Semantic Event Detection Using Cross-Domain Knowledge

Abstract

The present invention relates to a method that facilitates semantic event classification of a set of image records related to an event. The method utilizes an event detector system to provide: extracting a plurality of visual features from each image record; wherein the visual features comprise dividing the image record into a number of regions in which the visual features are extracted; using the The visual features generate a plurality of concept scores for each image record, wherein each concept score corresponds to a visual concept, and each concept score represents the probability that the image record includes the visual concept; the concept score generation based on the image record corresponds to the event and providing the feature vector to an event classifier that identifies at least one semantic event classifier corresponding to the event.

Description

technical field [0001] The present invention relates to classifying digital content records, such as digital still images or video. In particular, the invention relates to the classification of digital content records based on semantic event detection. Background technique [0002] The advent of low-cost consumer electronics imaging technology has resulted in a dramatic increase in the number of digital images captured by the average user. In fact, as various forms of electronic storage have become cheaper over time, users have tended to take more digital still images and video, and to keep digital still images and video that they would have otherwise discarded. As such, the average user faces increasing difficulties in properly identifying and categorizing digital images for storage and later retrieval. Typically, such identification and classification is usually performed manually, which is an extremely time-consuming process for the user. [0003] As just one example, ...

AI Technical Summary

Problems solved by technology

However, there are fundamental problems with semantic event detection as follows: first, practical systems need to be able to handle both digital still images and videos, since both digital still images and videos usually exist in the image corpora of real users; second, practical systems need to accommodate real different semantic content in the user corpus, thus making it ideal to provide systems that include general methods for detecting different semantic events rather than specific individual methods for detecting each specific semantic event; finally, practical systems need to be robust against Errors in identification and classification

Images