Method and apparatus for reduction of visual content

[0001] This invention relates generally to visual information processing.

[0002] Video information can be provided in many forms including sequences of video images that, when displayed in a sufficiently rapid fashion, present convincingly moving images. In some circumstances, video information (and particularly moving video information) can comprise a voluminous quantity of information. For example, a videotape that includes the output of a surveillance camera in a store can store and represent hours of surveillance information. There are times when a reviewer will wish to determine when and if something of interest has been recorded on such a videotape. In the past, such a reviewer must usually either review the entire tape in real time playback, or utilize a fast-forward/preview feature to attempt to locate the video information of interest in a more rapid fashion.

[0003] There are times when such techniques do not produce acceptable results, either because the review requires too much time or because the information of interest is not reliably noted. In an attempt to address these concerns, other prior art techniques seek to provide a corresponding video sequence summary that essentially provides a temporal thumbnail representation of the original video sequence. Such a summary typically requires reduced storage requirements and can be viewed in a reduced amount of time as compared to the original video sequence. When the summarization technique utilizes only time-based frame/video content reduction, however, the resultant summary may well exclude video content of importance and/or concern to a future viewer. Some content-based abridgement techniques have been suggested, but in general many such techniques tend to be quite computationally complex (with some even requiring multiple iterations of the content-reduction process). Other techniques seem better attuned to classifying a given video sequence instead of yielding a resultant video summary. As a result, a content-based mechanism or process to automatically yield a generally useful video summary of a given plurality of graphic images in a relatively non-complex fashion and at reasonable cost remains unmet.

[0014] Skilled artisans will appreciate that elements in the figures are illustrated for simplicity and clarity and have not necessarily been drawn to scale. For example, the dimensions of some of the elements in the figures may be exaggerated relative to other elements to help to improve understanding of various embodiments of the present invention. Also, common but well-understood elements that are useful or necessary in a commercially feasible embodiment are typically not depicted in order to facilitate a less obstructed view of these various embodiments of the present invention.

[0015] Generally speaking, pursuant to these various embodiments, an estimator provides a first significance metric for at least some of the frames in an original plurality of sequential frames of video information. The estimator identifies those frames having a corresponding significance metric that has at least a predetermined relationship with respect to at least a first video significance threshold value as being key frames (for example, in a preferred embodiment, the estimator identifies as key frames those frames having a significance metric that at least equals a predetermined threshold). The estimator then determines a cumulative significance value for at least some of the frames that fall between a pair of the key frames. In a preferred embodiment, the estimator essentially integrates over time a second corresponding significance metric for each such frame. These resulting values are then compared against a set of iteratively increasing thresholds to identify frames of interest. A video summary generator then generates a video summary comprising key frames and at least some of these other frames of interest (presuming, of course, that the video content contains sufficient information of interest and/or that the thresholds are not set so high as to essentially exclude all candidate content). The resultant summary will typically include visual content of interest and value while deleting relatively static content.

[0016] The degree of effective time compression can be scaled relatively easily by appropriate selection of the various thresholds noted above. Depending upon the needs of a given application, the iteratively increasing thresholds can increase in a linear fashion or a non-linear fashion. In addition, the increases can proceed pursuant to a fixed schedule and/or can increase pursuant to a content-sensitive dynamic scheme.

[0017] In a preferred embodiment, the significance metrics reflect at least one of changes in position/motion and/or color in the visual content from one frame to another. Such a metric therefore tends to reflect either movement and/or a change of scene. Generally speaking, such a process can be effected automatically and in near real-time to the original image capture, thereby potentially reducing storage requirements and facilitating rapid review by an observer without necessarily requiring extensive pre-processing.

[0018] Referring now to FIG. 1, as already noted, a plurality of frames that contain at least video information are provided 11. In a preferred embodiment, these frames will comprise frames of video information that will represent an animated scenario when displayed sequentially in accordance with well-understood prior art technique. It should be understood, however, that these embodiments may also be useful with other kinds of graphic information. Therefore these embodiments should not be viewed as being limited to standard motion video methodologies.

