)} (R) may become a triangle. However, it must not be a figure whose area is 0, that is, one point or one line segment. Figure 2 (R)
2A and FIG. 2D satisfy the bijection condition when the original is a quadrangle, FIG. 2B, FIG. 2C, and FIG.
(E) is not satisfied.

In an actual implementation, the following conditions may be further imposed in order to easily guarantee that the mapping is surjective. That is, each pixel on the boundary of the start point image is mapped to the pixel occupying the same position in the end point image. That is, f (i, j) = (i,
j) (where i = 0, i = 2 ^m- 1, j = 0, j = 2 ^m
-1 on the four lines). Hereinafter, this condition is also referred to as “additional condition”.

[1.3.2] Energy of mapping [1.3.2.1] The energy of the cost mapping f with respect to the luminance of the pixel is defined. The goal is to find the map that minimizes the energy. The energy is mainly determined by the difference between the brightness of the pixel of the starting point image and the brightness of the pixel of the corresponding end point image. That is, the energy C ^{(m, s)} _{(i, j) at} the point (i, j) of the map f ^{(m, s} ₎
Is determined by the following equation.

[0038]

[Equation 7] Here, V (p ^{(m, s)} _{(i, j)} ) and V (q
^{(M, s)} _{f (i, j)} ) is the pixel p ^{(m, s)}
The luminance of _{(i, j)} and q ^{(m, s)} _{f (i, j)} . The total energy C ^{(m, s)} of f is one evaluation formula for evaluating matching.
^It can be defined by the sum of ^{(m, s) and} _{(i, j)} .

[0039]

[Equation 8] [1.3.2.2] Cost of Pixel Position for Smooth Mapping In order to obtain a smooth mapping, another energy Df for the mapping is introduced. This energy is independent of the brightness of the pixel, p ^{(m, s)} _{(i, j)} and q ^{(m , s)}
Determined by the position of _{f (i, j)} (i = 0, ..., 2 ^m
-1, j = 0, ..., 2 ^m -1). Energy D ^{(m, s )} _{(i, j)} of the map f ^{(m, s) at} the point (i _{, j)}
Is defined by the following equation.

[0040]

[Equation 9] However, the coefficient parameter η is a real number of 0 or more, and

[Equation 10]

[Equation 11] And here,

[Equation 12] And f (i ', for i'<0 and j '<0
j ') is set to 0. E ₀ is (i, j) and f (i,
It depends on the distance of j). E ₀ prevents pixels from being mapped to pixels that are too far away. However, E ₀ is replaced later with another energy function. E ₁ guarantees the smoothness of the map. E ₁ represents the distance between the displacement of p _{(i, j) and} the displacement of its neighbors. Based on the above consideration, energy D _f, which is another evaluation expression for evaluating matching, is determined by the following expression.

[0041]

[Equation 13] [1.3.2.3] Total energy of mapping Total energy of mapping, that is, a comprehensive evaluation formula related to the integration of a plurality of evaluation formulas is defined by λC ^{( m, s)} _f + D ^{(m, s)} _f . Here, the coefficient parameter λ is a real number of 0 or more. The purpose is to detect the state where the comprehensive evaluation formula has an extreme value, that is, to find the map that gives the minimum energy shown by the following formula.

[0042]

[Equation 14] Note that for λ = 0 and η = 0 the map is a unit map (ie all i = 0, ..., 2 ^m
F for −1 and j = 0, ..., 2 ^m −1
^{(M 2 , s)} (i, j) = (i, j)). As will be described later, in the base technology, since the case of λ = 0 and η = 0 is evaluated first, the mapping can be gradually transformed from the unit mapping. Temporarily change the position of λ in the comprehensive evaluation formula to C
^{If (m, s)} _f + λD ^{(m , s)} _f is defined,
When λ = 0 and η = 0, the comprehensive evaluation formula is C ^{(m, s)} _f
Then, the pixels that are originally unrelated to each other are associated with each other simply because the brightness is close, and the mapping becomes meaningless. Even if the map is transformed based on such a meaningless map, it makes no sense at all. Therefore, consideration is given to how to give the coefficient parameter so that the unit map is selected as the best map at the start of the evaluation.

The optical flow is the same as in this base technology.
Consider differences in pixel brightness and smoothness. However, optical flow cannot be used for image conversion. This is because only the local movement of the object is considered.
A global correspondence can be detected by using the singularity filter according to the base technology.

[1.3.3] Determining Map by Introducing Multi-Resolution Map f that gives the minimum energy and satisfies the bijection condition
_min is obtained using a multi-resolution hierarchy. Compute the mapping between the starting and ending sub-images at each resolution level. Starting from the top (coarse level) of the hierarchy of resolutions, the mapping for each resolution level is determined, taking into account the mappings of other levels. The number of mapping candidates at each level is limited by using a higher, or coarser level of mapping. More specifically, when determining the mapping at a certain level, the mapping obtained at one coarser level is imposed as a kind of constraint condition.

First,

[Equation 15] , P ^{(m-1, s)} _{(i ', j')} , q
^{(M-1, s)} _{(i ', j')} are respectively p ^{(m, s)}
_{(I, j)} and q ^{(m, s)} are called parents of _{(i, j)} . [X] is the maximum integer that does not exceed x. Also, p ^{(m, s )} _{(i, j)} , q ^{(m, s)}
Let _{(i, j)} be p ^{(m-1, s)} _{(i ', j ')} , q ^{(m-1, s), respectively.}
_It is called a child of _{(i ', j')} . Function parent
(I, j) is defined by the following equation.

[0046]

[Equation 16] The mapping f ^{(m, s)} between p ^{(m, s)} _{(i, j)} and q ^{(m, s)} _{(k, l} ⁾ is determined by performing the energy calculation and finding the minimum one. To be done. f ^{( m, s)}
The value of (i, j) = (k, l) is f ^{(m-1, s)} (m =
By using 1, 2, ..., N), it is determined as follows. First, the condition that q ^{(m, s)} _{(k, l)} must be inside the following quadrangle is used to narrow down maps with high reality among maps that satisfy the bijective condition.

[0047]

[Equation 17] However, here

[Equation 18] Is. The quadrilateral determined in this way is p ^{(m, s)}
We call it an inherited quadrilateral of _{(i, j)} . In the inherited quadrilateral, find the pixel that minimizes the energy.

FIG. 3 shows the above procedure. In the figure, the pixels A, B, C, and D of the start point image are A ′, B ′, C ′, and C ′ of the end point image at the m−1th level, respectively.
It is projected to D '. The pixel p ^{(m, s)} _{(i, j)} is a pixel q existing inside the inherited quadrangle A′B′C′D ′.
⁽ M ^{, s)} must be mapped to _{f (m) (i, j)} . From the above consideration, the mapping from the m-1th level mapping to the mth level mapping is performed.

