JP2839132B2 - Image cut point detection method and apparatus - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method and apparatus for detecting a video cut point, and more particularly, to a method and apparatus for detecting a cut point (a point at which a scene is switched) from a plurality of image data columns.

[0002]

2. Description of the Related Art A point at which a scene is switched by turning on / off a camera in a video or by a joint (fade, wipe, etc.) in video editing is called a cut point. Camera movement (pan, zoom, etc.) and movement of objects in the image are not considered cuts. Such detection of an image cut point is also called scene change detection, and various methods have been conventionally proposed.

As a typical method, two images I _t adjacent temporally in a series of images captured, by calculating the difference in luminance values of each corresponding pixel in I _t-1, the Assuming that the sum of the absolute values over the entire screen (inter-frame difference) is D (t), when D (t) is greater than a given threshold, t (t)
Is a cut point (Otsuji, Totomura, Oba:
"Moving image browsing using luminance information", IEICE technical report, IE90-103, 1991). In this case, instead of the inter-frame difference, pixel change area, the luminance histogram difference, block-specific color correlation, and chi ² test amount D (t)
(Otsuji, Totomura: "Study of Automatic Video Cut Detection Method", Technical Report of the Institute of Television Engineers of Japan, Vo
l.16, No.43, pp.7-12). This method has a problem that even if there is a strong movement of an object or flash light in an image, it is erroneously detected as a cut.

There is also a method of performing threshold processing on a result obtained by applying various time filters to D (t) instead of performing threshold processing on D (t) as it is (K. Otsuji snd Y. Tonomur).
a: "Projection Detecting Filter for Video Cut Detec
tion "Proc. of ACM Multimedia 93, 1993, pp. 251-257).
This method is characterized in that erroneous detection is unlikely to occur even if there is a strongly moving object or flash light in the video.

[0005]

The above prior art has the following problems. The first problem is that temporally slow scene changes cannot be detected. This is because, in the conventional cut point detection method, the amount representing the rate of change of the scene is calculated from only two frames that are temporally coincident, and a long-term change in the scene is hardly reflected.

A typical example of a slow scene change is an effect (special effect) such as a fade-in, fade-out, dissolve, or wipe inserted at a video editing stage. Fade-in is a technique for gradually increasing the video signal level to gradually show an image, and fade-out is a technique for gradually lowering the video signal level and gradually erasing the image. When the dissolve moves from the scene A to the scene B, the scene B is lowered while the video signal level of the scene A is lowered.
Is a technique for increasing the video signal level of a scene, and the scene is changed so that the scene A and the scene B are fused to each other. The wipe is a technique for displaying the image of the scene B while erasing the image of the scene A from part of the image. With these effects, the scene changes slowly (for example, it takes about one second until the scene completely changes), so two images that are temporally adjacent to each other (the time interval is 1/30 second) The degree) could not be detected. This is because the difference between two images that are temporally adjacent to each other in a slow scene change is so small that it is caused by noise.
This is because it is difficult to determine whether the change is caused by a scene change.

A second problem is that if a flash light appears in an image, it is erroneously detected as a cut. When the flash is turned on, the brightness level of the image rises for a moment, so that when the flash is turned on, the difference between the previous image and the image in which the flash light appears is sharply increased. Therefore, in the conventional cut point detection method in which a sudden increase in the luminance difference is regarded as a cut point, the flash light is mistaken for the cut point.

A third problem is that a cut point of a telecine-converted image cannot be correctly detected (it is easy to miss). The telecine conversion is a conversion for matching the number of frames of a film (24 frames per second) with the number of frames of a television signal (30 frames per second) when an image photographed on a film is broadcast on a television. There are several methods for telecine conversion, but in one method, one frame is added to every four frames. Since the frame to be inserted at this time is created by combining the fields of the previous and next images,
The change in the image becomes small, and the conventional cut point detection method is liable to cause cut detection omission for the same reason as the dissolve described above.

As for the second and third problems, there is a method of preventing erroneous detection and omission of detection by applying a time filter. However, when designing this time filter, it is necessary to check in advance what kind of image the image is, and it is not suitable for real-time processing, and it is erroneous that flash light appears in the telecine-converted image. There is a problem that detection occurs.

[0010]

SUMMARY OF THE INVENTION In view of the above-mentioned problems of the prior art, it is an object of the present invention to detect a scene change which is slow in time and to display a telecine-converted image, an image including flash light, and the like. An object of the present invention is to provide an image cut point detection method and apparatus that can be integrated and processed in real time.

In order to achieve the above object, an image of the present invention is provided.
According to the image cut point detection method, the input image data
At the current time t and before the J frame_t, I_t-1,
…, I _tJAll sets of images satisfying 0 ≦ i <j ≦ J
Data I_ti, I_tjCalculate the distance d (t-i, t-j) between
Create a table and set the distance d (t-i, t-
j)_CFrom (t-J) to t
Scene change rate C (t-j_C) To the frame (t
-j_C), The scene change rate
Rate C (t-j_C) Is compared with a predetermined threshold,
(T-j_C) Is determined as a cut point.

A video cut point detecting apparatus according to the present invention comprises: a buffer memory means for sequentially buffering at least J + 1 frames of image data which are continuous in time; the image data _{I t, I t- 1, ...} , the I _tJ, 0 ≦
image-to-image distance calculating means for calculating the distance between the image data of the two images _Iti and _Itj in the range of i < _j≤J , and a distance table for storing the image-to-image distance calculated by the image-to-image distance calculating means Means, by referring to the distance table means, a scene change rate calculating means for calculating a change in the image from the frame (tJ) to t as a scene change rate in the frame (tj _C ); and the calculated scene change rate. Is compared with a predetermined threshold value to determine whether or not the frame (tj _C ) is a cut point.

In the present invention, the scene change rate C (tj _C ) is taken into consideration not only by comparing two images that are temporally adjacent to each other as a screen change amount, but also by taking into account the change amount between images that are temporally separated.
Is calculated. Thereby, the cut point can be detected stably. Now, assuming that the scene is switched at the time of * in the image among the image data strings A, B, *, D, and E, A and B before the cut are images similar to each other, and * after the cut. , D, and E are also similar to each other, so that the distances d (A, B), d (*, E), and d (D, E) are small, while the distance d ( A, *), d (A, D),
d (A, E), d (B, *), d (B, D), d (B, E) increase. If the scene change rate C (*) is obtained using this property, cut detection can be performed.

In this way, fade, dissolve,
Even when the scene changes slowly over time, such as a wipe, the scene change can be detected. That is, image sequences A, B, *, including scene changes slowly in time.
For D and E, the screen change amounts d (A, D), d (A, E), d (B, D), and d (B, E) between two images separated in time are Since it takes a large value, a scene change can be detected by actively using this.

Further, even if the flash light appears in an image, the flash light is not erroneously regarded as a cut. The image data sequence including the flash light is represented by A, B, *,
Assuming that the image * is brighter than the other images A, B, D and E due to the flash light,
The distance d (A, B) between the image A and the image B is small,
Since the brightness level of the image * is higher than that of D, d
(B, *) and d (*, D) have large values. Some conventional methods determine a point where d (B, *) becomes larger than a certain threshold value as a cut point, and thus have a disadvantage that flash light is erroneously recognized as a cut point. The fact that d (B, *) increases at the cut point is the same, but the characteristic property of flash light is that d (*, D) has a large value. The flash light can be distinguished, so that erroneous recognition can be prevented.