[0019] One or more significance metrics are then determined 12 for at least some of these frames (and preferably all of the frames). In a preferred embodiment, the significance metric represents detection of a change to position, color, or both as pertains to the graphic content of a given frame. Preferably, the metric represents a comparison between the graphic content of a given frame and a previous frame in the sequence of frames (most preferably, the given frame will be compared with a most-recent sequentially previous frame). Other indicia (such as changes to texture) could also be used, alone or in combination, with these criteria as desired. In general, the indicia should be representative of either a scene change and/or motion of one or more depicted objects. So selected, the metric should tend, in a preferred approach, to yield a higher value as the degree and/or overall quantity of movement or other changes from one frame to the next increases.

[0020] The process then provides for identification 13 of frames that appear to represent a significant change in content as compared to a previous frame. For example, the first frame that represents a change of scene in an edited presentation will tend to represent a significant change of visual content as compared to the last frame of the previous scene. As another example, consider a surveillance film of a point-of-sale location in a store. The first frame when a patron first enters the scene will often represent a significant visual change from the preceding frame. Such frames are identified by comparing the significance metric determined above with a first video significance threshold value. This first video significance threshold value can be set as desired and appropriate to a given application, but in general should preferably be set high enough to typically permit accurate identification of such frames that appear to capture the initiation of a significant scene change and/or action sequence. For purposes of this description, such frames are referred to as key frames.

[0021] It will be observed that a first frame in a sequence of frames has no preceding frame or frames with which it can be compared. In a preferred embodiment, the first frame in a sequence of frames will nevertheless always be treated as a key frame. There are situations, however, when one might wish to operate in a different fashion. For example, if the opening frame is substantially identical to the next 10,000 frames, virtually any of those frames could be selected as a key frame with substantially similar results being expected with respect to the resultant summary produced below. In general, however, and especially for ease of automation and reduced computational complexity, selection of the first frame as a key frame will tend to support effective summarization results.

[0022] Cumulative visual significance values are then determined 14 for most (and preferably all) of the frames (or at least those frames that were not previously identified as being key frames). In a preferred approach, these cumulative visual significance values are accumulated with respect to frames that intervene between two key frames. Also preferably, these values are accumulated with respect to all of the frames that so intervene. A more detailed example will be provided below with respect to one approach to forming such cumulative significance values.

[0023] The process then identifies 15 those particular frames that have a corresponding cumulative visual significance value that at least equals a predetermined threshold value. In a preferred embodiment, a series of thresholds that increase in an incremental fashion are used to identify the frames of interest in this way. For example, a first threshold having value X may be used to identify a first frame in a sequence of frames following a given key frame that at least equals this threshold X. Upon identifying this frame, the threshold can then be increased by a known amount (for example, the threshold can be doubled to thereby provide a threshold of 2X). The identification process can then be continued to determine if any subsequent frame (prior to the next key frame) will at least equal this new threshold 2X. Upon identifying such a frame, the threshold can then be increased again (for example, by tripling the threshold to provide a threshold of 3X) and the comparison process continued. In this way, frames having a cumulative visual significance value that at least equals an increasing threshold value are identified between each pair of key frames.

[0024] Those identified frames, along with the key frames, are then selected 16 in a preferred embodiment as frames to be used to comprise a visual summary of the original sequence of frames. For example, these selected frames can be displayed in accord with the original frame rate for the original sequence of frames. Since this process typically results in the removal of a considerable number of frames (i.e., all frames that are not key frames and/or that are not otherwise selected), the corresponding resultant summary video will be viewable in a considerably shortened period of time.

[0025] Notwithstanding a significant compression of time that can be achieved with such a process, the resultant summary video tends to be inclusive of content of interest. This occurs at least in part because the frames that are selected for inclusion in the summary tend to either be frames that begin a new scene or frames that otherwise represent a considerable change from a previous frame(s). As a result, frames that contain redundant visual information tend to be excluded from the summary while frames that include new visual information tend to be included. In some settings, such as reviewing a surveillance tape, the resultant time compression can be significant while simultaneously preserving considerable content of potential interest.

[0026] These various embodiments can be effected through various means. Referring now to FIG. 2, a particular illustrative embodiment will be described. FIG. 2 depicts a high level block diagram of a system 20 that can support generation of video summaries as per the above embodiments. A Frame Visual Significance and Cumulative Visual Significance Estimator 21 takes the input video frames along with weighting factors (w1 and w2) to compute visual significance values VSV and cumulative visual significance values CVSV as noted above. The VSV and CVSV results are input to a Video Summary Generator 22 along with threshold values (t1 and t2) to generate a corresponding video summary.