The energy E ₀ defined above is replaced by the following equation in order to calculate the submapping f ^{(m, 0} ) at the m-th level.

[0050]

[Formula 19] Further, the following equation is used to calculate the submapping f ^{(m, s)} .

[0051]

[Equation 20] This gives a map that keeps the energy of all submaps low. According to Expression 20, the submappings corresponding to different singular points are associated within the same level so that the submappings have a high degree of similarity. Formula 19 is f
^{(M, s)} shows the distance between (i, j) and the position of the point where (i, j) is to be projected when it is considered as a part of the m-1th level pixel.

If there is no pixel satisfying the bijective condition inside the inherited quadrilateral A'B'C'D ', the following measures are taken. First, a pixel whose distance from the boundary line of A′B′C′D ′ is L (initially L = 1) is examined. Among them, if the one with the smallest energy satisfies the bijection condition, this is selected as the value of f ^{(m, s)} (i, j). L is increased until such a point is found or L reaches its upper limit L ^(m) max. L ^(m) max is fixed for each level m.
If such a point is not found at all, the third condition of bijection is temporarily ignored, and a mapping such that the area of the transformed quadrangle becomes zero is accepted, and f ^{(m, s)} (i , J). If none of the conditions is found, then the bijective first and second conditions are removed.

The approximation method using multi-resolution is indispensable for determining the global correspondence between images while avoiding that the mapping is affected by image details. It is impossible to find the correspondence between pixels that are far apart without using an approximation method based on multiresolution. In that case, the size of the image must be limited to an extremely small size, and only an image with a small change can be handled. Furthermore, since smoothness is required for the normal mapping, it is difficult to find the correspondence between such pixels. This is because the energy of mapping from a pixel with a distance to a pixel is high. According to the approximation method using multi-resolution, it is possible to find an appropriate correspondence relationship between such pixels. This is because those distances are small at the upper level (coarse level) of the hierarchy of resolution.

[1.4] Automatic determination of optimum parameter value One of the main drawbacks of the existing matching technology is the difficulty of parameter adjustment. In most cases, adjustment of parameters is done manually, and it is extremely difficult to select the optimum value. With the method according to the base technology, the optimum parameter value can be determined completely automatically.

The system according to the base technology comprises two parameters, λ and η. In short, λ is the weight of the difference in the luminance of the pixels, and η is the rigidity of the mapping. The initial values of these parameters are 0. First, η is fixed to 0 and λ is gradually increased from 0. When the value of λ is increased and the value of the comprehensive evaluation expression (Expression 14) is minimized, the value of C ^{(m, s)} _f for each submapping generally decreases. This basically means that the two images must match better. However, when λ exceeds the optimum value, the following phenomenon occurs.

1. Pixels that should not be corresponded to each other are erroneously corresponded to each other simply because they have similar brightness. 2. As a result, the correspondence between pixels becomes strange,
The map begins to collapse.

3. As a result, in Equation 14, D
^{(M, s)} _f tends to increase rapidly. 4. As a result, the value of Equation 14 tends to increase rapidly, so that _f ^{(m, s)} _f is suppressed so as to suppress the rapid increase.
^{(M, s)} changes, resulting in an increase in C ^{(m, s)} _f . Therefore, the threshold value at which C ^{(m, s)} _f turns from decreasing to increasing is detected while maintaining the state that Expression 14 takes the minimum value while increasing λ, and sets λ to the optimum value at η = 0. Next, η is gradually increased and the behavior of C ^{(m, s)} _f is inspected, and η is automatically determined by the method described later. That η
Λ is also determined corresponding to.

This method resembles the operation of the focus mechanism of the human visual system. In the human visual system, images of both eyes are matched while moving one eye. When the object is clearly recognizable, its eyes are fixed.

[1.4.1] Dynamic determination of λ λ is increased from 0 by a predetermined step size, and the submapping is evaluated each time the value of λ changes. As in Equation 14, the total energy is defined by λC ^{(m, s)} _f + D ^{(m, s)} _f . D ^{(m, s)} _{f in} Expression 9 represents the smoothness, and theoretically has the minimum value in the case of the unit mapping, and E ₀ and E ₁ increase as the mapping becomes distorted. Since E ₁ is an integer, the minimum step size of D ^{(m, s)} _f is 1. Therefore, the current change (decrease ₎ of λC ^{(m, s)} _{(i, j} )
If is not greater than 1, then the total energy cannot be reduced by changing the mapping. Because D ^{(m, s)} _f increases by 1 or more as the mapping changes, λ
This is because the total energy does not decrease unless C ^{(m, s)} _{(i, j)} decreases by 1 or more.

Under this condition, it is shown that C ^{(m, s)} _{(i, j)} decreases in the normal case as λ increases.
The histogram of C ^{(m, s)} _{(i, j} ) is described as h (l). h (l) is energy C ^{(m, s)} _{(i, j)}
Is the number of pixels with l ² . Since λl ² ≧ 1 holds, consider the case of l ² = 1 / λ, for example. When λ slightly changes from λ ₁ to λ ₂ ,

[Equation 21] A pixels shown by

[Equation 22] It changes to a more stable state with the energy of. Here, it is assumed that the energies of these pixels are all zero. In this expression, the value of C ^{(m, s)} _f is

[Equation 23] Show that only changes, and as a result,

[Equation 24] Is established. Since h (l)> 0, usually C
^{(M, s)} _f decreases. However, when λ tries to exceed the optimum value, the above-mentioned phenomenon, that is, the increase of C ^{(m, s)} _f occurs. The optimum value of λ is determined by detecting this phenomenon.

When H (h> 0) and k are constants,

[Equation 25] Assuming that

[Equation 26] Holds. At this time, if k ≠ −3,

[Equation 27] Becomes This is a general expression for C ^{(m, s)} _f (C is a constant).

When detecting the optimum value of λ, the number of pixels that violates the bijection condition may be checked for safety. Here, when determining the mapping of each pixel, it is assumed that the probability of breaking the bijection condition is p ₀ . in this case,

[Equation 28] Therefore, the number of pixels that violates the bijection condition increases at the rate of the following equation.

[0063]

[Equation 29] Therefore,

[Equation 30] Is a constant. If h (l) = Hl ^k is assumed,
For example,

[Equation 31] Becomes a constant. However, when λ exceeds the optimum value, the above value increases rapidly. By detecting this phenomenon, B ₀ λ
^{3/2 +} value of ^k / ^{2/2 m} examines whether exceeds the abnormal value _{B 0thres,} it is possible to determine the optimal value of lambda. Similarly, by checking whether or not the value of B ₁ λ ^{3/2 + k /} _2/2 ^m exceeds the abnormal value B _1thres , the increase rate B _{1 of} pixels that violates the third condition of bijection is confirmed. To do. The reason for introducing the factor 2 ^m will be described later. This system is not sensitive to these two thresholds. These thresholds can be used to detect undue distortions of the map that are missed by observing the energy C ^{(m, s)} _f .

