Abstract According to Support Vector Machine (SVM) theory, samples located near the decision boundary typically contain more classification information than those further away. This paper explores boundary enhancement for face detection by proposing an algorithm based on SVM and an Improved Reduced Set (IRS) method. The IRS algorithm incorporates Image Euclidean Distance (IMED) into the kernel function to improve iterative approximation performance. This allows for the generation of informative virtual samples near the boundary to expand and improve the training set distribution. Experimental results on the MIT+CMU frontal face test set demonstrate that training an AdaBoost-based face detector with this enhanced set significantly improves detection performance.
Introduction
In machine learning tasks like face detection, the quality and distribution of training data are paramount. Standard datasets often lack sufficient examples near the decision boundary—the "edge cases" that distinguish a face from a near-face. SVM theory identifies these critical samples as Support Vectors (SVs). This research addresses the sample distribution problem by synthesizing high-value virtual samples to fill the gap near the decision boundary, thereby helping classifiers learn more precise boundaries.
Methodology
The proposed framework utilizes a nonlinear method to generate synthetic training examples through two primary components: the Improved Reduced Set (IRS) algorithm and the Image Euclidean Distance (IMED) metric.
Workflow
The process begins by training an initial SVM on a base dataset to identify support vectors. The IRS algorithm is then employed to generate virtual "Reduced Set Vectors" (RSVs) that approximate the decision function. By embedding IMED—which accounts for spatial correlations between pixels—into the kernel, the generated samples become more robust and representative of real-world image variations.
These virtual samples are added back into the original training set to create an "Enhanced Set," which is then used to train a final, efficient classifier such as AdaBoost.
Experimental Results
The effectiveness of the IRS-IMED approach was evaluated using the MIT+CMU frontal face test set. The researchers compared the approximation performance of different distance metrics within the IRS framework.
As shown in the approximation error curves, the IMED-RSV and TD-RSV methods reach a stable approximation of the decision boundary much faster and with lower error than the standard Euclidean Distance (ED-RSV). This improved approximation translates directly to better classifier performance. The final detector achieved a 90.73% detection rate on the MIT+CMU set with only 48 false alarms, consistently outperforming models trained on the original, unenhanced dataset.
Conclusion
This paper demonstrates that generating synthetic boundary samples using SVM and IRS with spatial distance metrics can significantly boost the generalization of face detectors. While the generation process is computationally intensive, the resulting high-quality training data allows for the creation of more robust final classifiers. Future work may explore more efficient distance measures and the application of this technique to multi-class problems.
Summary updated