Method for selecting frames from video sequences based on incremental improvement
The present application is directed to new methods for automatically and effectively combining multiple metrics to select interesting, or key, frames from a video sequence. In one aspect, the methods include selecting a reference frame from a video sequence and iteratively computing the improvement or gain in the combined metrics for candidate frames with respect to that reference frame. In another aspect, a master list of reference key frames is determined, quality metrics are calculated for each of a plurality of candidate frames associated with each reference key frame, and each reference key frame in the master list is replaced with a candidate frame that yields the greatest aggregate quality improvement over that reference key frame.
1. A method of selecting key frames from a video sequence, comprising:
determining a master list of reference key frames using a key frame extraction algorithm; and
for each reference key frame in the master list:
selecting a plurality of candidate frames adjacent to the reference key frame:
evaluating the plurality of candidate frames on one or more quality metrics;
selecting from the plurality of candidate frames a candidate key frame that yields an improvement over the reference key frame as measured by the one or more quality metrics; and
replacing the reference key frame in the master list with the candidate key frame,
wherein evaluating the plurality of candidate frames on one or more quality metrics comprises:
calculating a first quality metric score for each of the plurality of candidate frames;
selecting a first candidate frame from among the plurality of candidate frames, wherein the first candidate frame has the highest score for the first quality metric:
calculating a second quality metric score for each of the plurality of candidate frames;
selecting a second candidate frame from among the plurality of candidate frames, wherein the second candidate frame has the highest score for the second quality metric;
calculating a first aggregate quality improvement of the first candidate frame over the reference key frame; and
calculating a second aggregate quality improvement of the second candidate frame over the reference key frame.
2. The method of claim 1 , wherein selecting from the plurality of candidate frames a candidate key frame that yields an improvement over the reference key frame as measured by the one or more quality metrics comprises:
if the first aggregate quality improvement is greater than the second aggregate quality improvement and the first aggregate quality improvement indicates that the first candidate frame has a higher quality than the reference key frame, selecting the first candidate frame as the candidate key frame; and
if the second aggregate quality improvement is greater than the first aggregate quality improvement and the second aggregate quality improvement indicates that the second candidate frame has a higher quality than the reference key frame, selecting the second candidate frame as the candidate key frame.
3. A method of selecting key frames from a video sequence, comprising:
determining a master list of reference key frames using a key frame extraction algorithm;
for each reference key frame in the master list:
determining a plurality of candidate frames associated with the reference key frame;
calculating a first quality metric score for each of the plurality of candidate frames;
selecting a first candidate frame from among the plurality of candidate frames, wherein the first candidate frame has the highest score for the first quality metric;
calculating a second quality metric score for each of the plurality of candidate frames;
selecting a second candidate frame from among the plurality of candidate frames, wherein the second candidate frame has the highest score for the second quality metric;
calculating a first aggregate quality improvement of the first candidate frame over the reference key frame;
calculating a second aggregate quality improvement of the second candidate frame over the reference key frame;
if the first aggregate quality improvement is greater than the second aggregate quality improvement and the first aggregate quality improvement indicates that the first candidate frame has a higher quality than the reference key frame, replacing the reference key frame in the master list with the first candidate frame; and
if the second aggregate quality improvement is greater than the first aggregate quality improvement and the second aggregate quality improvement indicates that the second candidate frame has a higher quality than the reference key frame, replacing the reference key frame in the master list with the second candidate frame.
4. The method of claim 3 , wherein calculation of the first quality metric for a frame is based on the sharpness of the frame.
5. The method of claim 3 , wherein calculation of the second quality metric for a frame comprises:
detecting regions in the frame in which a human face appears; and
evaluating the quality of the detected regions.
6. The method of claim 3 , further comprising the following steps:
designating a subset of reference key frames in the master list as final frames prior to performing any quality improvement replacements;
creating a final list of key frames, wherein the final list comprises the final frames from the master list as modified by the quality improvement replacements; and
adding to the final list of key frames any additional key frames from the master list for which the quality improvement replacements accomplished an aggregate improvement greater than a predefined threshold.