In the experiment, when the submapping f ^{(m, s)} is calculated, if λ exceeds 0.1, the calculation of f ^{(m, s)} is stopped and the calculation of f ^{(m, s + 1)} is started. did. λ> 0.
When 1, the difference of only “3” in the luminance 255 level of the pixel affected the calculation of the submapping, and λ> 0.1.
This was because it was difficult to get the correct result when.

[1.4.2] Histogram h (l) C^{(M, s)} _fThe inspection depends on the histogram h (l)
Absent. H (l) during bijection and the inspection of its third condition
Can be affected by. Actually (λ, C^{(M, s)} _f)
When lotted, k is usually around 1. In the experiment, k = 1
Using B₀λ ^TwoAnd B₁λ^TwoWas inspected. If k is true
If the value of is less than 1, then B₀λ^TwoAnd B₁λ^TwoBecomes a constant
Without factor λ^{(1-k) / 2}Gradually increase according to
To do. If h (l) is a constant, for example, the factor is λ
^1/2Is. However, such a difference is a threshold B
_0thresCan be absorbed by setting
You can

Here, the center of the starting point image is (x
₀ , y ₀ ) and a circular object of radius r.

[Equation 32] On the other hand, the end point image has the center (x ₁ , y ₁ ),
It is assumed that the object has a radius r.

[Expression 33] Here, it is assumed that c (x) has a form of c (x) = x ^k . If the centers (x ₀ , y ₀ ) and (x ₁ , y ₁ ) are far enough,
The histogram h (l) has the form:

[Equation 34] When k = 1, the image shows an object with a sharp border embedded in the background. This object has a dark center and becomes brighter toward the edges. When k = -1, the image represents an object with ambiguous boundaries. This object is brightest in the center and darkens towards the periphery. The general object does not lose its generality, even though it is considered to be somewhere in between these two types of objects. Therefore, k can cover most cases with −1 ≦ k ≦ 1, and it is guaranteed that Equation 27 is generally a decreasing function.

It should be noted that r is affected by the resolution of the image, that is, r is proportional to 2m, as can be seen from equation 34. For this reason [1.4.1]
Introduced a factor of 2 m.

[1.4.3] The dynamic determination parameter η of η can be automatically determined by the same method. First, η = 0, and the final mapping f at the finest resolution is obtained.
Calculate ⁽ⁿ⁾ and energy C ⁽ⁿ⁾ _f . Subsequently, η is increased by a certain value Δη, and again the final map f ⁽ⁿ⁾ and the energy C ⁽ⁿ⁾ _{f at} the finest resolution are obtained.
Recalculate. This process is continued until the optimum value is obtained.
η indicates the rigidity of the map. Because it is the weight of the following formula.

[0069]

[Equation 35] When η is 0, D ⁽ⁿ⁾ _f is determined independently of the immediately preceding submapping, and the current submapping is elastically deformed and excessively distorted. On the other hand, when η is a very large value, D
^(N) _f is almost completely determined by the immediately preceding submapping. At this time, the submapping is very rigid, and the pixels are projected on the same place. As a result, the map becomes a unit map. When the value of η gradually increases from 0, C ⁽ⁿ⁾ _f gradually decreases as described later. However, if the value of η exceeds the optimum value,
As shown in, energy begins to increase. In the figure, the X axis is η and the Y axis is C _f .

With this method, the optimum value of η that minimizes C ⁽ⁿ⁾ _f can be obtained. However, as a result of various factors affecting the calculation as compared with the case of λ, C ⁽ⁿ⁾ _f changes with a small fluctuation. In the case of λ, the submapping is recalculated once each time the input changes by a small amount.
This is because in the case of, all submappings are recalculated. Therefore, it is not possible to immediately determine whether or not the obtained value of C ⁽ⁿ⁾ _f is the minimum. When a candidate for the minimum value is found, it is necessary to search for the true minimum value by setting a finer interval.

[1.5] The range of f ^{(m, s)} can be expanded to R × R in order to increase the degree of freedom when determining the correspondence between super-sampling pixels (R is a real number). set). In this case, the brightness of the pixel of the end point image is interpolated,

[Equation 36] An f ^{(m, s)} with a luminance at is provided. That is, super sampling is performed. In the experiment, f
^{(M, s)} is allowed to take integer and half-integer values,

[Equation 37] Is

[Equation 38] Given by.

[1.6] Normalization of Pixel Luminance of Each Image When the start point image and the end point image include very different objects, it is difficult to use the original pixel luminance as it is for the calculation of mapping. This is because the energy C ^{(m, s)} _f related to the brightness becomes too large because the difference in the brightness is large, and correct evaluation is difficult.

For example, consider a case where a human face and a cat face are matched. The cat's face is covered in hair, with very bright and very dark pixels. in this case,
To calculate the submapping between two faces, the subimage is first normalized. That is, the brightness of the darkest pixel is set to 0 and the brightness of the brightest pixel is set to 255, and the brightness of other pixels is obtained by linear interpolation.

[1.7] Implementation A recursive method is used in which the calculation proceeds linearly according to the scanning of the starting point image. First, the value of f ^{(m, s)} is determined for the uppermost leftmost pixel (i, j) = (0,0). Next, increase each i by 1
^{(M, s)} Determine the value of (i, j). When the value of i reaches the width of the image, the value of j is increased by 1 and i is returned to 0.
Thereafter, as the start point image is scanned, f ^{(m, s)} (i,
j) is decided. When the pixel correspondences are determined for all points, one mapping f ^{(m, s)} is determined.

Corresponding point q for a certain p _{(i, j)
If f (i, j)} is determined, then the corresponding point qf _{(i, j + 1)} of p _{(i, j + 1)} is determined. At this time, q
_{Since the position of f (i, j + 1 )} satisfies the bijection condition,
Limited by the position of _{qf (i, j)} . Therefore, in this system, the higher the priority is, the higher the corresponding point is determined. When (0,0) always has the highest priority, an extra bias is added to the final mapping required. In the base technology, in order to avoid this state, f
^{(M, s)} is determined by the following method.

First, when (s mod 4) is 0, (0,
Start from 0) and gradually increase i and j while deciding. When (s mod 4) is 1, the right end point of the top row is set as the starting point, i is decreased, and j is increased. When (s mod 4) is 2, the right end point of the bottom row is set as the starting point, and i and j are decreased and determined. (S
When mod 4) is 3, the leftmost point of the bottom line is set as the starting point, and i is increased and j is decreased. Since the concept of submapping, that is, the parameter s, does not exist in the nth level having the smallest resolution, s = 0 and s =
Two directions were calculated consecutively as being 2.

