A weight adjustment strategy to prevent cascade of boosted classifiers from overfitting

Abstract

We propose a weight adjustment strategy to prevent a cascade of boosted classifiers from overfitting and to achieve an improved performance. In cascade learning, overfitting often occurs due to the iterative applications of bootstrapping. Since false positives that the previous classifier misclassifies are collected as negative examples through bootstrapping, negative examples more similar to positive examples are prepared as stages go on, and thus classifiers become tuned to the positive examples. When overfitting occurs, the classifier cascade shows performance degradation more in the detection rate than in the false alarm rate. In the proposed strategy, the imbalance between the detection rate and the false alarm rate is evaluated by computing the weight ratio of positive examples to negative examples and it is compensated by adjusting the weight ratio prior to boosting at each stage. Experimental results confirm the effectiveness of the proposed strategy. For experiments, face and pedestrian classifier cascades were trained by employing previous approaches and the proposed strategy. By employing the proposed strategy, the detection rate of classifier cascades was significantly improved for both face and pedestrian.