Further, no omission of detection occurs even for an image subjected to telecine conversion. Assuming that images A, B and images *, D, E are different scenes among A, B, *, D, E... To which telecine conversion has been applied, image * becomes image B due to telecine conversion. In some cases, the image is a composite image obtained by superimposing D. In this case, it is often impossible to detect a scene change despite the change.
This is the distance between adjacent images d (A, B), d (B, *),…,
This is because d (D, E) does not increase rapidly like a normal cut and does not exceed a predetermined threshold. However, also in this case, the distance d (B, D) between the two images straddling the image *
It can be understood that the time * is the cut point if attention is paid to the fact that becomes larger.

[0017]

An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing the basic configuration of the present invention. In FIG. 1, an image data stream I captured from an image signal source (not shown) at an appropriate sample rate is shown.
_0, I _1, ... are inputted to the buffer memory 11. Here, the image data captured for each screen is referred to as one frame of image data, and a series of frame numbers are assigned to the image data of the series of frames. The image sampling rate, image format, and image size may be arbitrary. That is, the NTSC standard video signal is transmitted at 30 frames /
Sampling may be performed at sec or sampling at a coarser sampling rate. Also,
The image data sequence is an analog signal such as NTSC,
It may be a digital signal. The external image signal source may be an image file stored in a storage device such as a hard disk or a CD-ROM. Buffer memory 11 the image data frame sequence _{I 0, I 1, ...,} I t are sequentially input, always present time t and, then going back to the front J point J + 1 pieces of image data I _t, I _t-1, ..., Buffer _ItJ temporarily. The buffer memory 11 is configured by, for example, a shift buffer array, buffers input image data while sequentially shifting the image data for each frame, and discards old image data. Therefore, the buffer memory 11 always holds a predetermined number of image data in the order of input. Note that the buffer memory 11 may store data obtained secondarily by processing image data. For example, it may store luminance histogram or color histogram data, or may store edge information. Here, these are collectively referred to as image data.

The image data I _ti of two frame image distance calculating unit 12 every time the image data I _t is input and I _tj
, I.e., the image distance d (ti, tj) is calculated for i = 0 and j = 1, 2,... The inter-image distance is a quantity representing a so-called “difference” between two images. If the two images _Iti and _Itj are exactly the same, d (ti, tj) = 0, and the “difference” The larger "is, the larger the value. Image distance calculation unit 12
Are stored in the distance table memory unit 13. In the inter-image distance calculation unit 12,
Image data I _t of the new frame in the buffer memory 11
Is stored in the buffer memory 11
, And calculates the inter-image distance with the images It ₁ , It ₂ ,... Of the other frames, and updates the contents of the distance table memory unit 13. The distance table memory unit 13 temporarily stores, as a table, d (ti, tj) for all pairs of i and j satisfying 0 ≦ i <j ≦ J at a time point (frame number) t. In the following description, the distance table memory unit 13
It is assumed that the distance table created in the table is also represented by the same reference numeral 13. FIG. 1 shows a case where J = 3. The new I _t was calculated by the distance calculation unit 12 each time it is input to the buffer memory 11 i = 0, j = 1,2 , ..., between images in J distances d (t, t-
j) is written into the position (0, j) in the distance table 13 (corresponding to the address in the table memory unit 13), and the previous position (i, j) in the distance table 13 is written.
The distance data d (ti, tj) specified by
To the position (i + 1, j + 1) within
+1) is discarded if it is out of the range of 0 ≦ i + 1 <j + 1 ≦ J.

The scene change rate calculating section 14 calculates a scene change rate C (tj _C ) with reference to the distance table 13. The scene change rate C (tj _C ) is a quantity representing the certainty of a so-called cut point, and takes a value of −1 or more and 1 or less in the embodiment.
It is configured so that the closer to 1, the higher the probability of a cut,
When C (tj _C ) is larger than a certain threshold, the determination unit 15 determines that the time t is a cut point. Scene change rate C (t
function defining the -j _C) chooses the optimal function according to the type of cut, as will be described later.

The operations of the inter-image distance calculation unit 12, the distance table memory unit 13, and the scene change rate calculation unit 14 will be described below based on specific examples. Now, the image data string _{I t-7, I t-} 6 for each frame is obtained in the buffer memory 11 to the present time t, ..., I _t is assumed to become the sequence shown in FIG. In FIG. 2, frames It _-7 , It _-6 , I
_t-5 and It _-4 are images in which the entire screen is black (the luminance value of all pixels is the minimum 0). _{I t-3, I t-} 2, I t-1, I t is the entire screen white (B _MAX luminance value maximum of all pixels). That is, image I
frame number (corresponding to time) t-3 corresponding to the _t-3 is considered to be a cut point t _C.

First, some examples of functions that define the image distance calculated by the inter-image distance calculation unit 12 will be described below. In the first embodiment of the distance function, the sum d (ti, tj) = Σ _{x, y} | I _i (x, y) of all pixel positions (x, y) of the absolute value of the luminance difference between frames −I _j (x, y) ｜ / (B _MAX × P) (1) is used. Here, P represents the total number of pixels on one screen, and I _i (x, y) and I _j (x, y) represent the luminance values of the pixels (x, y) of the images _Iti and _Itj , respectively. The luminance value is represented by an intensity corresponding to the luminance. When the image is a color image, the luminance value is represented by each intensity of the R, G, and B image signals. In this example, the sum of the differences between frames is normalized by the brightness when all pixels have the maximum brightness _BMAX . Therefore, the distance is in the range of 0 ≦ d (ti, tj) ≦ 1. For example, if B _MAX = 255, the distance between the images I ₀ and I ₄ in FIG. 2 is d (t, t-4) =-4
_{x, y} | 0− _BMAX | / ( _BMAX × P) = 1.0 is calculated.

A second embodiment of the distance function is a method for comparing the luminance values or the color histograms of two images. The luminance value histogram is defined as the number of pixels belonging to each luminance class (accumulated value or frequency, when the luminance of each pixel in an image of one frame is classified into predetermined N levels of luminance. Call). Images _Iti and
The histograms of the luminance values of _Itj are expressed as H _i (n) and H
_{Assuming that j} (n), n = 1, 2,..., N, the distance function is expressed by the following equation: d (ti, tj) = _{ | H _i (n) −H _j (n) | Expressed from 1 to N. Here, N is the number of steps in the histogram. For color histogram, R, G, a histogram of the B signal levels respectively _{H R (n), H G} (n), when the H _B (n), the inter-image distance is expressed by the following equation.