7. The method of claim 3 , wherein determining a plurality of candidate frames associated with a reference key frame includes:
determining whether any two pluralities of candidate frames share one or more candidate frames; and
modifying one or both of the pluralities of candidate frames such that no candidate frame is shared by two pluralities of candidate frames.
8. A method of selecting a final list of key frames from a video sequence, comprising:
determining a master list of reference key frames using a key frame extraction algorithm;
designating a subset of reference key frames in the master list as final frames;
for each reference key frame in the master list:
determining a plurality of candidate frames associated with the reference key frame;
calculating a first quality metric score for each of the plurality of candidate frames;
selecting a first candidate frame from among the plurality of candidate frames, wherein the first candidate frame has the highest score for the first quality metric;
calculating a second quality metric score for each of the plurality of candidate frames;
selecting a second candidate frame from among the plurality of candidate frames, wherein the second candidate frame has the highest score for the second quality metric;
calculating a first aggregate quality improvement of the first candidate frame over the reference key frame:
calculating a second aggregate quality improvement of the second candidate frame over the reference key frame;
if the first aggregate quality improvement is greater than the second aggregate quality improvement and the first aggregate quality improvement indicates that the first candidate frame has a higher quality than the reference key frame, replacing the reference key frame in the master list with the first candidate frame;
if the second aggregate quality improvement is greater than the first aggregate quality improvement and the second aggregate quality improvement indicates that the second candidate frame has a higher quality than the reference key frame, replacing the reference key frame in the master list with the second candidate frame;
creating a final list of key frames, wherein the final list comprises the final frames from the master list as modified by the quality improvement replacements; and
adding to the final list of key frames any additional key frames from the master list for which the quality improvement replacements accomplished an aggregate improvement greater than a set threshold.
9. The method of claim 8 , wherein calculation of the first quality metric for a frame is based on the sharpness of the frame.
10. The method of claim 8 , wherein calculation of the second quality metric for a frame comprises:
detecting regions in the frame in which a human face appears; and
evaluating the quality of the detected regions.
11. The method of claim 8 , wherein determining a plurality of candidate frames associated with a reference key frame includes:
determining whether any two pluralities of candidate frames share one or more candidate frames; and
modifying one or both of the pluralities of candidate frames such that no candidate frame is shared by two pluralities of candidate frames.
12. A method of selecting key frames from a video sequence, comprising:
determining a master list of reference key frames using a key frame extraction algorithm; and
for each reference key frame in the master list:
selecting a plurality of candidate frames adjacent to the reference key frame;
evaluating the plurality of candidate frames on a first quality metric;
selecting from the plurality of candidate frames a candidate key frame that yields an improvement over the reference key frame as measured by the first quality metric;
replacing the reference key frame in the master list with the candidate key frame;
selecting a plurality of candidate frames adjacent to the reference key frame;
evaluating the plurality of candidate frames on a second quality metric;
selecting from the plurality of candidate frames a candidate key frame that yields an improvement over the reference key frame as measured by the second quality metric;
replacing the reference key frame in the master list with the candidate key frame;
wherein calculation of the first quality metric for a frame is based on the sharpness of the frame, and wherein calculation of the second quality metric for a frame comprises detecting regions in the frame in which a human face appears and evaluating the quality of the detected regions.
13. A method of selecting key frames from a video sequence, comprising:
determining a master list of reference key frames using a key frame extraction algorithm; and
for each reference key frame in the master list:
selecting a plurality of candidate frames adjacent to the reference key frame;
evaluating the plurality of candidate frames on one or more quality metrics;
selecting from the plurality of candidate frames a candidate key frame that yields an improvement over the reference key frame as measured by the one or more quality metrics;
replacing the reference key frame in the master list with the candidate key frame;
designating a subset of reference key frames in the master list as final frames prior to performing any quality improvement replacements;
creating a final list of key frames, wherein the final list comprises the final frames from the master list as modified by the quality improvement replacements; and
adding to the final list of key frames any additional key frames from the master list for which the quality improvement replacements accomplished an aggregate improvement greater than a predefined threshold.