The present invention relates to MPEG-type video, and more particularly to scene cut detection within video bitstream.
MPEG-based video sequences can be divided into scenes that may vary from less than a second to minutes or more in length. Detection of these divisions or scene cuts makes possible a wide variety of value added features. For example, a frame can be selected from each scene to create a storyboard summarizing the contents of a video recording. Further, a video editor can be used to manipulate the video recording on a scene-by-scene basis, for example, re-ordering scenes or applying image-processing techniques to each frame in a scene.
MPEG video sequences include three types of frames, Intra-Frames (I), Inter-Frames (P), and Bi-Directional Frames (B). I frames encode a still image using a method similar to JPEG encoding. P frames are predicted from a previous I or P frame. B frames are predicted both from a previous I or P frame and a next I or P frame. These three types of frames are encoded using a Discrete Cosine Transform (DCT), which organizes redundancy in spatial directions between frames. However, for I frames, the DCT information is derived directly from an image sample, whereas for the P and B frames, the DCT information is derived from a residual error after prediction.
Each frame is divided into a plurality of macroblocks. Each macroblock includes information related to a plurality of luminance blocks, e.g., Y1, Y2, Y3 and Y4, and a plurality of chrominance blocks, e.g., one U and one V in a YUV system. Each of these blocks includes a plurality of pels, or picture elements, e.g., an 8×8 block.
When video has been encoded into an MPEG-like bitstream, scene cut detection can be done without fully decoding the bitstream, which increases the speed of video processing. Additional information, such as, macroblock encoding types, motion vectors, and DCT coefficients, may also be extracted from the bitstream without fully decoding the bitstream.
One method for scene cut detection is performed as follows:
In another proposed method for scene cut detection: for I frames, a difference between color histograms built from DC coefficients is used, and combined with information about the ratio of the number of macroblocks without motion compensation to the number with motion compensation. The proposed method looks for a peak in a plot versus frame number. For B frames, the ratio of forward to backward predictions is determined. In all cases, a local adaptive threshold technique is implemented to identify peaks.
Yet another method makes use of histograms for all frames (I, P, and B) built from Discrete Cosine (DC) coefficients with motion compensation.
However, no known system or method currently exists for scene cut detection based on global examination of all of the predictions within a subgroup of pictures (GOP).
Scene cuts are detected based upon global examination of all of the predictions within a sub-GOP in accordance with the principles of the present invention. The number of coded macroblocks for each of a plurality of potential scene cuts are first detected. The number of predicted macroblocks for each of the plurality of potential scene cuts is then determined. The ratio of the number of coded macroblocks to the number of predicted macroblocks for each of a plurality of frames corresponding to the potential scene cuts is next determined. A frame having the desired ratio is selected. The ratio is compared to a threshold value. A scene cut is detected if it is determined that the ratio satisfies the threshold value.
Preferred embodiments of the present invention will be described below in more detail, with reference to the accompanying drawings, in which:
In accordance with the principles of the present invention, scene cut detection is performed by collecting data on the number of forward and backward predicted macroblocks in each Inter-Frame (P) or Bi-Directional (B) frame, analyzing them on a sub-GOP level (each anchor frame along with preceding B frames) without considering each frame independently. It is to be understood that the present invention may be implemented in various forms of hardware, software, firmware, special purpose processors, or a combination thereof. In one exemplary embodiment, the present invention is implemented in software as an application program tangibly embodied on a program storage device. The application program may be uploaded to, and executed by, a machine comprising any suitable architecture. Preferably, the machine is implemented on a computer platform having hardware such as one or more central processing units (CPU), a random access memory (RAM), and input/output (I/O) interface(s). The computer platform also includes an operating system and micro-instruction code. The various processes and functions described herein may either be part of the micro instruction code or part of the application program (or a combination thereof) which is executed via the operating system. In addition, various other peripheral devices may be connected to the computer platform such as an additional data storage device(s) and a printing device(s).
