IMAGE PATTERN RECOGNITION SYSTEM AND METHOD

摘要

An image pattern recognition method detects a pattern in a sequence of video images or individual images from detected interest points. Feature vectors are extracted with video data from video regions around the interest points. A forest of decision trees is used to compute a set of bin values in histograms with bins corresponding to leaf nodes of the decision trees. Each bin value is a sum of contributions computed for individual interest points. Non-binary decision functions are used to compute the contributions and node dependent scale values are used to compute the arguments of the non-binary decision functions. The node dependent scale values may be computed from standard deviations of feature values found for the nodes, multiplied by a factor that is common to the nodes. This factor may be adjusted by feedback so that it can be set differently for different detection classes.

主权项

1: An image pattern recognition system for detecting a pattern from at least one image based on detected interest points in the at least one image, the pattern recognition system comprising
a feature vector extractor configured to extract feature vectors, each feature vector comprising data derived from video content in a region in a predetermined relation to a location of a respective detected interest point in the at least one image; a bin value computing module, configured to compute a set of bin values for bins associated with leaf nodes of decision trees of a forest of decision trees, the bin value being computed for each bin by summing contributions corresponding to the feature vectors, each contribution being computed dependent on selections of feature vector components and thresholds associated with non-leaf nodes along a path through the decision tree to the leaf node associated with the bin; a classifier, configured to compute a pattern detection result by comparing the set of bin values to reference sets of values,wherein at least a first non-leaf node from the non-leaf nodes of the decision trees has an associated first scale value and at least a second non-leaf node from the non-leaf nodes of the decision trees has an associated second scale value, the bin value computing module being configured to use a soft decision function in the computation of the contributions for the bins associated with leaf nodes at ends of paths through the decision tree that include the first non leaf node, the contributions being computed in proportion to a result of applying the soft decision function to a ratio of the first scale value and a difference between the threshold and a feature value taken from the feature vector according to the selection of the feature vector component associated with the first non-leaf node, the second scale value being used in the computation of the contributions for the bins associated with leaf nodes at ends of paths through the decision tree that include the second non leaf node, the system comprising
a processing system configured to compute the first and second scale value by determining a first and second measure of statistical spread of the feature values taken from the feature vectors of a training set according to the selection of the feature vector component associated with the first and second non-leaf node respectively, and setting the first and second scale value to a product of a common factor and the first and second measure of statistical spread respectively.