In an actual implementation, a penalty is given to a candidate that violates the bijection condition, so that the bijective condition f ^{(m, s)} (from the candidates (k, l) is satisfied as much as possible. i, j) (m = 0, ..., N) values were chosen. Candidate energies that violate the third condition D (k,
l) is multiplied by φ, while candidates that violate the first or second conditions are multiplied by ψ. This time φ = 2, ψ = 100000
Was used.

In order to check the above-mentioned bijection condition, the following test was conducted in determining (k, l) = f ^{(m , s)} (i, j) as an actual procedure. That is, f
^{(M, s) For} each lattice point (k, l) included in the inherited quadrangle of (i, j), it is confirmed whether the z component of the outer product of the following equation is 0 or more.

[0079]

[Formula 39] However, here

[Formula 40]

[Formula 41] (Here, the vector is a three-dimensional vector, and the z axis is defined in the orthogonal right-handed coordinate system). If W is negative, then D ^{(m, s)} _{(k, l )} has ψ
Give a penalty by multiplying by and try to avoid choosing as much as possible.

FIGS. 5A and 5B show the reason for inspecting this condition. FIG. 5A shows a candidate with no penalty, and FIG. 5B shows a candidate with a penalty. Map f ^{(m, s)} for adjacent pixel (i, j + 1 ⁾
When determining (i, j + 1), if the z component of W is negative, no pixel satisfies the bijection condition on the starting image plane. This is because q ^{(m, s)} _{(k, l)} crosses the boundary line between adjacent quadrilaterals.

[1.7.1] In the submapping order implementation, σ (0) = 0, σ (1) = 1, σ (2) = 2, σ when the resolution level is even.
(3) = 3, σ (4) = 0 is used, and when odd, σ
(0) = 3, σ (1) = 2, σ (2) = 1, σ (3) =
0 and σ (4) = 3 were used. This appropriately shuffled the submap. In addition, originally there are four types of submappings,
s is any one of 0-3. However, actually s = 4
Was performed. The reason will be described later.

[1.8] Interpolation calculation After the mapping between the start and end images is determined, the corresponding
The brightness of the matching pixel is interpolated. In the experiment, tri-linear
Interpolation was used. Square p in the starting image plane
_{(I, j)}p_{(I + 1, j)}p_{(I, j + 1)}p
_{(I + 1, j + 1)}Is a quadrangle q on the end image plane
_{f (i, j)}q_{f (i + 1, j)}q_{f (i, j + 1)}q
_{f (i +1, j + 1)}Suppose it is projected to. Easy
Therefore, the distance between the images is set to 1. Distance from the starting image plane
Pixels r (x,
y, t) (0≤x≤N-1, 0≤y≤M-1) is
Required as required. First, the position of pixel r (x, y, t)
(However, x, y, tεR) is calculated by the following equation.

[0083]

[Equation 42] Subsequently, the brightness of the pixel at r (x, y, t) is determined using the following equation.

[0084]

[Equation 43] Here, dx and dy are parameters and change from 0 to 1.

[1.9] Mapping when a constraint condition is imposed Up to now, the determination of the map when no constraint condition exists has been described. However, when the correspondence relationship between the specific pixels of the start point image and the end point image is defined in advance, the mapping can be determined under the constraint condition.

The basic idea is that the starting point image is roughly transformed by a rough mapping in which a specific pixel of the starting point image is moved to a specific pixel of the ending point image, and then the mapping f is accurately calculated.

First, a rough mapping is determined in which a specific pixel of the starting point image is projected onto a specific pixel of the ending point image, and other pixels of the starting point image are projected onto appropriate positions. That is, a pixel close to a specific pixel is a map that is projected in the vicinity of the place where the specific pixel is projected. Here, the rough mapping of the m-th level is described as F ^(m) .

The rough map F is determined by the following procedure. First, the mapping is specified for some pixels. N _s pixels for the starting image,

[Equation 44] When specifying, determine the following values.

[Equation 45] The displacement amounts of the other pixels in the starting point image are p _{(ih, jh)} (h
= 0, ..., _Ns −1) is an average obtained by weighting the displacement. That is, the pixel p _{(i, j)} is projected onto the following pixels of the end point image.

[0089]

[Equation 46] However, here

[Equation 47]

[Equation 48] And

Subsequently, the energy D ^{(m, s)} _{(i, j)} of the mapping f is changed so that the candidate mapping f close to F ^(m) has less energy. To be exact, D
^{(M , s)} _{(i, j)} is

[Equation 49] Is. However,

[Equation 50] And κ and ρ ≧ 0. Finally, f is completely determined by the automatic map calculation process described above.

Where f ^{(m, s)} (i, j) is F
^(M) When they are sufficiently close to (i, j), that is, their distance is

[Equation 51] It should be noted that E ₂ ^{(m, s)} _{(i, j)} becomes 0 when it is within. The reason I defined it that way is
This is because as long as each f ^{(m, s)} (i, j) is sufficiently close to F ^(m) (i, j), its value is automatically determined so as to settle at an appropriate position in the end point image. For this reason, the starting point image is automatically mapped to match the ending point image without having to specify the exact correspondence in detail.

[2] The flow of processing by each elemental technique of the concrete processing procedure [1] will be described. Figure 6
Is a flowchart showing the overall procedure of the base technology. As shown in the figure, first, a process using a multi-resolution singularity filter is performed (S1), and then a start point image and an end point image are matched (S2). However, S2 is not essential,
Processing such as image recognition may be performed based on the characteristics of the image obtained in S1.

FIG. 7 is a flow chart showing the details of S1 in FIG. Here, it is assumed that the start point image and the end point image are matched in S2. Therefore, first, the starting point image is hierarchized by the singular point filter (S1
0), obtaining a series of starting point hierarchical images. Subsequently, the end point image is layered by the same method (S11) to obtain a series of end point hierarchical images. However, the order of S10 and S11 is arbitrary, and the starting point hierarchical image and the ending point hierarchical image can be generated in parallel.

FIG. 8 is a flow chart showing the details of S10 of FIG. The size of the original starting point image is 2 ⁿ × 2 ⁿ . Since the starting point hierarchical image is created in ascending order of resolution, the parameter m indicating the resolution level to be processed is set to n (S100). Next, the m-th level images p ^{(m, 0)} , p ^{(m, 1)} ,
Singular points are detected from p ^{(m, 2 )} and p ^{(m, 3)} using a singular point filter (S101), and the images at the ^{(m-1) th} level p ^(m-1,0) and p ^(m- ), respectively. ^1,1) , p
^{(M-1 , 2)} and p ^{(m-1, 3)} are generated (S10 ^).
2). Since m = n here, p ^{(m, 0)} = p
^{(M, 1)} = p ^{(m, 2)} = p ^{(m, 3)} = p ⁽ⁿ⁾ , and four types of sub-images are generated from one starting point image.