D (ti, tj) = Σ ｛| H _Ri (n) −H _Rj (n) | + | H _Gi (n) −H _Gj (n) | + | H _Bi (n) −H _Bj ( n) | {(3)} is from n = 1 to N. The third example of realizing the distance function is a statistic representing the deviation from the center of the histogram of the difference between the pixels of the images _Iti and _Itj from d.
(ti, tj). First, two images _Iti , I
A difference δ (x, y) = I _i (x, y) −I _j (x, y) between _tj is obtained for all pixels. Here, I _i (x, y) and I _j (x, y) represent the luminance value at the pixel (x, y) of the images _Iti and _Itj , respectively. Examples of the histogram of the difference are shown in FIGS. 3A and 3B. FIG. 3A shows that the images _Iti , _Itj belong to the same scene,
That is, a typical histogram shape in a case where a cut point is not included between two images is shown. As shown
Since the change in the luminance value is small in most of the pixels even if the object in the video slightly moves, the difference is concentrated near 0, and the cumulative value in the histogram decreases as the distance from 0 increases. In contrast, in Figure 3B, an image I _ti
An example of a histogram shape when a cut point is included between _Itj is shown. As shown in the figure, unlike FIG. 3A, the histogram does not have a steep peak near 0, but has a high-tailed and wide peak. Therefore, a function representing the distance may be defined so that the higher the base of the histogram, the greater the distance between images. As a specific example of a statistic in which d (ti, tj) increases as the tail of the histogram increases, the standard deviation ε of the difference δ (x, y) may be used as a distance function as in the following equation.

D (ti, tj) = ε = ｛Σ _{x, y} [δ (x, y) −δ ′] ² ｝ ^1/2 (4) where δ ′ is δ (x, y ) Means the average. Further, the following equation d (ti, tj) as a fourth distance function _{= CNT PIX {| δ (x} , y) |> ε} as expressed by (5), I _ti and I _tj of the difference [delta] ( The number of pixels whose absolute value of (x, y) exceeds the standard deviation ε may be counted as the distance. In the example of this distance function, a case has been described to use the standard deviation of the difference ε between the two images I _ti and I _ti should seek image distance, prior m frames from these picture I _tim and I _TJM Standard deviation ε _m of the difference δ (x, y) between
Similarly, the difference between the image I _tim and the image I _tjm as expressed by the following equation d (ti, tj) = CNT _PIX ｛| δ (x, y) |> ε _m ｝ (5 ′) the value which the absolute value has counted the number of pixels exceeding the epsilon _m of may be the inter-image distance. m is a predetermined constant integer of 1 or more, and m =
When it is set to 0, it represents the same distance function as the equation (5).

When the inter-image distance is obtained using the above-described standard deviation of the inter-image difference as the fifth distance function, the distance function may be determined as follows. That is, the distance to the set I _ti and I _tj of the image to be obtained and the standard deviation epsilon _k k = 1 of the luminance difference between the set I _tik and I _tjk thereof than k frames before the image [delta] (x, y) To k = n, and the average value T = Ｔ _k ε _k / n of their standard deviations ε _k
The use as with epsilon _m of the above equation (5 ') as represented by the following formula d (ti, tj) = CNT PIX {δ (x, y)> T} (6), an image I _ti The number of pixels whose absolute value of the difference δ (x, y) between _Itj exceeds T is counted and defined as the distance. According to this method, there is an effect that an image change caused by a camera shake or a sudden movement of a camera at the time of shooting is cut to prevent erroneous detection.

Further, as a sixth distance function, the image _Iti is divided into blocks of a predetermined size as shown in FIG.
Each block is checked for a match with any part of the image _Itj, blocks that do not match any part are counted, and the count value is defined as a distance d (ti, tj). That is, a picture I _ti blocks which means that the distance by counting the number of blocks to detect whether the moved to any position on the image I _tj, could not be moved detected. Any The function defining the matching between the region, for example, one vector the sequence of the luminance values of p × q pixels of the block on the image I _tj each block (e.g. p × q pixels) in an image I _ti and then, the luminance value sequence of a region to be compared with the same size of the image I _tj and another vector, these two distances between vectors equation _{D = Σ p, q {b} j (p, q) -b _i (p, q)｝ ² , if the distance D is less than or equal to a predetermined threshold D _th , it is determined that the block and the compared area match, and if D exceeds D _th , no match is made Is determined. image
The region to be compared on _Itj is sequentially shifted at a predetermined pixel pitch every time the matching distance is checked, over the entire surface of the image _Itj . Similarly for other block of the image I _ti whether there is a portion that matches the image I _tj be determined respectively, the number of blocks not take the matching with the distance between the two images. Therefore, the distance between images can be expressed as the following equation: d (ti, tj) = CNT _BLK {D> D _th } (7)

Next, the distance table 13 will be described. The distance table 13 calculates the inter-image distance d (ti, tj) obtained by the inter-image distance calculation unit 12 as J (i = 0,..., J-1) × J (j = 1,
.., J), and the size of this table is called J. The distance table 13 is developed, for example, in an area secured on a memory in the computer, and a data area that is no longer necessary for effectively using the memory is overwritten with new distance data for use.

For the image data sequence shown in FIG.
Expression (1) for all pairs of integers i and j in the range of <j ≦ 7
FIG. 5 shows the value of the distance d (ti, tj) between the images calculated in the above as a table obtained at time t. Since d (ti, tj) generally has symmetry and d (ti, tj) = d (tj, ti) holds, it is not necessary to calculate the area of i> j in the distance table of FIG. Further, as will be described later, for example, in the distance table of FIG. 5, only d (ti, tj) satisfying 0 <ji ≦ 3 may be obtained.

FIG. 6 is a diagram showing characteristics of the distance table of FIG. When the time point, that is, the frame number (tj _C ) is the cut point (j _C = 4 in the example of FIG. 5), the distance table is
It is characterized in that the value of the distance between images in the area A1 has a larger value than the value of the distance between images in the areas A2 and A3. Region A1 is a region that satisfies i <j _C ≦ j, the inter-image distance in the lower left corner of the area corresponds to _{d (tj C -1, tj C} ). Region A2 is a region that satisfies i <j ≦ j _C -1, region A3 is a region that satisfies j _C ≦ i <j.

FIG. 7 shows another example of the image data sequence, in which flash light is shown in the image It _-4 corresponding to the frame (t-4). Expression for this image data sequence
The distance table obtained at the time t according to (1) is as shown in FIG. FIG. 9 is a view for explaining the characteristics of this distance table. When the flash light is reflected in the frame (tj _F ), the value of the distance between the images in the areas A1, A2, and A3 is changed to the area A.
Compared to values on the i = j _F and j = j _F indicated by 4 takes a small value. FIG. 10 shows still another example of the image data sequence, in which a cut is included in the image sequence subjected to the telecine conversion processing. In FIG. 10, the image of the frame (t-4) is the image I.
It is synthesized by superimposing _t-3 and It _-5 ,
The even-numbered fields are black and the odd-numbered fields are white. FIG. 11 shows a distance table at time t according to equation (1) for the image data sequence of FIG. 10, and FIG. 12 is a diagram showing characteristics of this distance table. A telecine blur image such as the image It _-4 is a frame (tj _T ) (j _{T in} FIG. 10).
= 4), as shown in FIG. 12, the areas A2 and A3
Values of image distance takes a smaller value than the value of the area A1, the value of the inter-image distance on i = j _T and j = j _T indicated by a region A4 is in the value and the region A1 in the region A2, A3 Take a value in the middle of the values.

FIG. 13 shows an example of an image data string including a fade. 13, the image _{I t-7, I t-} 6 is an image of the entire screen black image _{I t-2, I t-} 1, I t is the entire screen is white image, that during the image The screen gradually becomes brighter at It _-5 , It _-4 , and It _-3 . FIG. 14 shows an equation for this image data sequence.
FIG. 15 shows a distance table at time t according to (1), and FIG. If a slow scene change occurs in the image with time t, the area A shown in FIG.
Tends to decrease as the distance from the upper right corner increases.