It is to be further understood that, because some of the constituent system components and method steps depicted in the accompanying figures may be implemented in software, the actual connections, interconnections, or functional relationships between the system components (or the process steps) may differ depending upon the manner in which the present invention is programmed. Given the teachings of the present invention provided herein, one of ordinary skill in the related art will be able to contemplate these and similar implementations or configurations of the present invention.
For each P or B frame, the number of macroblocks employing forward prediction and the number of macroblocks employing backward prediction (found only in B frames) are recorded. The data for each anchor frame, Intra-frame (I) or P, and all immediately preceding B frames is examined.
If a macroblock is predicted improperly, the residual after motion compensation can be large, and large DCT coefficients are needed in order to compensate for the prediction. For these predictions, there is a high (about greater than 50%) likelihood that the prediction is spurious. That is, an anchor-frame macroblock on which the prediction is based may have no real association with the macroblock in the predicted frame, but the anchor-frame may be similar enough that a prediction could be made from it. For this reason, such a macroblock is ignored in gathering statistics. More particularly, when counting macroblocks with forward or backward prediction, the total energy (entropy) of the DCT coefficients is examined. The total energy of the DCT coefficients is the sum of the squares of the non-DC coefficients. If this energy is above a threshold value, the macroblock is not counted, either for the number of forward or backward predictions (nFwd and nBwd respectively), or for the number of coded macroblocks (nCode). For MPEG-2, one exemplary threshold value has been found to be about 500.
For an anchor frame, the pattern of predictions for each frame in the sub-GOP are analyzed. Consider
According to an embodiment of the present invention, if a scene cut occurs at 108, few forward-predicted macroblocks would be present in frames B1 102, B2 103, or P3 104, because frame I/P0 101 is unrelated to any of these frames. Likewise, if a scene cut occurs at 109, few backward-predicted macroblocks would be present in frame B1 102 because frame B1 102 is unrelated to frame P3 104 from which these predictions are made; and we expect few forward-predicted macroblocks in frame B2 103 or P3 104 because frame I/P0 101 is unrelated to these frames.
Thus, a measure of the likelihood of a scene cut at 108 is the total of the number of forward-predicted macroblocks in frames B1 102, B2 103, or P3 104. A measure of the likelihood of a scene cut at 109 is the total of the number of backward-predicted macroblocks in frame B1 102 plus the number of forward-predicted macroblocks in frame B2 103 and P3 104; and similarly for the likelihood of a scene cut at 110. The likelihood that there is a scene cut at the corresponding point increases as the total number of forward and backward-predicted macroblocks decreases.
Referring to
Referring now to
For each field f in the sub-GOP, it is determined whether the field f is in an I picture, in accordance with step 307. If the field f is in an I picture, then the process is continued for the next field in accordance with step 308. Upon determining that field f is not in an I picture, then the process loops over the potential sites for a scene cut in the sub-GOP, that is, each site f′ between two fields. The value of codedMBs[f′] is incremented by ncode[f] for each potential scene cut f′ in the sub-GOP, in accordance with step 309. Further, the value of predMBs[f′] is incremented by nFwd[f] for each potential scene cut f′ before or equal to f, in accordance with step 310.
In accordance with step 311, it is determined whether the last field has been reached, and if not the process proceeds with the next field, in accordance with step 308. Upon determining that the last field has been reached, a field f for which the ratio of predMBs[f] to coded MBs[f] (e.g., predMBs[f] to coded MBs[f]) is at a minimum is determined, in accordance with step 312. For the field f determined to have the minimum ratio, if predMBs[f] divided by coded MBs[f] is less than a predRatioThresh in accordance with step 313, the process detects and reports a scene cut between this field and the preceding field, in accordance with step 314, else no scene cut is detected in the sub-GOP, in accordance with step 315.