FIG. 9 shows a part of the m-th level image and the m-th image.
The correspondence relationship of a part of a 1-level image is shown. Numerical values in the figure indicate the brightness of each pixel. In the figure, p ^{(m, s)} is p
^{(M, 0) ~p (m} , 3) intended to symbolize the four ^images, when generating the ^{images p (m-1,0),} p
Consider ^{(m, s)} to be p ^{(m, 0)} . [1.2]
According to the rule shown in, p ^(m-1,0) is, for example, "3" out of the four pixels included in the block in which luminance is entered in the figure, and p ^(m-1,1) is "8". , P
^{(M-1 , 2)} is "6" and p ^{(m-1, 3)} is "10"
, And replaces this block with the acquired single pixel. Therefore, the size of the sub-image at the (m-1) th level is 2 ^m-1 × 2 ^m-1 .

Next, decrement m (S1 in FIG. 8).
03), and confirm that m is not negative (S10
4) Returning to S101, a sub-image with coarse resolution is generated next. As a result of this repeated processing, S = 0 ends when m = 0, that is, when the 0th level sub-image is generated. The size of the 0th level sub-image is 1 × 1.

FIG. 10 exemplifies the starting point hierarchical image generated in S10 in the case of n = 3. Only the first starting point image is common to the four series, and thereafter, the sub-images are independently generated according to the types of singular points. The process of FIG. 8 is common to S11 of FIG. 7, and the end point hierarchical image is also generated through the same procedure. With the above, S in FIG.
The process according to 1 is completed.

In the base technology, in order to proceed to S2 in FIG.
Prepare for the hatching evaluation. Figure 11 shows the procedure
There is. As shown in the figure, first, multiple evaluation formulas are set.
(S30). Regarding the pixels introduced in [1.3.2.1]
Energy C^{(M, s)} _fAnd [1.3.2.2]
Energy D related to the smoothness of the entered map^{(M, s)}
_fIs that. Next, we integrate these evaluation formulas and
A total evaluation formula is established (S31). Derived in [1.3.2.3]
Total energy input λC^{(M, s)} _f+ D^{(M, s)} _f
Is that, using η introduced in [1.3.2.2]
Then, ΣΣ (λC^{(M, s)} _{(I, j)}+ ΗE₀
^{(M, s)} _{(I, j)}+ E₁ ^{(M , S)} _{(I, j)})
(Expression 52) is obtained. However, the sum is the same for i and j
0, 1 ... 2 respectively^mCalculate with -1. Matching above
The evaluation is ready.

FIG. 12 is a flow chart showing the details of S2 in FIG. As described in [1], the matching between the starting point hierarchical image and the ending point hierarchical image is performed between images having the same resolution level. In order to achieve good global matching between images, matching is calculated in order of coarse resolution. Since the starting point hierarchical image and the ending point hierarchical image are generated using the singular point filter, the position and brightness of the singular point are clearly preserved even at the coarse level of resolution, and the result of global matching is better than before. It will be very good.

As shown in FIG. 12, first, the coefficient parameter η is set to 0 and the level parameter m is set to 0 (S20). Subsequently, matching is calculated between each of the four sub-images at the m-th level in the start point hierarchical image and the four sub-images at the m-th level in the end point hierarchical image, satisfying the bijective conditions, and 4 types of submappings f ^{(m, s)} (s = 0,1,2,3) that minimize
1). The bijection condition is checked using the inherited quadrilateral described in [1.3.3]. At this time, as shown in Expressions 17 and 18, the sub-mappings at the m-th level are constrained to those at the (m-1) -th level, so that matching at a coarser resolution level is sequentially used. This is a vertical reference between different levels. Note that m = 0 now and there is no coarser level, but this exceptional processing will be described later with reference to FIG.

On the other hand, horizontal reference within the same level
Done. As in Expression 20 of [1.3.3], f
^{(M, 3)}Is f^{(M, 2)}And f^{(M, 2)}Is f
^{(M, 1)}And f ^{(M, 1)}Is f^{(M, 0)}To that
It is decided to be similar. The reason is that the type of singularity is
Even if they are different, they are originally the same start image and end image.
Situation in which the submappings are completely different because they are included
Is unnatural. As you can see from Equation 20,
The closer the images are to each other, the smaller the energy
Goodness is considered good.

Since there is no submapping that can be referred to at the same level for f ^{(m, 0) to} be determined first, one coarse level is referred to as shown in Expression 19. However, in the experiment, after obtaining up to f ^{(m, 3),} the procedure was taken to update f ^{(m, 0)} once with this as a constraint condition. This is equivalent to substituting s = 4 into Equation 20 and making f ^{(m, 4)} a new f ^{(m, 0)} . f ^{(m, 0)}
This is to avoid the tendency that the degree of association between f and ⁽ f ^{(m, 3))} becomes too low, and this measure improved the experimental results. In addition to this measure, in the experiment, the shuffling of the submapping shown in [1.7.1] was also performed. This is also to keep close the degree of association between submappings that are originally determined for each type of singularity. Further, the point of changing the position of the starting point according to the value of s in order to avoid the deflection depending on the starting point of the processing is as described in [1.7].

FIG. 13 is a diagram showing how the submapping is determined at the 0th level. At level 0, each sub-image consists of only one pixel, so four sub-maps
^{All f (0, s )} are automatically determined as unit maps. Figure 1
FIG. 4 is a diagram showing how a submapping is determined at the first level. At the first level, each sub-image consists of 4 pixels. In the figure, these four pixels are shown by solid lines.
Now, when searching the corresponding point of the point x of p ^{(1, s)} in q ^{(1, s)} , the following procedure is performed.

1. The upper left point a, the upper right point b, the lower left point c, and the lower right point d of the point x are obtained at the first level resolution. 2. A pixel to which the points a to d belong at one coarse level, that is, the 0th level is searched for. In the case of FIG. 14, points a to d belong to pixels A to D, respectively. However, the pixels A to C are virtual pixels that do not originally exist. 3. Plot the corresponding points A ′ to D ′ of the pixels A to D, which have already been obtained at the 0th level, in q ^{(1 , s)} . Pixels A ′ to C ′ are virtual pixels, and are pixels A to C, respectively.
Shall be in the same position as. 4. The corresponding point a ′ of the point a in the pixel A is considered to be in the pixel A ′, and the point a ′ is plotted. At this time, the position of the point a in the pixel A (lower right in this case) and the point a ′
Are assumed to occupy the same position in pixel A '. Plot corresponding points b ′ to d ′ in the same way as in 5.4,
Make an inherited quadrilateral at points a'-d '. 6. Find the corresponding point x ′ of the point x so that the energy is minimized in the inherited quadrilateral. The candidates of the corresponding point x ′ may be limited to those in which the center of the pixel is included in the inherited quadrangle, for example. In the case of FIG. 14, all four pixels are candidates.