It should be noted that the distance data d (ti, tj) at each position (i, j) in the distance table obtained at time t is d (t-1). -
As is clear from the fact that it can be rewritten as (i + 1, t-1-j + 1), the distance data of the position (i-1, j-1) in the distance table created at the last time point (t-1) Is the same as That is, every time the time t is updated, the inter-image distance data d (ti, tj) of the position (i, j) in the distance table is
(i + 1, j + 1) and the latest distance data d (t,
t-1), d (t, t-2), ..., d (t, tJ) are table positions (0,1), (0,
2),…, (0, J). Data in which the shift destination position (i + 1, j + 1) due to this data shift is out of the table is discarded. Therefore, at each time point t, the distance between images that needs to be newly calculated for creating a distance table is d (t, t-1), d (t, t-2),..., D (t, tJ). Only. [Scene Change Rate] Next, the scene change rate calculation unit 14 will be described. The scene change rate C (tj _C ) is calculated by a predetermined function using the data of the inter-image distance table obtained from the image up to the J frame before obtained at time t.
The scene change rate C (tj _C ) calculated for the time t is an evaluation amount for determining whether or not the image j _C frames before the current time is a cut, and is compared with a predetermined threshold value C _th to determine If it is larger, it is determined to be a cut. The scene change rate calculation unit 14 calculates the scene change rate using the fact that the distance table 13 has the above-described features for each of the cut point, flash light, telecine bokeh, and slow scene change, and the determination unit 15 calculates scene change rate C (tj _C) compared with the threshold C _th frame (tj _C) determines whether it is cut. 5 shows an example of a function that defines a scene change rate in the scene change rate calculation unit 14. j _c is 0
It is a predetermined constant integer satisfying <j _c ≦ J.

As can be seen from the example of the distance table based on the image sequence including the instantaneous cuts shown in FIG. 5 and the distance table based on the telecine transformed image sequence including the cuts illustrated in FIG.
When there is a scene change point (cut point), the inter-image distance value in the area A1 of the table is large, and the inter-image distance value in the areas A2 and A3 is small. In the distance table in the case where the fade of FIG. 14 is in the image sequence, FIG.
When the regions A1, A2, and A3 similar to FIG. 13 are applied, the same tendency is found. On the other hand, in the distance table based on the image sequence including the flash shown in FIG.
The distance values in 1, A2 and A3 are all substantially the same. Therefore, focusing on these areas A1, A2, A3,
A distance table of size J as shown in FIG. 16 from the image up to the J-th frame before the current time t is set to 0 ≦ i <j ≦ J and ji ≦
g within the range. g is a predetermined constant integer satisfying 0 <g ≦ J, and FIG. 16 shows a case where g = j _C -1. The value of j _C is, for example, the smallest integer exceeding J / 2. In this distance table, the function of the scene change rate is defined as follows.

C (tj _C ) = {a (j _C ) −βMAX [b (j _C ), b ′ (j _C )]} / S (8) a (j _C ) = Σ _{i, j} d (ti , tj) for 0 ≦ i <j _C ≦ j ≦ J, ji ≦
g b (j _C ) = Σ _{i, j} d (ti, tj) for 0 ≦ i <j ≦ j _C −1, ji ≦
g b ′ (j _C ) = Σ _{i, j} d (ti, tj) for j _C ≦ i <j ≦ J, ji ≦ g a (j _C ) is the distance between all the images indicated by crosses in the area A1. The sum of the distances, b (j _C ) is the sum of all the inter-image distances indicated by △ in the area A2, and b ′ (j _C ) is the sum of all the inter-image distances indicated by the △ in the area A3 Is the sum of _{MAX [b (j C),} b '(j C)] is b
This means that the larger value of (j _C ) and b '(j _C ) is taken. S is the number of distance data in the area A1 (the number of crosses)
In the example of FIG. 16, S = (j _C -1) j _C / 2.
If J is an odd number, the number of data in the area A2 is equal to the number of data in the area A3. β is the number of data S in the area A1 and the area A
This is a coefficient for equalizing the number of data in 2 or A3. According to the settings of the areas A1, A2, and A3 in FIG.
Is odd, β = 1 because the number of data in A1, the number of data in A2, and the number of data in A3 are all equal.

When g = J, the upper right corner of the distance table
The whole area including (0, J) will be used. In that case, S
= J is _{C (Jj C +1), β} = 2 (Jj C +1) / (j C -1) is about 2. Since the area A1 of the distance table shown in FIGS. 2, 10, and 13 shows a large distance value over almost the entire surface, the position (0, J) of the upper right corner is included by selecting g such that g <J. Even if a corner region having a certain size is removed, the feature that the distance value in the region A1 is larger than the distance values in the regions A2 and A3 is maintained. Therefore, the cut point can be detected by Expression (8). As described above, the distribution of the distance value shifts one step toward the lower right every time the time t is updated. Therefore, when g <J, it is necessary to create a distance table every time the time t is updated. The required distance between images that must be newly calculated is g of d (t, t-1), d (t, t-2), ..., d (t, tg)
It is possible to reduce the calculation amount without reducing the cut detection ability. If the value of g is selected to be small, the necessary amount of calculation is reduced accordingly. However, if the value of g is too small, the error of the cut detection increases. Therefore, the range of j _C −1 ≦ g <J is preferably set.

If the scene change rate C (tj _C ) obtained by substituting the data of the distance table shown in FIG. 16 obtained at the time point t into the equation (8) is larger than a predetermined threshold value C _th, j _It is determined that a scene change (cut) has occurred in the image of the frame (tj _C ) before _C. For example illustrating an example of a J = 5, g = j area on _C = 3 in the distance when the table A1, A2, A3 in FIG. 17. The number S of data in the area A1 is 6,
Since the number of data of both A2 and A3 is 3, β = 2. C (tj _C ), a (j _C ), b (j _C ) and b ′ (j _C ) are shown in FIG.
Is given by

C (tj _C ) = {[a (j _C ) −2MAX [b (j _C ), b ′ (j _C )]} / 6, (9) a (j _C ) = d (t−2) , t-3) + d (t-1, t-3) + d (t, t-3) + d (t-2, t-4) + d (t-1, t-4) + d (t-2, t _{-5), b (j C)} = d (t, t-1) + d (t, t-2) + d (t-1, t-2), b '(j C) = d (t-3, t-4) + d (t-3, t-5) + d (t-4, t-5) where a (j _C ) is d (ti,
tj), and b (j _C ) is d (ti, tj) belonging to the area A2.
And b ′ (j _C ) is the sum of d (ti, tj) belonging to the area A3. Therefore, when equation (1) is used as the distance function between images, the scene change rate C (tj _C ) is always −1 ≦ C (tj _C ) ≦ 1
Value in the range.

As described above, when the frame (tj _C ) is the cut point, a (j _C ) becomes larger than the values of b (j _C ) and b '(j _C ), so that C (tj _C ) Has a positive value. Stated another way, a (j _C ) is the distance between the images of all sets of two frames with a frame interval of 3 or less, each one selected from three before and three after six consecutive images. If the scene has changed significantly at time (tj _C ), a
(j _C ) should have a large value. b (j _C ) is the frame
(tj _C ) and the total of the three images before that are the sums of the distances between the two images in all sets, and are the frames (t
-j _C ) The smaller the motion of the previous scene, the smaller the value. In b '(j _C) three images after the frame (tj _C), an amount obtained by adding the distance between two images of all the sets, the movement of the scene appearing after the frame (tj _C) The smaller the value, the smaller the value.

