Video summarization using sparse basis function combination
A method for determining a video summary from a video sequence including a time sequence of video frames, comprising: defining a global feature vector representing the entire video sequence; selecting a plurality of subsets of the video frames; extracting a frame feature vector for each video frame in the selected subsets of video frames; defining a set of basis functions, wherein each basis function is associated with the frame feature vectors for the video frames in a particular subset of video frames; using a data processor to automatically determine a sparse combination of the basis functions representing the global feature vector; determining a summary set of video frames responsive to the sparse combination of the basis functions; and forming the video summary responsive to the summary set of video frames.
1. A method for determining a video summary from a video sequence including a time sequence of video frames, comprising:
a) determining a global feature vector representing the entire video sequence;
b) selecting a plurality of subsets of the video frames;
c) determining a frame feature vector for each video frame in the selected subsets of video frames;
d) defining a set of basis functions, wherein each basis function is associated with the frame feature vectors for the video frames in a particular subset of video frames;
e) using a data processor to automatically determine a sparse combination of the basis functions representing the global feature vector;
f) determining a summary set of video frames responsive to the sparse combination of the basis functions; and
g) forming the video summary responsive to the summary set of video frames;
wherein the sparse combination of the basis functions is determined by finding a sparse vector of weighting coefficients .alpha. by minimizing D(y,Aα) subject to the constraint ||α|| 0 <m, where y is the global feature vector, A is a matrix representation of the set of basis functions, D(°) is a dissimilarity function which returns a number that represents the dissimilarity between its two arguments, m is the number of non-zero elements in α, ||°|| 0 denotes l 0 norm;
wherein the video frames in the summary set of video frames corresponding to the largest weighting coefficients are merged in a manner that preserves the temporal order to form the video summary.
2. The method of claim 1 wherein the summary set of video frames are the video frames included in the subsets of video frames corresponding to the basis functions included in the sparse combination of basis functions.
3. The method of claim 1 wherein temporal order is preserved in the subsets of video frames.
4. The method of claim 1 wherein the subsets of the video frames are selected by randomly sampling video frames from the video sequence while preserving the temporal order.
5. The method of claim 1 wherein the video frames in the video sequence are clustered into groups of similar frames, and wherein the subsets of the video frames are selected based on the groups of similar frames.
6. The method of claim 1 wherein the video frames in the video sequence are analyzed to determine a quality score, and wherein video frames having higher quality scores are preferentially included in the subsets of the video frames.
7. The method of claim 6 wherein the quality score for a video frame is determined based on detecting the presence of one or more faces in the video frame, estimating a noise level for the video frame, or estimating a sharpness level for the video frame.
8. The method of claim 1 wherein the dissimilarity function applies a set of pre-defined operations to change one of its arguments into the other and determines a dissimilarity value by summing cost values associated with the application of the pre-defined operations.
9. The method of claim 1 wherein the pre-defined operations include substitution, insertion and deletion.
10. The method of claim 1 wherein the video frames in the summary set of video frames are merged in a manner that preserves the temporal order to form the video summary.
11. The method of claim 1 wherein the sparse combination of the basis functions is formed by determining weighting coefficients for each basis function, wherein only a minority of the weighting coefficients are non-zero.
12. The method of claim 11 wherein less than 10% of the weighting coefficients are non-zero.
13. The method of claim 11 wherein the number of non-zero weighting coefficients is selected to correspond to a specified time duration for the video summary.