The above is the procedure for determining the corresponding point of the point x. Similar processing is performed for all other points to determine the submapping. At the second level and above, it is considered that the shape of the inherited quadrangle gradually collapses, so that a situation occurs in which the pixels A ′ to D ′ are spaced apart as shown in FIG. 3.

Thus, if four sub-maps of a certain m-th level are determined, m is incremented (S2 in FIG. 12).
2) Make sure that m does not exceed n (S2
3) and returns to S21. Hereinafter, each time the process returns to S21, the sub-mapping of the finer resolution level is obtained, and when the process finally returns to S21, the n-th level mapping f ⁽ⁿ⁾ is determined. Since this mapping is fixed for η = 0, f
^(N) Write (η = 0).

Next, the maps for different η should be obtained.
Shift η by Δη and clear m to zero (S2
4). New η is the predetermined search cutoff value η_maxBeyond
Confirm that it is not (S25), return to S21,
map f with respect to η^(N)Calculate (η = Δη). This place
The process is repeated, and in S21, f^(N)(Η = iΔη) (i =
0, 1, ...) η is η_maxWhen exceeded
Proceeding to S26, the optimum η = η by the method described later._optDecide
And f^(N)(Η = η_opt) Is finally a map f ^(N)
And

FIG. 15 is a flow chart showing the details of S21 of FIG. This flowchart determines the submapping at the m-th level for a certain η. When determining the submappings, the precondition technique independently determines the optimum λ for each submapping.

As shown in the figure, first, s and λ are cleared to zero.
(S210). Next, for λ at that time (and
And implicitly for η) the submapping f that minimizes the energy
^{(M, s)}(S211), and this is f
^{(M, s)}Write (λ = 0). Also the mappings for different λ
To obtain, shift λ by Δλ, and set a new λ
Search cutoff value λ_maxConfirm that it does not exceed (S
213), the process returns to S211 and f is repeated in the subsequent processing.
^{(M, s)}(Λ = iΔλ) (i = 0,1, ...)
It λ is λ _maxWhen it exceeds, the process proceeds to S214 and the optimum
λ = λ_optAnd f^{(M, s )}(Λ = λ_opt)
Finally the map f^{(M, s)}(S214).

Next, in order to find another submapping at the same level, λ is cleared to zero and s is incremented (S215). It is confirmed that s does not exceed 4 (S216), and the process returns to S211. When s = 4, f ^{(m, 3 )} is updated using f ^{(m, 3)} as described above,
Finish the submapping determination at that level.

FIG. 16 shows that f ^{(m, s)} (λ = iΔλ) (i = 0,
It is a figure which shows the behavior of the energy C ^{(m, s)} _f corresponding to 1, ...). As described in [1.4], when λ increases, C ^{(m, s)} _f usually decreases. However, when λ exceeds the optimum value, C ^{(m, s)} _f starts to increase. Therefore, in the base technology, λ when C ^{(m, s)} _f has a minimum value is determined as λ _opt . Even if C ^{(m, s)} _f becomes small again in the range of λ> λ _opt as shown in the figure, the map is already broken at that point and it does not make sense, so pay attention to the first minimum point. Good. λ _opt is independently determined for each submapping, and finally f ⁽ⁿ⁾ is also determined.

On the other hand, FIG. 17 is a diagram showing the behavior of the energy C ⁽ⁿ⁾ _f corresponding to f ⁽ⁿ⁾ (η = iΔη) (i = 0, 1, ...) Obtained while changing η. . Again, when η increases, C ⁽ⁿ⁾ _f usually decreases, but when η exceeds the optimum value, C ⁽ⁿ⁾ _f starts to increase. So C
^(N) Determine η when _f takes a minimum value as η _opt .
It can be considered that FIG. 17 is an enlarged view of the vicinity of zero on the horizontal axis of FIG. Once η _opt is determined, f ⁽ⁿ⁾ can be finally determined.

As described above, according to the base technology, various advantages can be obtained. First, since it is not necessary to detect edges, it is possible to solve the problems of the edge detection type prior art. Also,
A priori knowledge of the objects contained in the image is not required, and the corresponding points are automatically detected. The singular point filter can maintain the brightness and position of the singular point even at a coarse resolution level, and is extremely advantageous for object recognition, feature extraction, and image matching. As a result, it is possible to construct an image processing system that significantly reduces manual work.

Note that the following modified technique can be considered for the base technique. (1) In the base technology, the parameters are automatically determined when matching is performed between the starting point hierarchical image and the ending point hierarchical image, but this method does not perform matching between hierarchical images but ordinary two images. Available in all cases.

For example, between two images, the energy E ₀ related to the difference in pixel luminance and the energy E ₁ related to the positional shift of pixels are used as evaluation formulas, and the linear sum E _{t ot} = αE ₀ + E _{1 of these is used.} Is a comprehensive evaluation formula. Paying attention to the vicinity of the extreme value of this comprehensive evaluation formula, α is automatically determined. That is, a map that minimizes E _tot is obtained for various α. Among these mappings, α when E ₁ has a minimum value with respect to α is determined as the optimum parameter. The map corresponding to that parameter is finally regarded as the optimum matching between both images.

Other than this, there are various methods for setting the evaluation formula. For example, 1 / E ₁ and 1 / E ₂ that have a larger evaluation result may be used. . The total evaluation formula does not necessarily have to be a linear sum, and n
The sum of multiplications (n = 2, 1/2, -1, -2, etc.), a polynomial, an arbitrary function, etc. may be appropriately selected.

The parameter may be either only α, two cases of η and λ as in the base technology, or more cases. If the parameter is 3 or more, change it one by one and decide.

(2) In the base technology, after the mapping is determined so that the value of the comprehensive evaluation formula is minimized, one evaluation formula C ^{(m, s)} _f constituting the comprehensive evaluation formula becomes minimum. The points were detected and the parameters were determined. However, instead of such a two-step process, it may be effective to simply determine the parameters so that the minimum value of the comprehensive evaluation formula is minimized depending on the situation. In that case, for example, αE ₀ + βE ₁ may be used as a comprehensive evaluation formula, and a constraint condition of α + β = 1 may be provided to treat each evaluation formula equally. This is because the essence of automatic parameter determination is to determine the parameters so that the energy is minimized.

(3) In the base technology, four types of sub-images relating to four types of singular points are generated at each resolution level. However, naturally, one, two, and three types among the four types may be selectively used. For example, if there is only one bright point in the image, a corresponding effect should be obtained even if a hierarchical image is generated only with f ^{(m, 3)} related to the maximum point.
In that case, different submappings at the same level are not necessary, which has the effect of reducing the amount of calculation regarding s.

(4) In this base technology, when the level advances by one with the singularity filter, the number of pixels becomes 1/4. For example, it is possible to use 3 × 3 as one block and search for a singular point in the block. In that case, the pixel becomes 1/9 when the level advances by one.

