Haar-features training parameters analysis in boosting based machine learning for improved face detection

Haar features have been used in most of the works in literature as key components in the task of object as well as face detection. Training of Haar features is an important step in the development of overall machine learning based face detection system. In this work, we have done investigation in the variations of a number of training parameters during AdaBoost based machine learning of Haar features with respect to the size of training images. A number of observations have been drawn based on the variations of these parameters during the training process. Training parameters, true detection rate and figure of merit have been used as weighing parameter. These parameters have been used in the formation of detection cascade for the improvement in the original AdaBoost framework used for machine learning of Haar features. Statistical analysis based on variations of training parameters has been done for selection of efficient learners from a large pool of available features which further are cascaded to make the strong classifiers. We have been able to achieve the maximum percentage reduction in training time of 47.20 corresponding to training image's size of 25×25 pixels with the help of statistical screening of Haar features prior to AdaBoost. The proposed system has been developed using training dataset from Center for Biological & Computational Learning (CBCL) and successfully tested on facial images from datasets. The datasets considered were WIDER FACE Detection Benchmark, Annotated Faces in the Wild (AFW) and Face Detection Dataset and Benchmark (FDDB). The results achieved were promising in terms of training time computed in the range of 0.31-2.83hrs. The maximum detection rate is 96.21% and minimum detection time is 16.95 ms.