A comparative study of a combinatorial machine learning approach to face detection using a very small training dataset

In recent years, machine learning algorithms have improved the prediction rates in object detection task, but with high computational cost in the training stage. Haar cascade classifier is a cheap method widely used in object detection, but its resulting predictions contain a significant number of false positives when the model was not trained with a large dataset. In this work, we propose a method to detect faces in images by using combinatorial widely used machine learning algorithms. The main goal of our approach was to reach acceptable prediction rates with models trained with a very small training dataset. In this way, we present a practical implementation with a statistical comparison between different prediction models. The models were tested with a fixed dataset and then compared by using standard evaluation metrics. Furthermore, the challenging Face Detection Data Set and Benchmark (FDDB) was used for performance evaluation. The experimental results showed that our proposed method reach similar prediction rates than some the state-of-the-art methods, even with better false positive rate, once trained with a dataset 97.20% smaller.