(5) When the starting point image and the ending point image are in color, they are first converted into a black and white image and the mapping is calculated.
The color image at the starting point is converted using the mapping obtained as a result. As another method, a submapping may be calculated for each of the RGB components.

[Embodiment Regarding Image Editing] A specific example of image editing partially utilizing the above-described prerequisite technology will be described. Figure 1
Reference numeral 8 shows the configuration of the image editing apparatus 10 according to the embodiment.
The image editing apparatus 10 includes an image input unit 12 for inputting a key frame of a moving image to be edited (hereinafter, also referred to as “original key frame”) from a network or a recording device, and a key for storing the original key frame. In the entire column of the frame storage unit 14 and the original key frame and the key frame (indicated by “KF” in the figure) including the new key frame described later, if necessary, image matching by the pre-requisite technique between adjacent key frames. And a matching processor 16 for generating corresponding point information (indicated as “MF” in the figure) by calculating the intermediate point and an intermediate frame (indicated as “IF” in the figure) by interpolation calculation based on the generated corresponding point information and the key frame based on the base technology ) Is generated, and the key frame and the generated intermediate frame are sequentially arranged. And a display unit 20 for displaying a picture. However, the display unit 20 has a function of displaying a key frame and an intermediate frame one by one in accordance with an instruction from an image editor 24, which will be described later, and allows a user to edit an image.

The intermediate frame generated by the intermediate image generator 18 is also sent to the intermediate frame buffer 22 and stored therein. The image editor 24 is a functional block having an image editing function, and has functions such as changing the color of an image, trimming, enlarging / reducing, and various filtering processes. The image editor 24 is the intermediate frame buffer 22.
The user can edit the intermediate frame stored in.

The registration unit 26 registers or stores the edited intermediate frame (denoted as "EIF" in the drawing) in the key frame storage unit 14 as a new key frame (denoted as "NKF" in the drawing). The registration unit 26 may automatically or forcibly register the intermediate frame when any edit is made to the intermediate frame, or the confirmation unit 28.
May once ask the user a question such as "Do you want to register this intermediate frame as a key frame?", And wait for the user's answer to it before registering.

The stream generation unit 30 finally determines the determined key frames, that is, all the original key frames and the newly registered key frames, and the final corresponding points recalculated by the matching processor 16 for these key frames. The information is acquired, and the encoded data CDS is generated by incorporating the information. This data is sent to the network via the communication unit 32, received by an image reproducing device (not shown), and stored in a storage device (not shown).

FIG. 19 is a flow chart showing an example of processing by the image editing apparatus 10. When the editing is started, the original key frame is first read by the image input unit 12 and stored in the key frame storage unit 14 (S10). Subsequently, the matching processor 16 calculates corresponding point information between adjacent key frames (S12), and sends the information to the intermediate image generation unit 18. The intermediate image generation unit 18 generates an intermediate frame by interpolation calculation based on the key frame and the corresponding point information (S14).

Next, it is confirmed whether or not the operation mode of the image editing apparatus 10 is currently the "editing mode" (S1).
6). Image editor 2 in edit mode or normal mode
Suppose it is specified from 4. In the normal mode (Y in S16), the key frames and the intermediate frames are arranged in the order of display time, and are normally reproduced on the display unit 20 at a rate of 30 sheets per second (S18). After that, the stream generation unit 3
It is encoded with 0 (S20) and output. At the time of encoding, it is not necessary to restart the matching processor 16 because the original key frames and the corresponding point information generated for them are valid unless a new key frame is added.

On the other hand, in the edit mode (N of S16),
After the key frames and the intermediate frames are arranged in the order of display time, the frames are sequentially reproduced one by one (S22). While confirming the frame to be reproduced, the user stops the reproduction via the image editor 24 when an intermediate frame that requires editing appears, and actually proceeds to the editing work (Y in S24). When the user manually edits (S28) and this is completed, the confirmation unit 28 inquires whether or not to register the intermediate frame as a new key frame, and when "register" is selected (S30). Y), the registration unit 26 actually stores the key frame as the key frame storage unit 14
Is additionally registered to (S32). The registration unit 26 may forcibly register the edited intermediate frame as a new key frame without the confirmation by the confirmation unit 28.

After that, if the editing is completed (Y of S26), the stream generation unit 30 performs encoding (S3).
2). At this time, the matching processor 16 recalculates corresponding point information for the finally obtained key frame, that is, the entire original key frame and the new key frame through a processing path (not shown), and stores the same in the encoded data. It

When N is selected in each of S24 and S26, that is, when the editing is not started in the frame-by-frame display and the edited intermediate frame is not registered as a key frame, the process proceeds to S26. Unless the editing is completed, the process proceeds to reproduction and display of the next frame (S22) and the process is repeated.

FIG. 20 shows the structure of another image editing apparatus 100. In the figure, the same components as those in FIG. 18 are designated by the same reference numerals, and the description thereof will be appropriately omitted. In FIG. 18, the intermediate frame edited by the user is upgraded to the key frame, but in FIG. 20, the existing intermediate frame is upgraded to the key frame. When shooting a moving image, for example, 30 frames per second can be obtained. The apparatus of FIG. 18 only needs to determine the original key frame for every ten such frames, for example, and the section lacking in it is an intermediate frame generated by the intermediate image generation unit 18 (this is also called a virtual intermediate frame). I got a new keyframe by modifying (Yobu). However, if you keep the original video, you can upgrade the intermediate frames that actually exist between the key frames (also called actual intermediate frames) to key frames without changing the virtual intermediate frames. Is enough.

For this reason, the original moving image is stored in any place, and the user extracts the actual intermediate frame at the time position corresponding to the virtual intermediate frame whose image quality is a problem from the moving image as a key frame. Promote.

The new structure in FIG. 20 is replaced with a frame storage unit 102 for collectively holding not only key frames but also all or a number of frames necessary for realizing this apparatus, and an image editor 24. The instruction acquisition unit 104 is a user interface that receives an “image quality improvement instruction” from the user.

The frame storage unit 102 is provided with a key frame table (not shown) in addition to the frames. The frame number of the key frame is described in this table. For example, initially, "0, 9, 19, 2
9 ,. ． ． , Every 10 sheets are designated as original keyframes. Matching processor 16
Reads out the key frame with reference to this table.

On the other hand, it is assumed that the user has found a virtual intermediate frame having a problem in image quality while the moving image is being reproduced on the display unit 20. At this time, the user uses the instruction acquisition unit 1
Give instructions via 04. This instruction is sent to the registration unit 26, the virtual intermediate frame displayed at that time is specified, and subsequently, the two key frames from which this virtual intermediate frame is generated are specified as the section.

If these keyframes are frame 39
And the frame 49, the registration unit 26 registers
49], the number of the existing intermediate frame is specified, and this is stored in the frame storage unit 102 as “new key frame NK.
Notify as "F". The frame storage unit 102 adds the number to the table. After that, the matching processor 1
6 generates corresponding point information including the key frame,
This is reflected in the image generation by the intermediate image generation unit 18 and the encoded data generation by the stream generation unit 30.

