Method of automatically training a classifier hierarchy by dynamic grouping the training samples
AssigneeSeiko Epson Corporation
InventorsLihui Chen, Yang Yang
Life of this patent
Estimating the term…
Abstract
The present invention uses dynamic grouping to divide up training samples to train different classification nodes. At the beginning of the training, all samples are in the same group. A clustering process is applied in the feature space of the selected feature vectors with cluster indexes accumulated. The average of all the accumulated cluster indexes is used as the threshold for splitting the samples into two groups. When the splitting criterion is met, samples are split into two groups based on their similarity in the feature space.
Claims17
A method for training an object classifier hierarchy from a set of sample images, comprising:
- inputting a set of object image samples;
- and
- using a processor to:
- iteratively train an object classifier hierarchy as follows:
- (a) examine the object image samples in a feature space to extract feature vectors;
- (b) apply a clustering algorithm to the object image samples in the feature space of selected feature vectors;
- (c) split the object image samples into two groups using a splitting criterion based on the result of the application of the clustering algorithm;
- (d) assign one of the two groups of object image samples to train one node of the object classifier hierarchy and assign another of the two groups of object image samples to train another node of the object classifier hierarchy;
- and
- repeat steps (a)-(d) until a stopping threshold is met;
- wherein the splitting criterion is:
- f
- T
- FA
- >
- δ
- where f denotes a false alarm rate of a current ensemble classifier, TFA denotes a target false alarm rate, and δ≧1 is a threshold controlling a tolerance of non-perfect convergence.
A device for training an object classifier hierarchy from a set of sample images, comprising:
- a processor that:
- iteratively trains an object classifier hierarchy as follows:
- (a) examines object image samples in a feature space to extract feature vectors;
- (b) applies a clustering algorithm to the object image samples in the feature space of selected feature vectors;
- (c) splits the object image samples into two groups using a splitting criterion based on the result of the application of the clustering algorithm;
- (d) assigns one of the two groups of object image samples to train one node of the object classifier hierarchy and assign another of the two groups of object image samples to train another node of the object classifier hierarchy;
- and
- repeats steps (a)-(d) until a stopping threshold is met
- wherein the splitting criterion is:
- f
- T
- FA
- >
- δ
- where f denotes a false alarm rate of a current ensemble classifier, TFA denotes a target false alarm rate, and δ≧1 is a threshold controlling a tolerance of non-perfect convergence.
One or more tangible, non-transitory computer-readable media embodying instructions executable by a computer to perform a method for training an object classifier hierarchy from a set of sample images, the method comprising:
- inputting a set of object image samples;
- and
- using a processor to:
- iteratively train an object classifier hierarchy as follows:
- (a) examine the object image samples in a feature space to extract feature vectors;
- (b) apply a clustering algorithm to the object image samples in the feature space of selected feature vectors;
- (c) split the object image samples into two groups using a splitting criterion based on the result of the application of the clustering algorithm;
- (d) assign one of the two groups of object image samples to train one node of the object classifier hierarchy and assign another of the two groups of object image samples to train another node of the object classifier hierarchy;
- and
- repeat steps (a)-(d) until a stopping threshold is met;
- wherein the splitting criterion is:
- f
- T
- FA
- >
- δ
- where f denotes a false alarm rate of a current ensemble classifier, TFA denotes a target false alarm rate, and δ≧1 is a threshold controlling a tolerance of non-perfect convergence.
Dependencies are inferred from the claim text.