[0027] FIG. 3 presents a detailed representation of the Frame Visual Significance and Cumulative Visual Significance Estimator 21. In this embodiment, Color Layout Descriptor (CLD) and Motion Activity Descriptor (MAD) (as these descriptors are defined by the well known standard MPEG-7) Extractors 31A and 31B respectively process the input video frames to obtain representative CLD and MAD descriptions. Estimators 32A and 32B then generate visual significance value functions using the CLD and MAD descriptions independently, VSV_cld and VSV_mad respectively. The overall visual significance value for a given frame n, VSV[n], is computed in this embodiment as a weighted sum 33 of VSV_cld[n] and VSV_mad[n], as shown in equation (1).
VSV[n]=w1*VSV_—cld[n]+w2*VSV_—mad[n]  (1)
The weighting values w1 and w2 reflect the relative importance of VSV_cld and VSV_mad respectively and can be selected as appropriate to a given application or visual setting/context (for example, when motion information is more likely to accompany visual content of interest than changes to color information, these weighting factors can be altered accordingly).

[0028] VSV_cld[n] is computed as a distance between the CLD of a current frame n and a just-previous frame n−1 (the latter being provided by an appropriate delay unit 34). The distance computation between CLDs can be as specified in the MPEG-7 Visual specification as understood by those skilled in these arts.
VSV_—cld[n]=Dis tan ce(CLD [n], CLD [n−1])  (2)

[0029] VSV_mad[n] is computed, in this embodiment, by the estimator 32B as the variance of the set of motion vectors, MV, within the frame n: $\begin{array}{cc}\mathrm{VSV\_mad}\left[n\right]=\left(E\lfloor {\mathrm{MV}\left(n\right)}^2\rfloor -E^2\left[\mathrm{MV}\left(n\right)\right]\right)/N_{\mathrm{MAD}}& \left(3\right)\end{array}$
Again, the above reflects normalization using N_MAD as a normalization factor by a second normalization unit 36. Generally, the value of N_MAD is a function of frame size. In a preferred embodiment, N_MAD is computed as the square root of the area of the frame. Where frameHeight is the height of the video frame in pixels and frameWidth is the width of the video frame in pixels: $\begin{array}{cc}{N}_{\mathrm{MAD}}=\sqrt{\mathrm{frameHeight}\times \mathrm{frameWidth}}& \left(4\right)\end{array}$

[0030] A cumulative VSV generator 37 computes the cumulative visual significance value function, CVSV[n], from the VSV[n] function: $\begin{array}{cc}\mathrm{CVSV}\left[n\right]=\{\begin{array}{c}\sum _{k=\mathrm{n0}}^n\mathrm{VSV}\left[n\right],\mathrm{When}\mathrm{frame}n\mathrm{is}\mathrm{not}a\mathrm{key}\mathrm{frame}\\ 0,\mathrm{When}\mathrm{frame}n\mathrm{is}a\mathrm{key}\mathrm{frame},\mathrm{n0}=n\end{array}& \left(5\right)\end{array}$
In equation (4), n0 corresponds to the last frame, less than n, that is identified as a key frame by analyzing VSV[n].

[0031] Referring now to FIG. 4, appropriate key frames and intervening enhancement frames are selected by processing VSV[n] and Cumulative Visual Significance CVSV[n].

[0032] Key frames are chosen when a first threshold test 41 indicates that VSV[n]>t1 (t1 being the first threshold described earlier). This corresponds to a significant change in frame content that typically corresponds to a scene change. The corresponding cumulative visual significance CVSV[n] is also set to zero when a key frame is so detected. This helps to prevent CVSV[n] over flow. Enhancement frames are chosen when a second threshold test 42 indicates that the increment of CVSV[n]>t2. This identifies an enhancement frame for a given amount of visual significance change. The values of the thresholds t1 and t2 should preferably correspond inversely to the number of key frames and enhancement frames directly. In general, with larger threshold values, the fewer the number of identified key frames and enhancement frames. Conversely, lower threshold values identify more key frames and enhancement frames.

