Related Experiment Videos
A novel feature selection methodology for automated inspection systems.
Hugo C Garcia1, Jesus Rene Villalobos, Rong Pan
1L3, Electro-Optical Systems, Tempe, AZ 85281, USA. Hugo.Garcia@L-3com.com
Summary
This study introduces a novel feature selection method using misclassification error estimation, reducing computational time for classification algorithms. It offers direct assessment of feature benefits, improving inspection and classification accuracy.
Area of Science:
- Machine Learning
- Statistical Pattern Recognition
Background:
- Traditional feature selection relies on discriminant metrics like Wilks' Lambda.
- Estimating classification error often requires computationally intensive methods like simulation or cross-validation.
Purpose of the Study:
- To propose a new feature selection methodology using misclassification error estimation.
- To improve the efficiency and directness of feature selection in classification algorithms.
Main Methods:
- Utilizes a stepwise variable selection procedure.
- Employs an estimation of the misclassification error rate (MER) as the primary metric.
- Derives MER using densities of a constructed function representing the conditional distribution of the quadratic discriminant function estimate.
Main Results:
- Achieves significant savings in computational time for classification error estimation compared to traditional methods.
- Provides a direct estimation of expected misclassification error during the feature selection process.
Conclusions:
- The proposed methodology offers an efficient alternative for feature selection.
- Enables immediate assessment of the value of additional features in classification tasks.