For example, based on the distance table shown in FIG.
In order to apply specific numerical values to (9), the area pattern map shown in FIG.
(i, j) = (4,5) is superimposed so as to coincide with the position (i, j) = (6,7) in the table shown in FIG. ) At the scene change rate C (tj _C ) (however, FIG.
J _C = 3) a calculated as follows in.

A (3) = 0 + 0 + 0 + 1 + 1 + 1 = 3, b (3) = 0 + 1 + 1 = 2, b '(3) = 0 + 0 + 0 = 0, C (t-1-3) = (3-2 × 2) /6=-1/6=-0.16 At the next time point t, the area pattern map in FIG. 17 is shifted one step to the upper left. As a result, positions (1, 2),
Data 0, 0, 1 in (1, 3), (1, 4) are written as the latest distance data at positions (0, 1), (0, 2), (0, 3) in FIG. The scene change rate in this state is calculated as follows.

A (3) = 1 + 1 + 1 + 1 + 1 + 1 = 6, b (3) = 0 + 0 + 0 = 0, b '(3) = 0 + 0 + 0 = 0, C (t−3) = (6−2 × 0) /6=1.0 In these two calculation examples, the value of the scene change rate C (t−4) calculated from the distance table at the time (t−1) is negative. Yes, no cut. On the other hand, the scene change rate C (t−3) calculated from the distance table at time t is 1
For example, if the threshold value C _th is set to 1/16, it means that a cut is detected in the frame (t−3) three times before at the time point t. Note that the threshold value C _th = 1/16 is a value determined by an experiment so as to minimize the determination error.

As can be seen from this example, C (tj _C ) takes a value close to 1 at the cut point. Further, even if, for example, camera shake or panning (operation for moving the camera horizontally) is present in the video, it is not uncommon for the camera to erroneously detect it as a cut point. Because even if there is shake or bread a
(j _C ) has a large value, but at the same time, b (j _C ) and b '(j _C ) have a large value.
This is because (tj _C ) becomes smaller.

In order to apply Equation (9) to FIG. 17 to detect a cut in an actual series of input images, the area A1 in FIG. , A2, and A3 as long as the distance d (ti, tj) between the images is obtained, and each time the time t is updated, FIG.
Only the distance between the images at the middle positions (0,1), (0,2), (0,3) is newly calculated, and all the data held so far need only be shifted to the lower right by one step.

Next, by applying the area pattern map of FIG. 17 to the data of FIG. 8 in the case where the flash light shown in FIG. Shows the results calculated for the time points (t-1), t, and (t + 1). C (t-1-3) = (3-2 × 2) /6=-1/6=-0.16 C (t-3) = (3-2 × 2) /6=-1/6=-0.16 C (t + 1-3) = (2-2 × 2) /6=−2/6=−0.3 As can be seen from the calculation results, in both cases at the time (t−1) and t, FIG. since bract 3 of the four "1" on the three "1" and j = j _F on i = j _F in the area A4 indicated cancel, the scene change rate C (t-4), C (t-3) is smaller than the threshold value 1/16, so that the flash light is not erroneously detected as a cut point. In the conventional method, d (t−3,
Since t-4) has a large value, the frame (t-4) is often erroneously determined to be a cut point.

When the telecine blur image is included in the image sequence as shown in FIG. 10, the distance value in the region A4 shown in FIG. 12 is an intermediate value, but the distance value in the region A1 is intermediate. Since the value is large, C (tj _C ) takes a value close to 1. As described above, the area pattern map of FIG. 17 is superimposed on the distance table of FIG.
When C (tj _C ) is calculated according to (9), C (t-1-3) = (4.5-2 × 1) /6=5/12=0.4 C (t-3) = (4.5-2 × 1) /6=5/12=0.4 C (t + 1-3) = (2-2 × 2) /6=-1/3=-0.3 Frame because C (t-3)> 1/16 and C (t-4)> 16
It is correctly determined that (t-3) and (t-4) are cut points. That is, this example indicates that the cutting is continued over a plurality of frames. D (t-3, t-4), d in FIG.
When the distance between images is small as in (t-4, t-5), the conventional method often cannot detect frames (t-4) and (t-5) as cut points.

As shown in FIG. 13, when the image sequence includes a plurality of slowly changing images such as fades, the value becomes larger toward the upper right of the area A shown in FIG. C (tj _C ) calculated by the equation (9) by superimposing the area pattern map of FIG. 17 on the table at the time (t-1), t, (t + 1) is as follows. C (t-1-3) = (3.5-2 × 0.8) /6=0.31 C (t-3) = (3.0-2 × 1.2) /6=0.1 C (t + 1-3) = (2.1- 2 × 1.0) /6=0.016, and it can be correctly detected that the scene has changed.

The equation for defining the scene change rate C (tj _C ) calculated by the scene change rate calculation section 14 is as follows: C (tj _C ) = ｛a (t) −β [b (t) + b ′ ( t)]｝ / S, (10) a (j _C ) = Σ _{i, j} d (ti, tj) for 0 ≦ i <j _C ≦ j ≦ J, ji ≦
g b (j _C ) = Σ _{i, j} d (ti, tj) for 0 ≦ i <j ≦ j _C -1, ji ≦
g b ′ (j _C ) = Σ _{i, j} d (ti, tj) for j _C ≦ i <j ≦ J, ji ≦ g

FIG. 18 is a processing flowchart of an embodiment of a method for detecting a video cut point in an input image sequence in real time according to the present invention. Here, real-time processing means that video cut point detection processing can be performed while displaying, recording, and transferring video at a standard speed. The time t every time one image is input and the frame number of a series of images are the same, and the distance table 13
Are J = 5 and g = j _C = 3 as shown in FIG.

In step S1, time t is initialized to zero. That is, the processing start time is regarded as 0. In step S2, a process of reading the image at time t into the buffer memory 11 is performed. This process includes, for example, a process of converting a video signal into digital data that can be processed by A / D conversion as preprocessing. In step S3, the data d (ti, tj) of each position (i, j) in the data areas A1, A2, A3 in the distance table is shifted to the position (i + 1, j + 1), and the image distance In the calculation unit 12, for j = 1, 2, 3
The distance d (t, tj) between the images is obtained by (1), and the distance table 1
3 are stored at the positions (0,1), (0,2), (0,3). In this case, not only the distance d (t, t-1) between temporally adjacent images,
This method differs from the conventional method in that a distance between images separated in time such as d (t, t-2) and d (t, t-3) is calculated and used. In step S4, the scene change rate calculation unit 14 uses the data in the distance table 13 to calculate the scene change rate C (t-3) using, for example, equation (9). In step S5, C (t-3)
Is larger or smaller than the threshold value 1/16, and if C (t-3)> 1/16, the frame (t-3) is determined to be the cut point (step S6), and C is determined.
If (t−3) ≦ 1/16, it is determined that the frame (t−3) is not a cut point (step S7). Thereafter, the time t is increased by 1 in step S8, and the process returns to step S2.

