Systems, methods, and media for on-line boosting of a classifier

摘要

Systems, methods, and media for on-line boosting of a classifier are provided, comprising: receiving a training sample; for each of a plurality of features, determining a feature value for the training sample and the feature, using the feature value to update a histogram, and determining a threshold for a classifier of the feature; for each of the plurality of features, classifying the training sample using the threshold for the classifier of the feature and calculating an error associated with the classifier; selecting a plurality of best classifiers from the classifiers; and, for each of the plurality of best classifiers, assigning a voting weight to the one of the plurality of best classifiers.

主权项

1. A system for on-line boosting of a classifier, comprising:
a hardware processor that is configured to:
access a selection of a plurality of image features for the classifier;receive an online training sample not used to select the plurality of image features;for each of the plurality of image features:
determine a feature value for the online training sample and the feature using the online training sample;determine one of a plurality of bins of a histogram that corresponds to the online training sample based on the feature value for the online training sample, wherein:
when the feature value falls below a lower range of the plurality of bins, a lowest bin in the plurality of bins is determined to be the one of the plurality of bins; andwhen the feature value falls above an upper range of the set of bins, a highest bin in the plurality of bins is determined to be the one of the plurality of bins;use the feature value to update the one of the plurality of bins of the histogram; anddetermine a threshold for a classifier of the feature;for each of the plurality of image features, classify the training sample using the threshold for the classifier of the feature and calculate an error associated with the classifier;select a plurality of best classifiers from the classifiers, wherein each of the plurality of best classifiers have not been previously selected as a best classifier; andfor each of the plurality of best classifiers, assign a voting weight to the one of the plurality of best classifiers that contributes to a final classifier, wherein the final classifier is a linear classifier that is a weighted combination of the plurality of best classifiers.