Distribution-Based Face Detection using Calibrated Boosted Cascade Classifier

This study presents a distribution-based face detection. The proposed face detection integrates information over image space and scale because the face can be recognized even if a small warp is applied. We must calibrate raw classifier outputs for information integration. Two salient contributions of this paper are calibration for a boosted cascade classifier and a distribution-based face detection. We first propose a new calibration method that converts the raw classifier outputs into posterior probabilities. This calibration method enables us to obtain a "Face Likelihood Distribution "from an input scene. The second contribution is the distribution-based face detection, which evaluates the local face likelihood distribution over space and scale. In contrast, existing methods evaluate the face similarity value only at that point, which causes misdetection. Experimental results using 170 input scenes, which might or might not include faces, show that the proposed method improves the detection rate by about 5%.

1. Introduction

Face detection is a key technique in computer vision. It serves an important role in widely various applications such as surveillance systems, video tracking applications and biometrics [9].