Discriminant analysis and adaptive wavelet feature selection for statistical object detection

We utilize the discriminant analysis to select wavelet features for efficient object detection. The analysis applies to the Bayesian classifier and is extended to the case of boosting. Based on the error analysis under the Bayesian decision rule, we reduce the number of coefficients involved in detection to lower the computational cost. Using a hidden Markov tree model to describe the pattern distributions, we introduce the concept of error-bound-tree to relate feature selection to error reduction. The scheme selects discriminative features that are adaptive to the pattern and allows the detector to reach a decision faster.