The number of the new key frame by the registration unit 26 may simply be a convenient one such as the center of the above-mentioned section. For example, for [39,49], "35" is notified to the frame storage unit 102. Of course, the number of the virtual intermediate frame that originally caused the problem may be directly notified to the frame storage unit 102. In either case, it is basically that a new key frame is additionally registered in the key frame section sandwiching the frame in question.

The image editing technique has been described above. It should be understood by those skilled in the art that the above specific examples are merely examples, and that various modifications are present. Here is an example.

In FIG. 18, the image input unit 12 inputs only the key frame. However, when the corresponding point information exists outside in advance, the corresponding point information may be input in addition to the key frame.

In FIG. 18, there are various methods for editing by interaction between the image editor 24 and the user. There are many known techniques for stopping display in an intermediate frame and processing the frame with retouching software. The image editor 24 may have such an arbitrary function.

Although the stream generation unit 30 incorporates the original key frame and the new key frame in the encoded data, it may simply output the original key frame and the new key frame. Furthermore, it is also effective to output a new key frame. This is because it is possible to obtain the new key frame, externally generate corresponding point information, and generate encoded data.
From this point of view, when outputting a new key frame, it is desirable to output the information specifying the frame, for example, the frame number.

In FIG. 20, the frame storage unit 102 holds a large number of frames. However, in order to reduce the capacity of the frame storage unit 102, an intermediate frame necessary at a necessary timing is appropriately acquired from the network or the like via the image input unit 12. May be.

[Brief description of drawings]

1A and 1B are images obtained by applying an averaging filter to the faces of two persons, and FIG.
(C) and FIG. 1 (d) are images of p ^{(5, 0)} required by the base technology for the faces of the two persons, and FIGS. 1 (e) and 1 (f) are the faces of the two persons. Images of p ^(5,1) required by the base technology, Fig. 1 (g) and Fig. 1 (h)
Is the p required by the base technology for the faces of two people.
The images of ^{(5, 2)} , Fig. 1 (i) and Fig. 1 (j) are p ^{(5, 3)} obtained by the base technology for the faces of two persons.
3 is a photograph of a halftone image in which each of the images is displayed on the display.

2 (R) is a diagram showing an original quadrangle, FIG.
(A), FIG. 2 (B), FIG. 2 (C), FIG. 2 (D), FIG.
(E) is a figure which shows each inherited quadrangle.

FIG. 3 is a diagram showing a relationship between a start point image and an end point image and a relationship between an mth level and an m−1th level by using an inherited quadrangle.

FIG. 4 is a diagram showing a relationship between a parameter η and energy C _f .

5 (a) and 5 (b) are diagrams showing how to determine from a cross product calculation whether or not a mapping regarding a certain point satisfies a bijection condition.

FIG. 6 is a flowchart showing the overall procedure of the base technology.

FIG. 7 is a flowchart showing details of S1 in FIG.

FIG. 8 is a flowchart showing details of S10 of FIG.

FIG. 9 is a diagram showing a correspondence relationship between a part of an m-th level image and a part of an m-1th level image.

FIG. 10 is a diagram showing a starting point hierarchical image generated by the base technology.

FIG. 11 is a diagram showing a procedure of preparation for matching evaluation before proceeding to S2 of FIG.

FIG. 12 is a flowchart showing details of S2 in FIG.

FIG. 13 is a diagram showing how a submapping is determined at the 0th level.

FIG. 14 is a diagram showing how a submapping is determined at the first level.

FIG. 15 is a flowchart showing details of S21 in FIG.

FIG. 16 is a diagram showing a behavior of energy C ^{(m, s)} _f corresponding to f ^{(m, s)} (λ = iΔλ) obtained by changing λ for a certain f ^{(m, s)} .

FIG. 17 shows f ⁽ⁿ⁾ (η = determined by changing η
energy C corresponding to iΔη) (i = 0, 1, ...)
It is a figure which shows the behavior of ⁽ⁿ⁾ _f .

FIG. 18 is a configuration diagram of an image editing apparatus according to an embodiment.

FIG. 19 is a flowchart showing a procedure of editing processing by the apparatus of FIG.

FIG. 20 is a configuration diagram of another image editing apparatus according to the embodiment.

[Explanation of symbols]

10,100 image editing device, 12 image input unit,
16 matching processor, 18 intermediate image generation unit, 24 image editor, 26 registration unit, 28 confirmation unit, 30 stream generation unit, 104 instruction acquisition unit.

Claims (18)

[Claims] 1. An image editing method, wherein when an intermediate frame generated based on a key frame is edited, the edited intermediate frame is registered as a new key frame. 2. A step of receiving a user's edit of an initial intermediate frame generated based on a key frame, and a step of inquiring of the user whether or not to register the edited intermediate frame as a new key frame. An image editing method comprising: 3. The method according to claim 2, further comprising the step of forcibly registering the edited intermediate frame as a new key frame, instead of the inquiring step. 4. The method according to claim 2, further comprising: obtaining corresponding point information between the keyframes. 5. The method according to claim 4, further comprising: generating the initial intermediate frame by interpolation calculation based on the corresponding point information. 6. The method according to claim 2, further comprising the step of regenerating intermediate frames, including new keyframes that have been registered. 7. The method of claim 6, further comprising: displaying the regenerated intermediate frame; and re-accepting edits to the displayed intermediate frame. 8. The method according to claim 4, further comprising: generating an encoded data stream including a plurality of keyframes including a new registered keyframe and corresponding point information. 9. An editor for accepting user's editing of an initial intermediate frame generated based on a key frame, and a confirmation unit for inquiring of the user whether to register the edited intermediate frame as a new key frame. And an image editing apparatus comprising: 10. The apparatus according to claim 9, further comprising a registration unit that forcibly registers the edited intermediate frame as a new key frame, instead of the confirmation unit. 11. The apparatus according to claim 9, further comprising a matching processor that acquires corresponding point information between the key frames. 12. The method according to claim 9, further comprising an input unit that acquires corresponding point information between the key frames from the outside.
The device according to. 13. The apparatus according to claim 11, further comprising an intermediate image generation unit that generates the initial intermediate frame by interpolation calculation based on the corresponding point information. 14. The apparatus according to claim 13, wherein the intermediate image generation unit regenerates an intermediate frame including a new registered key frame. 15. The apparatus of claim 14, wherein the editor accepts edits for the regenerated intermediate frame again. 16. The stream generation unit further includes a stream generation unit that generates encoded data in a form including a plurality of key frames including a registered new key frame and corresponding point information.
The method described in 2. 17. A computer program causing a computer to execute the image editing method according to claim 2. 18. A computer program for causing a computer to execute the image editing method according to claim 3.