Analysis of the data influence on the training of Haar cascades for face detection
Elizaveta Rudinskaya, Парингер Рустам Александрович · 2021 International Conference on Information Technology and Nanotechnology (ITNT) · 2021
This paper is devoted to the problem of the sampling influence on the result of the detector, which is based on the Haar cascades. Trained Haar cascades give more stable detection results. On the given data sets, the spread of correct detection decreased by an average of 1.55 times. Based on the results obtained the criteria for selecting the training sample are formulated. The best results for training were obtained: 28-29 stages of the cascade training, with the maximum false positive rate being 0.5-0.51, the minimum hit rate for each stage being 0.999.