[0050] Note that when using equation (9) described above obtaining the scene changing ratio C (tj _C), as can be seen from the processing flow of FIG. 18, the first time at time t of (tj _C) since it is cut point determination, but temporally delayed only cut detection j _C image occurs, applications there is no problem. In addition, actually, the calculation of the scene change rate in step S4 and the subsequent processes in steps S5 to S7 are not executed until the areas A1, A2, and A3 of the table shown in FIG. 17 are filled with data.
Steps S2, S3 and S8 are repeated a predetermined number of times, and after these areas of the table are filled with data, step S4
To S7 can be performed effectively. In the case of FIG. 17, when an image of 3 × 5 frames is further input after the first image frame is input, three areas of the distance table 13 are filled with the distance data. Normally, after the start of the video, several tens
Since there is no cut in the frame, there is no problem even if the cut is not detected for several tens of frames from the first frame.

An example of another function for defining the scene change rate will be described below. As described with reference to FIGS. 5, 8, 11, and 14, the inter-image distance data shown in the distance table has a characteristic distribution depending on the type of scene change existing in the image sequence. Accordingly, a table of distance data distribution representative of one or more types of scene changes (cuts) to be detected is prepared in the scene change rate calculation unit 14 in advance. This is called a template. FIG. 19A,
19B and 19C are examples of an instant cut detection template T _CT (i, j), a fade detection template T _FD (i, j), and a flash detection template T _FL (i, j), respectively. This is the case where J = 5. The scene change rate C (tj _C ) shown here is defined by the similarity between the distance table as shown in FIG. 20 and the template of the same size J created each time one frame of image data is input. . The degree of similarity with the template is represented, for example, by a cross-correlation coefficient R (d, T). In that case, it can be expressed by the following equation.

C (tj _C ) = R (d, T) = Σ _{i, j} ｛d (ti, tj) −d _AV ｝ ｛T (i, j) −T _AV ｝ / (σ _d σ _T ) ^{1 / 2} (11) d _AV = 1 / SΣ _{i, j} d (ti, tj): Average value of d (ti, tj) T _AV = 1 / SΣ _{i, j} T (i, j): T (i, Average value of j) σ _d = 1 / SΣ _{i, j} ｛d (ti, tj) −d _AV ｝ ² : d (ti, tj)
Σ _T = 1 / Sj _{i, j} ｛T (i, j) −T _AV ｝ ² : variance of T (i, j) where i, j are all satisfies 0 ≦ i <j ≦ J Is a set, S is the number of the set, and Σ _{i, j} represents the sum of all the sets. Also in this case, the area of the template and the distance table may be limited to ji ≦ g, 0 <g <J.

The scene change rate calculation unit 14 calculates the distance table shown in FIG. 20 for each input image frame and the distance table shown in FIG.
The similarity R (d, T) with one or more desired templates shown in 19B, 19C, etc. is calculated as a scene change rate by equation (11), and when the value is larger than a predetermined threshold value _Rth . If so, it is determined that the corresponding scene change is in the frame (tj _C ).

FIG. 21 shows a processing flow for detecting what kind of scene change (cut) exists in which frame in the input image sequence using the above-mentioned plural templates. The basic procedure is as shown in FIG. The same is true. However, in this example, the flash is not a cut, but is added as one of the detection targets in the image sequence. In step S1, time t is initialized to 0, and in step S2, image data at the current time t is fetched. In step S3, the distance data d (ti, tj) at all positions (i, j) in the distance table is shifted to the position (i + 1, j + 1).
The image interval distance d (t, tj), j = 1, 2,..., J is calculated using the J + 1 image frames preceding the J frame held in the buffer memory 11 (FIG. Calculated by
The distance table at the present time t is obtained by writing to the position (0, j) in the table. In step S4, FIG.
The similarity between the 9A instantaneous cut detection template and the current distance table is calculated by equation (11). Step S
In step S5, the similarity is compared with a predetermined threshold value _{Rth. If the} similarity is equal to or less than the threshold value, it is determined that there is no instantaneous cut, and the process proceeds to step S6. The similarity between the fade detection template in FIG. Calculate according to 11). In step S7, the degree of similarity with the template for fade detection is compared with the threshold value _{Rth. If} the similarity is equal to or smaller than the threshold value, it is determined that there is no fade, and the process proceeds to step S8. Degree expression
It is calculated by (11). The similarity between the flash detection template at step S9 is compared with a threshold value R _th, determines that the flash was not as long below the threshold, step S1
The time t is updated with 0, and the process returns to step S2. Step S
5, S7, if the degree of similarity in any of S9, was greater than the threshold value R _th, corresponding scene change frame (t
determines that were in -j _C), the flow returns to step S2 to update the time in step S10.

Although the principle of the cut detection method of the present invention has been described above with reference to a specific example, the cut detection device that implements the cut detection method of the present invention is not limited to the configuration shown in FIG. It is. Figure 22 shows another embodiment of the cut detection apparatus, J + 1 frames buffer 11 ₀ to 11 _J are provided corresponding buffer memory 11 in FIG. 1, also the inter-image distance calculation section 12 of FIG. 1 the J inter-image distance calculation section 12 ₁ to 12 _J are provided corresponding to. The new image frame I _t before storing in the buffer, the image frame stored in the frame buffer 11 ₀ ~11 _J-1 transferred respectively to the buffer 11 ₁ to 11 _J (hence far into the buffer 11 _J until then image data that has been held is discarded), the new image frame I _t is written into the buffer 11 _0. When storage of image frames for all buffers is completed, the inter-image distance calculation section 12 ₁ to 12 _J an image frame I read from the frame buffer 11 ₁ to 11 _J, respectively
_t-1, I _t-2, ..., with I _tJ is given, the most recent image frame I _t at the current time t from the frame buffer 11 ₀ is supplied with a common, each inter-image distance d (t, t-1) , d (t, t-2),…, d
(t, tJ) is calculated using, for example, equation (1). These inter-image distances are given to the distance table memory unit 13 and are shown in FIG.
Position (0,1) on the distance table when J = 5 shown in Fig. 3
Write to the address area corresponding to (0,2), ..., (0, J). However, before the writing, the data d (ti, tj) at the position (i, j) on the existing table is shifted to the position (i + 1, j + 1) as shown by the arrow in FIG. I do. Data that will be located outside the table due to this shift is discarded. Using the distance table obtained at the time point t in this way, the scene change rate calculation unit 14 calculates the scene change rate C (tj _C ) based on the above-described desired scene change rate calculation method.
Is calculated. The determination unit 15 compares the given scene change rate with a threshold, and determines whether or not there has been a scene change based on the comparison result.

