Enhanced face detection performance based on multi-block local binary patterns and Dual - Threshold Haar Features

Abstract

In the field of computer vision, face detection is particularly important and in widespread use. Face detection is a specific case of object detection and is used as the primary module in face and gesture recognition systems, face tracking and more. Due to its potential importance, it is a hot topic in computer vision and is under extensive research with many proposed approaches and their variants that have shown increasingly better performance. The main issues that are addressed are detection ability and computational density which limit usability. The most notable advance in this area was introduced in 2001 by the influential work of Viola and Jones and is the basis of the work in this thesis. Many variants of the Viola and Jones framework have been proposed, most notably the variant that uses Multi-Block Local Binary Pattern (MB-LBP) features has shown the best results and is used in real applications. However, strengths and weaknesses are identified in this variant which has prompted for its adoption but with improvement. In this thesis, a comprehensive analysis is performed in a theoretical manner aided by runtime characteristics of the selected variant, to reveal weak points which are then collectively addressed and evaluated. This work will introduce a new feature called Dual-Threshold Haar Feature (DT-HF), which helps extract more information from the image and improves the discrimination ability of the detector, also resulting in much lower computational density. The DT-HFs exhibit best performance when used as a pre-detector designed to reject most non-face image regions quickly before being passed to the MB-LBPs for further analysis. DT-HFs enhance the discrimination power of the detector by significantly decreasing false positives. The work used the BioID and Caltech face detection benchmarking databases to reveal its performance characteristics and for comparison to the other variants. Quantitative evaluation showed favorable results of the proposed face detector with 12% higher detection rate over MB-LBP features, at a given false alarm of 15%. Moreover, the proposed detector processed the BioID face detection database in 11 seconds (128fps), gaining ×3 speedup over using MB-LBP alone. The testing was run on a single core of a 4th gen Intel mobile Core i7 processor.