[0033] The MUX unit 43 combines the identified key frames and enhancement frames to produce an overall video summary representation of the original video sequence.

[0034] In general, the thresholds t1 and t2 can be fixed, user specified, or computed to adapt locally to the video content. The thresholds t1 and t2 govern the temporal granularity of the key frames and the number enhancement frames within a video shot (i.e., in between key frames). In a preferred embodiment, threshold t2 will increment with each identified enhancement frame (until a key frame is again encountered, at which point the threshold can revert to the original lowest t2 value). In one embodiment, this threshold increments as a multiple of itself as already related above.

[0035] Without loss of generality, the video summaries may be encoded using known video compression methods such as H.263 and MPEG-4. These and other enabling embodiments can be realized through use of dedicated logic circuits and/or with programmable platforms such as microprocessors and microcontrollers as well understood in the art.

[0036] Referring now to FIG. 5, an example will be set forth. In FIG. 5, an original series 51 of twenty-eight frames is provided. In this illustration, these frames F1 through F28 each include visual information as captured in a temporal sequence as with a motion picture camera. It should be understood that only twenty-eight frames are shown in this example for purposes of simplicity and clarity; in general, considerably more frames are to be expected.

[0037] As related above, various embodiments provide a visual significance value for each of these frames. Such values are shown in FIG. 6 (in this example, a value has not been provided for the first frame F1 as explained above). Each of these values is then compared against a first threshold T1. In this example, three of the frames have a corresponding visual significance value that exceeds this threshold. In particular, frame F8 has a visual significance value 61 that exceeds this threshold T1, frame F14 has a visual significance value 62 that exceeds this threshold T1, and frame F22 has a visual significance value 63 that exceeds this threshold T1. As shown in FIG. 7, these frames (F8, F14, and F22) along with the first frame F1 are identified as key frames 71.

[0038] Referring now to FIG. 8, the cumulative visual significance values for each frame that intervenes between each pair of key frames 71 are determined. In this example, the cumulative visual significance value 81 for frame F2 equals the visual significance value as was determined for that same frame. The cumulative visual significance value 82 for frame F3 equals the sum of the visual significance value for frame F3 and the cumulative visual significance value for frame F2. This accumulation of values continues, such that the cumulative visual significance value 83 for frame F4 comprises the sum of the visual significance value for frame F4 and the cumulative visual significance value for frame F3, until the next key frame 71 has been reached. With each key frame 71, the cumulative visual significance value is reset to a starting point value (in this example, a value of zero).

[0039] Beginning with an initial frame, the cumulative visual significance values are compared against a second threshold T2. In this example, the cumulative visual significance value 87 for frame F6 comprises the first frame having a cumulative visual significance value to at least equal the second threshold T2. This identifies this frame F6 as being an enhancement frame to be included in the final summary. Pursuant to this illustrative embodiment, the threshold is now incremented to a higher value. In particular, in this embodiment, the process now uses a threshold value of 2T2 85. The cumulative visual significance value 88 that corresponds to frame F8 is the next value to at least equal the new threshold setting. The process then selects this frame F8 as another enhancement frame to be included in the summary. Frame F8 was also a key frame 71. The process now returns to the original second threshold value 84 and begins the process anew of comparing the next series of cumulative visual significance values with the incrementally increasing threshold values.

[0040] By proceeding in the above fashion, and as illustrated in FIG. 9, a number of frames 91 are selected for inclusion in a visual summary. In this example, frames F1, F6, F8, F11, F13, F14, F16, F18, F22, and F26 are so selected. A resultant visual summary would comprise these selected frames as presented in sequence without any intervening spaces. Such a visual summary will require less time to view as fewer frames are required for presentation. Notwithstanding this reduction in frames, considerable visual content of potential interest remains preserved as the processes described above tend to emphasize retention of frames that provide an indication of considerable motion or other metric of interest.

[0041] Those skilled in the art will recognize that a wide variety of modifications, alterations, and combinations can be made with respect to the above described embodiments without departing from the spirit and scope of the invention, and that such modifications, alterations, and combinations are to be viewed as being within the ambit of the inventive concept.