Semantic event detection for digital content records
View Patent ↗A system and method for semantic event detection in digital image content records is provided in which an event-level “Bag-of-Features” (BOF) representation is used to model events, and generic semantic events are detected in a concept space instead of an original low-level visual feature space based on the BOF representation.
1. A method for facilitating semantic event classification of a group of image records related to an event, the method comprising:
extracting a plurality of visual features from each of the image records;
generating a plurality of concept scores for each of the image records using the visual features, wherein each concept score corresponds to a visual concept and each concept score is indicative of a probability that the image record includes the visual concept;
producing a codebook of semantic events by determining a pairwise similarity between pairs of data points corresponding to training events and applying spectral clustering to group the data points into different clusters with each cluster corresponding to one code word, based on the determined pairwise similarities;
mapping the training events to the codebook to generate a Bag-of-Features (BOF) feature vector corresponding to each training event;
training an event classifier based on the BOF features vectors corresponding to the training events;
generating a feature vector corresponding to the event based on the concept scores of the image records;
i) wherein the generation of the feature vector includes calculating a pairwise similarity between the concept scores of the image records and concept scores for predetermined training data points to generate individual feature vectors for each of the image records; and
ii) mapping the individual feature vectors to the codebook of semantic events; and
supplying the feature vector to the event classifier that identifies at least one semantic event.
2. A method for facilitating semantic event classification as claimed in claim 1 , wherein the image records include at least one digital still image and at least one video segment.
3. A method for facilitating semantic event classification as claimed in claim 2 , wherein the extracting of a plurality of visual features includes extracting a keyframe from the at least one video segment and extracting the plurality of visual features from both the keyframe and the at least one digital still image.
4. A method for facilitating semantic event classification as claimed in claim 3 , wherein the generation of the concept scores includes generating initial concept scores for each keyframe and each digital still image corresponding to each of the extracted visual features.
5. A method for facilitating semantic event classification as claimed in claim 4 , wherein the generation of the concept scores further includes generating an ensemble concept score for each keyframe and each digital still image based on the initial concept scores.
6. A method for facilitating semantic event classification as claimed in claim 5 , wherein the ensemble concept scores are generated by fusing individual concept scores for each extracted visual feature for a given keyframe or a given digital still image.
7. A method for facilitating semantic event classification as claimed in claim 1 , further comprising tagging each of the image records with the event classifier.