초록 ▼

A system and method for determining a classifier to discriminate between two classes—object or non-object. The classifier may be used by an object detection program to detect presence of a 3D object in a 2D image (e.g., a photograph or an X-ray image). The overall classifier is constructed of

A system and method for determining a classifier to discriminate between two classes—object or non-object. The classifier may be used by an object detection program to detect presence of a 3D object in a 2D image (e.g., a photograph or an X-ray image). The overall classifier is constructed of a sequence of classifiers (or “sub-classifiers”), where each such classifier is based on a ratio of two graphical probability models (e.g., Bayesian networks). A discrete-valued variable representation at each node in a Bayesian network by a two-stage process of tree-structured vector quantization is discussed. The overall classifier may be part of an object detector program that is trained to automatically detect many different types of 3D objects (e.g., human faces, airplanes, cars, etc.). Computationally efficient statistical methods to evaluate overall classifiers are disclosed. The Bayesian network-based classifier may also be used to determine if two observations (e.g., two images) belong to the same category. For example, in case of face recognition, the classifier may determine whether two photographs are of the same person. A method to provide lighting correction or adjustment to compensate for differences in various lighting conditions of input images is disclosed as well. As per the rules governing abstracts, the content of this abstract should not be used to construe the claims in this application.

대표청구항 ▼

What is claimed is: 1. A method of generating a classifier, said method comprising: using a computer to perform the following steps: computing a wavelet transform of each of a plurality of two-dimensional (2D) images, wherein each said wavelet transform generates a corresponding plurality of transf