Note that the number of predicted macroblocks (predMBs) and field coded macroblocks (codedMBs) are determined for each field, even if the pictures are frame pictures, since a scene cut may occur between fields of a frame picture (if the frame picture is interlaced). The macroblocks of a frame picture are applied to each field of the frame in computing predMBs and codedMBs. That is, nCode[f′] is the total number of coded macroblocks in the frame when f′ refers to either field of the frame. Similarly, nFwd[f′] and nBwd[f′] represent the number of forward prediction macroblocks and backward prediction macroblocks in field f′, respectively. When f′ is in the same frame picture as f, predMBs[f] is incremented by nBwd[f], and codedMBs[f] is incremented by nCode[f′], only when f is the first field of the frame. The arrays for the second field are not incremented because forward predictions are assumed to apply to the first field only, and backward predictions to the second field only. If this were not done, a scene cut between fields of a frame picture would be masked by predictions across the scene boundary.
A good value for the prediction ratio threshold, preRatioThresh, has been found, through experimentation, to be 0.05. However, further experimentation and/or applications may determine that some value other than this value, or that a variable threshold value may be more appropriate in a specific application. Furthermore, the specific value chosen for AC energy threshold, AcThresh, may similarly be other than the exemplary value described here. It would be apparent to those skilled in the art to evaluate and modify these and other values to maximize the efficiency and accuracy of the invention, such modification still including the principles of the present invention as described herein.
The specific method described above is that which has been implemented and tested. It is not meant to limit the scope of this disclosure or the claims contained herein. An underlying principle of the present invention is the precise detection of a sharp scene cut by global examination of all the predictions within a sub-GOP. By seeking consistency of all the information available on predictions, false positives (spurious reports of scene cuts) may be reduced.
The principles of the present invention may be integrated with, for example, histogram-based approaches, to handle the cases of I-frame-only, and of a P picture followed by an I picture, in which cases this method cannot detect a scene cut. Extensions involving histograms can also be implemented to detect gradual scene transitions such as fades, dissolves, and wipes.
Having described embodiments for precise detection of a scene cut by global examination of all of the predictions within a sub-GOP, it is noted that modifications and variations can be made by persons skilled in the art in light of the above teachings. It is therefore to be understood that various changes and modifications may be made in the particular embodiments of the invention disclosed which are within the scope and spirit of the invention as defined by the appended claims. Having thus described the invention with the details and particularity required by the patent laws, what is claimed and desired protected by Letters Patent is set forth in the appended claims.
This application claims the benefit of U.S. Provisional Patent Application Ser. No. 60/324,890, filed Sep. 26, 2001.
Number | Name | Date | Kind |
---|---|---|---|
5642174 | Kazui et al. | Jun 1997 | A |
5724100 | Kuchibhotla | Mar 1998 | A |
5731835 | Kuchibholta | Mar 1998 | A |
5751378 | Chen et al. | May 1998 | A |
5774593 | Zick et al. | Jun 1998 | A |
5835149 | Astle | Nov 1998 | A |
5872598 | Legall et al. | Feb 1999 | A |
5949953 | Shirakawa et al. | Sep 1999 | A |
6025886 | Koda | Feb 2000 | A |
6081622 | Carr et al. | Jun 2000 | A |
6125147 | Florencio et al. | Sep 2000 | A |
6205174 | Fert et al. | Mar 2001 | B1 |
6229925 | Alexandre et al. | May 2001 | B1 |
6233278 | Dieterich | May 2001 | B1 |
6252905 | Pokrinchak et al. | Jun 2001 | B1 |
6259736 | Chujoh et al. | Jul 2001 | B1 |
6735253 | Chang et al. | May 2004 | B1 |
20010019631 | Ohsawa et al. | Sep 2001 | A1 |
Number | Date | Country | |
---|---|---|---|
20030058947 A1 | Mar 2003 | US |
Number | Date | Country | |
---|---|---|---|
60324890 | Sep 2001 | US |