According to the embodiment of FIG. 22, it is necessary to provide a plurality of frame buffers and inter-image distance calculation units, but it is possible to process image data at high speed in a pipeline manner. FIG. 24 shows another embodiment of the cut detection device. This embodiment has a computer configuration in order to more practically realize the operation principle configuration shown in FIG. 1. The image sequence storage unit 11, the inter-image distance calculation unit 12, the distance table memory unit 13, The operations of the scene change rate calculator 14 and the like are controlled by a calculation controller (CPU) 20 connected via a bus 19. Image data input unit 10
The original control of the input image data I _t the calculation control unit 20 1
The frame is fetched and transferred to the image sequence storage unit 11. The image sequence storage unit 11 is, for example, a main storage unit of a computer,
Buffer area for pre-allocated J + 1 frames 11 ₀
~ 11 _J, and holds the image of the latest J + 1 frame. As in the case of FIG.
When a new image frame is input to the input unit 10, first data I _t in the region 11 of the image sequence storage unit 11 ₀ ~11 _J-1 (not _{shown), ..., I t-J} + 1 each writing in the area 11 ₁ to 11 _J, and writes the latest image data from the input unit 10 to the region 11 _0. When writing to the image sequence storage unit 11 is completed, the calculation control unit 20 is an image frame from the storage unit 11 I _t and between given by reading a set of I _tj in the inter-image distance calculation section 12 image distance d (t, tj) and transferring the calculation result to the position (0, j) in the table (see FIG. 23) of the distance table memory unit 13 are sequentially repeated for j = 1, 2,. As a result, J new distance data are obtained in the distance table of the distance table memory unit 13.

[0057]

As described above, according to the present invention, not only two images which are temporally adjacent to each other as a screen change amount are compared, but also a change amount between images which are temporally separated is taken into consideration. Cut point detection method and apparatus that can detect a scene change that changes slowly in time by calculating the scene change rate, and can uniformly process telecine-converted images, images including flash light, etc. in real time There is an effect that can be provided.

[Brief description of the drawings]

FIG. 1 is a block diagram showing an operation principle configuration of a cut detection device of the present invention.

FIG. 2 is a diagram showing an example of an image sequence including a cut.

3A is a diagram illustrating an example of a difference histogram of two images without a cut, and FIG. 3B is a diagram illustrating an example of a difference histogram of two images with a cut.

FIG. 4 is a diagram for explaining block matching between two images.

FIG. 5 is a diagram showing an example of a distance table created from the image sequence shown in FIG. 2;

FIG. 6 is a view for explaining features of the distance table of FIG. 5;

FIG. 7 is a diagram showing an image sequence including a flash image.

FIG. 8 is a diagram showing an example of a distance table created from the image sequence shown in FIG. 7;

FIG. 9 is a view for explaining features of the distance table of FIG. 8;

FIG. 10 is a diagram showing an example of a telecine-converted image sequence including a cut.

FIG. 11 is a view showing an example of a distance table created from the image sequence shown in FIG. 10;

FIG. 12 is a view for explaining features of the distance table of FIG. 11;

FIG. 13 is a diagram showing an example of an image sequence including a fade.

FIG. 14 is a view showing an example of a distance table created from the image sequence shown in FIG. 13;

FIG. 15 is a view for explaining features of the distance table of FIG. 14;

FIG. 16 is a view for explaining an example of a method for obtaining a scene change rate from a distance table.

FIG. 17 is a view for explaining a method of obtaining a scene change rate using a distance table when J = 5.

FIG. 18 is a flowchart showing a processing procedure of a cut detection method according to the present invention.

19A is a diagram illustrating an instantaneous cut detection template, FIG. 19B is a diagram illustrating a fade detection template, and FIG. 19C is a diagram illustrating a flash detection template.

FIG. 20 is a diagram showing an example of a distance table to be compared with a template.

FIG. 21 is a flowchart showing a processing procedure of a method for detecting a desired scene change using a template.

FIG. 22 is a block diagram showing another configuration example of the cut detection device of the present invention.

FIG. 23 is a view for explaining creation of a distance table.

FIG. 24 is a block diagram showing still another configuration example of the cut detection device of the present invention.

Continuation of front page (56) References 1. Otsuji, Tonomura "Study of Automatic Video Cut Detection Method" Technical Report of the Institute of Television Engineers of Japan Vol. 43, pp. 7-12 [July 1992] 2. K. Otsuji and Y. Tonomura "Projection Defect Filter for Video Cut Detection" Proc. of ACM Multimedia 93, 1993, pp. 251-257 (1993) (58) Fields investigated (Int. Cl. ⁶ , DB name) H04N 5/262-5/28

Claims (22)