What is claimed is: 1. A method of generating a classifier, said method comprising: using a computer to perform the following steps: computing a wavelet transform of each of a plurality of two-dimensional (2D) images, wherein each said wavelet transform generates a corresponding plurality of transform coefficients; creating a plurality of candidate subsets of said transform coefficients, where each subset represents a group of image input variables that statistically depend upon each other; selecting a group of candidate subsets from said plurality of candidate subsets; and constructing said classifier based on a ratio of a plurality of Bayesian networks using said group of candidate subsets, wherein a Bayesian network comprises a probability distribution over a set of variables where statistical independence and conditional statistical independence exist among various combinations of the variables, wherein each of said Bayesian networks is a probability distribution representation derived from dependencies among said image input variables that statistically depend upon each other. 2. The method of claim 1, wherein corresponding transform coefficients constituting each of said plurality of candidate subsets are statistically dependent. 3. The method of claim 1, wherein said creating includes: creating said plurality of candidate subsets by minimizing an error in a log-likelihood ratio function. 4. The method of claim 1, wherein said constructing includes: creating a Bayesian network graph for said group of candidate subsets, wherein said Bayesian network graph is used by each of said plurality of Bayesian networks; representing variables in each conditional probability distribution at each node in each of said plurality of Bayesian networks by a pair of discrete valued functions; estimating each said conditional probability distribution within each of said plurality of Bayesian networks; and determining an object detection threshold for said ratio of said plurality of Bayesian networks. 5. The method of claim 4, wherein said creating said Bayesian network graph includes: forming a restricted Bayesian network for each combination of said plurality of candidate subsets; computing a log-likelihood ratio function for each said restricted Bayesian network; and evaluating a structure of each said log-likelihood ratio function using a criterion of area under an ROC (Receiver Operating Characteristic) curve. 6. The method of claim 5, wherein said evaluating said structure of each said log-likelihood ratio function includes choosing said group of candidate subsets that maximizes said area under said ROC curve. 7. The method of claim 5, wherein said restricted Bayesian network comprises: a first layer of nodes, wherein a node in said first layer of nodes corresponds to every variable that occurs in multiple candidate subsets; a second layer of nodes, wherein a node in said second layer of nodes corresponds to a candidate subset from said plurality of candidate subsets and consists of all variables in that subset except for those variables that occur in multiple candidate subsets; and a directed link from each first layer node to each second layer node where a combined group of variables from the first and second layer nodes connected by said directed link occur together within a candidate subset from said plurality of candidate subsets. 8. The method of claim 4, wherein said representing variables at each node includes: quantizing said variables in each said conditional probability distribution within each of said plurality of Bayesian networks using Tree-Structured Vector Quantization. 9. The method of claim 4, said estimating each said conditional probability distribution includes: estimating each said conditional probability distribution through an iterative AdaBoost procedure. 10. The method of claim 1, wherein said selecting includes: creating a plurality of candidate features from said plurality of candidate subsets; estimating probability distributions over each of said plurality of candidate features; evaluating probability of each of said plurality of 2D images on each of said probability distributions; and choosing said group of candidate subsets to maximize an area under an ROC (Receiver Operating Characteristic) curve using said probability evaluation. 11. The method of claim 1, wherein said constructing includes: constructing a sequence of classifiers, wherein a set of training examples are selected from said plurality of 2D images, and wherein said training examples representing each class differ in each stage in said sequence of classifiers. 12. The method of claim 11, wherein said set of training examples are selected for each stage by bootstrapping where training examples are collected from errors of an existing classifier sequence on a set of images from said plurality of 2D images. 13. The method of claim 1, wherein probability distributions forming each Bayesian network are created using an iterative Adaboost procedure. 14. A system for generating a classifier, said system comprising: means for computing a wavelet transform of each of a plurality of two-dimensional (2D) images, wherein each said wavelet transform generates a corresponding plurality of transform coefficients; means for creating a plurality of candidate subsets of said transform coefficients, where each subset represents a group of image input variables that statistically depend upon each other; means for selecting a group of candidate subsets from said plurality of candidate subsets; and means for constructing said classifier based on a ratio of a plurality of Bayesian networks using said group of candidate subsets, wherein a Bayesian network comprises a probability distribution over a set of variables where statistical independence and conditional statistical independence exist among various combinations of the variables, wherein each of said Bayesian networks is a probability distribution representation derived from dependencies among said image input variables that statistically depend upon each other. 15. The system of claim 14, wherein said means for constructing includes: means for creating a Bayesian network graph for said group of candidate subsets, wherein said Bayesian network graph is used by each of said plurality of Bayesian networks; means for representing variables in each conditional probability distribution at each node in each of said plurality of Bayesian networks by a pair of discrete valued functions; means for estimating each said conditional probability distribution within each of said plurality of Bayesian networks; and means for determining an object detection threshold for said ratio of said plurality of Bayesian networks. 16. The system of claim 15, wherein said means for creating said Bayesian network graph includes: means for forming a restricted Bayesian network for each combination of said plurality of candidate subsets; means for computing a log-likelihood ratio function for each said restricted Bayesian network; and means for evaluating a structure of each said log-likelihood ratio function using a criterion of area under an ROC (Receiver Operating Characteristic) curve. 17. The system of claim 15, wherein said means for representing variables at each node includes: means for quantizing said variables in each said conditional probability distribution within each of said plurality of Bayesian networks using Tree-Structured Vector Quantization. 18. The system of claim 14, wherein said means for selecting includes: means for creating a plurality of candidate features from said plurality of candidate subsets; means for estimating probability distributions over each of said plurality of candidate features; means for evaluating probability of each of said plurality of 2D images on each of said probability distributions; and means for choosing said group of candidate subsets to maximize an area under an ROC (Receiver Operating Characteristic) curve using said probability evaluation. 19. The system of claim 14, wherein said means for computing includes means for performing a lighting adjustment on one or more of said plurality of transform coefficients. 20. A non-transitory computer readable data storage medium having stored thereon program instructions, which, when executed by a processor, cause said processor to: compute a wavelet transform of each of a plurality of two-dimensional (2D) images, wherein each said wavelet transform generates a corresponding plurality of transform coefficients; create a plurality of candidate subsets of said transform coefficients, where each subset represents a group of image input variables that statistically depend upon each other; select a group of candidate subsets from said plurality of candidate subsets; and construct a classifier based on a ratio of a plurality of Bayesian networks using said group of candidate subsets, wherein a Bayesian network comprises a probability distribution over a set of variables where statistical independence and conditional statistical independence exist among various combinations of the variables, wherein each of said Bayesian networks is a probability distribution representation derived from dependencies among said image input variables that statistically depend upon each other. 21. A computer system, comprising: a computer having a processor; and a computer readable data storage medium having stored thereon program instructions which, when executed by said processor, configure said computer to perform the following: compute a wavelet transform of each of a plurality of two-dimensional (2D) images, wherein each said wavelet transform generates a corresponding plurality of transform coefficients; create a plurality of candidate subsets of said transform coefficients, where each subset represents a group of image input variables that statistically depend upon each other; select a group of candidate subsets from said plurality of candidate subsets; and construct a classifier based on a ratio of two Bayesian networks using said group of candidate subsets, wherein a Bayesian network comprises a probability distribution over a set of variables where statistical independence and conditional statistical independence exist among various combinations of the variables, wherein each of said Bayesian networks is a probability distribution representation derived from statistical dependencies among a set of variables.