Joint Integral Histogram based Adaboost for Face Detection System

Abstract

Face detection is a crucial step in many vision applications. Since the Viola and Jones face detector, many feature extraction approaches based Adaboost are proposed. This paper presents a novel approach to extract effective features for face detection system. Both LBP and Three Patch LBP (TPLBP) with joint integral histogram are used to extract features. The joint integral histogram was firstly proposed for stereo matching application. Its effectiveness has motivated us to apply it harnessing its advantages. The evaluation of the novel features based Adaboost was done using the CMU-MIT frontal face data set. Experimental results show that its performance is noteworthy especially for the earlier stages. In fact, with few numbers of the new features we can achieve the max detection (1) and reduced false positive rate (0.28).

Block-Based Sparse Integral Histogram with Interpolation

“…In cases where no reference points above and left of the block exist, exact value of these points can be determined by following: (4) Equation (4) is used when no reference point exists immediately left of the block. p l refers to a point located block width away to the left of p, p l-l refers to a point located block width away to the left of p l .…”

“…Chai [2] applied such technique to Particle filter to implement real -time user interface, and Zhang [5] improved computing time of stereo matching using joint integral histograms. It is also used other areas such as Adaboost [4] or Bilateral Filtering [6].…”