(57) [Claims] 1. A video cut point detection method for detecting a video cut point from a video data sequence includes the following steps: (a) a sequence of sequentially input image frames from a current frame t to a frame J before the current frame; A series of J + 1 images
_{I t, I t-1,} ..., with respect to I _tJ, calculated frame interval is J or less all the sets of two images I _ti, the distance d (ti, tj) between I _tj to each position (i, j), create a distance table having a distance d (ti, tj), where J is a constant integer of 3 or greater, and i, j are all values in a range satisfying 0 ≦ i <j ≦ J. is variable integer, (b) a scene change rate of j _C frame prior to the distance based on the inter-image distance on the table J-frame preceding image I from _tJ current image I _t to the current time t changes in image C (t
-j _C ), where j _C is a predetermined constant integer in a range satisfying 0 <j _C ≦ J, and (c) compare the scene change rate C (tj _C ) with a predetermined threshold. and, j _C-frame preceding image from the current time by the comparison result determines whether the cut point. 2. The method of claim 1 wherein said step (a)
Is obtained by adding across each corresponding position of the image I _tj before image I _ti and j frames before i frame from the present time t (x, y) the absolute value of the difference of the luminances of pixels in the entire screen A step of obtaining the sum of the differences as a value corresponding to the inter-image distance d (ti, tj). 3. The method of claim 1 wherein said step (a)
Obtains a luminance value histogram of the predetermined number of classes for the i-th frame preceding image I _ti and j frame preceding image I _tj from the present time t, all classes the absolute value of the difference between the frequency of the corresponding class of their histograms And calculating the sum of the differences obtained by adding the values as a value corresponding to the inter-image distance d (ti, tj). 4. The method of claim 1 wherein said step (a)
Obtains across the corresponding luminance differences of the pixel at each position (x, y) of the i-th frame preceding image I _ti and j frame preceding image I _tj from the current time t to the full screen, the standard deviation of their differences And a step of obtaining a value corresponding to the inter-image distance d (ti, tj). 5. The method of claim 1, wherein said step (a)
Is the image i frames before the current time t for which the distance is to be obtained.
Each corresponding position of the I _ti and j frame preceding image I _tj (x, y)
Is calculated over the entire screen, the standard deviation of the differences is calculated, pixels whose difference exceeds the standard deviation are counted over the entire screen, and the counted value is calculated as the image distance d ( ti, tj). 6. The method of claim 1, said step (a), each corresponding position of the set of image I _ti and I _tj should seek i and j frames previous distance from the present time t (x, y) At the corresponding positions (x, y) of the images I _tim and _Itjm m frames before, respectively, from the above set of images. The difference is obtained as a second difference over the entire screen, the standard deviation of the second difference is obtained, and the pixels whose absolute value of the first difference exceeds the standard deviation are counted over the entire screen. Calculating a count value as a value corresponding to the inter-image distance d (ti, tj),
The above m is a predetermined integer of 1 or more. 7. The method of claim 1, said step (a) each corresponding position of the set of image I _ti and I _tj should seek i and j frames previous distance from the present time t (x,
obtains a luminance difference of pixels in y) over a previous screen as a first difference, each corresponding position of the sets from each of the image k frames before the image I _tik and I _tjk (x,
The difference in luminance of the image in y) is obtained as a second difference over the entire screen, the standard deviation ε _k of the second difference is obtained, and the second difference and its standard deviation are obtained by k =
The average of n standard deviations obtained as a result of the repetition for each set of images of 1, 2,..., N is determined as the average standard deviation, and the pixel whose absolute value of the first difference exceeds the average standard deviation is calculated. Is calculated over the entire screen, and the count value is determined as a value corresponding to the inter-image distance d (ti, tj), where n is a predetermined integer of 2 or more. 8. The method of claim 1, wherein said step (a)
Is to divide the image _Iti i-frames before the current time t into blocks of predetermined p × q pixels, and to determine whether a portion matching each of the above blocks is in the image _Itj before the j-th frame. And repeatedly counting the number of unmatched blocks for the above-mentioned block, and obtaining the counted value as a value corresponding to the inter-image distance d (ti, tj). 9. The method of claim 1 wherein said step (a)
Shifts the data of the distance of the position (i, j) in the table at the time t-1 to the position (i + 1, j + 1), and at the current time t, the distance is d (t, t-1). ), d (t, t-2),…, d (t, tJ)
The calculated values are stored at positions (0,1), (0,2), ...,
Creating the table at time t by writing to (0, J). 10. The method of claim 1, wherein the steps are
In (a), the range for obtaining the distance between the images is ji ≦
Limited to the range satisfying g, where g is a predetermined constant positive integer smaller than J. 11. The method of claim 10, wherein said steps are:
(a) shifts the distance data at the position (i, j) in the table at the time t-1 to the position (i + 1, j + 1), and at the current time t, the distance is d (t, j). t-1), d (t, t-2), ..., d (t, tg), and calculate the calculated values at the positions (0,1), (0,2),
, (0, g) to create the table at time t. 12. The method of claim 11, wherein the steps
(b) shows that j _C ≦ j ≦ J, 0 ≦ i <j _C , j−
The first area is defined by i ≦ g, the sum of the distances in the first area is represented by a (j _C ), and the second area is defined by 0 ≦ i <j ≦ j _C −1 and ji ≦ g. And the sum of the distances in the second area
b (j _C ), the third area is defined by j _C ≦ i <j ≦ J, ji ≦ g, and the sum of the distances in the third area is expressed as b ′ (j _C ). C (tj _C) as a scene change rate _{= {a (j C) -βMAX} [b (j C), b '(j C)]} / S comprises the step of calculating, beta is the first region and the This is a predetermined coefficient for equalizing the number of data in the second or third area, and S represents the number of data in the first area. 13. The method of claim 11, wherein the steps are
(b) shows that j _C ≦ j ≦ J, 0 ≦ i <j _C , ji ≦
g defines the first area, the sum of the distances in the first area is represented by a (j _C ), and 0 ≦ i <j ≦ j _C −1 and ji ≦ g define the second area. , The sum of the distances in the second area is b
(j _C ), the third area is defined by j _C ≦ i <j ≦ J, ji ≦ g, and the sum of the distances in the third area is expressed as b ′ (j _C ). include _{C (tj C) = {a} (j C) -β [b (j C) + b '(j C)]} / S to calculate the process as the rate of change, beta is in the first region, said first 2
And a predetermined coefficient for equalizing the number of data in the third area, and S represents the number of data in the first area. 14. The method of claim 9 wherein said steps are
(b) is the same size as the distance table and at least one
Calculating the similarity between the predetermined desired cut detection template and the distance table obtained at time t as the scene change rate. 15. The method according to claim 14, wherein the similarity is a cross-correlation coefficient between the distance table and the template. 16. A video cut point detecting apparatus for detecting a video cut point from an image data sequence, comprising: a buffer memory means for sequentially buffering image data of at least J + 1 frames continuous in time; the image data I _t, I _t-1 from t before J frame, ..., the I _tJ, 0
An image distance calculating means for calculating a distance d (ti, tj) between image data of two images _{Iti, Itj} in a range of ≤ i <j ≤ J; and an image calculated by the image distance calculating means. Distance table means for storing an inter-distance; scene change rate calculating means for calculating a change in an image from frame (tJ) to t as a scene change rate in frame (tj _C ) with reference to the distance table means; j _C is 0 <
j _C ≦ J is a predetermined constant integer, and the calculated scene change rate is compared with a predetermined threshold to determine whether the frame (tj _C ) is a cut point. Determination means. The apparatus of claim 17 according to claim 16, the image data I _t of at least J + 1 frames of said buffer memory means from frame t to (tJ) before, I _t-1, ..., each time one frame one by one the I _tJ First to J + 1
It includes a frame buffer, the inter-image distance calculating means each image data I _t-1 from the second to J + 1 frame buffers, ..., given I _tJ, the image data I _t from said first frame buffer is common respectively given, each image data I _t and the image data I _{t -1,} ..., the distance between I _tJ d (t,
t-1),..., d (t, tJ), respectively, and the distance table means stores the distance data of the position (i, j) on the table. (i + 1, j + 1) and writing the distance data from the inter-image distance calculation unit at the positions (0, 1),..., (0, J) on the table, respectively. Is a means for creating 18. The apparatus according to claim 16, wherein said inter-image distance calculating means is means for calculating said inter-image distance in a range satisfying a further limited ji ≦ g.
It is a smaller predetermined constant positive integer. 19. The apparatus according to claim 18, wherein said distance table means comprises a position in said table at time t-1.
The distance data of (i, j) is shifted to the position (i + 1, j + 1), and the image distances d (t, t-1) and d ( By writing (t, t-2), ..., d (t, tg) at the positions (0,1), (0,2), ..., (0, g) of the table, It is a means to create. 20. The apparatus according to claim 19, wherein said scene change rate calculating means stores j _C ≦ j ≦ J, 0 ≦
The first area is defined by i <j _C and ji ≦ g, and the sum of distances in the first area is represented by a (j _C ), and 0 ≦ i <j ≦ j _C −1, ji
≦ g defines a second area, the sum of the distances in the second area is represented as b (j _C ), and j _C ≦ i <j ≦ J, ji ≦ g defines a third area, The sum of the distances in the third area is b ′ (j
_C ), means for calculating C (tj _C ) = {a (j _C ) −βMAX [b (j _C ), b ′ (j _C )]} / S as the scene change rate. Is a predetermined coefficient for making the number of data in the first area equal to the number of data in the second or third area, and S represents the number of data in the first area. 21. The apparatus according to claim 19, wherein said scene change rate calculating means stores j _C ≤j≤J, 0≤
The first area is defined by i <j _C and ji ≦ g, and the sum of distances in the first area is represented by a (j _C ), and 0 ≦ i <j ≦ j _C −1, ji
≦ g defines the second area, the sum of the distances in the second area is represented as b (j _C ), and j _C ≦ i <j ≦ J, ji ≦ g defines the third area, The sum of the distances in the third area is b ′ (j
_C ), means for calculating C (tj _C ) = {a (j _C ) −β [b (j _C ) + b ′ (j _C )]} / S as the scene change rate, where β is The first region and the second region
And a predetermined coefficient for equalizing the number of data in the third area, and S represents the number of data in the first area. 22. The apparatus according to claim 16, wherein said scene change rate calculating means has at least one predetermined kind of predetermined cut detection template of the same size as said distance table, and at time t. This is means for calculating the degree of similarity with the obtained distance table as the scene change rate.