Face Detection with Improved Local Binary Patterns in CUDA March 25, 2013 Jeaff Wang jeaff.wang@utoronto.ca Richard Abrich richard.abrich@utoronto.ca Abstract As mobile computing and user interactivity become more ubiquitous, accurate and fast fa- cial detection mechanisms are necessary. And with the development of accessible parallel computing, it becomes possible to leverage the power of parallel algorithms to increase both speed and accuracy of facial detection systems. In this paper, we propose and analyse one such system based on the Improved Local Binary Pattern. 1 Introduction 1.1 Motivation With the rapid proliferation of computer tech- nology in recent decades, face detection using computer vision techniques has proven its util- ity in many applications, ranging from video surveillance to authentication and beyond. Si- multaneously, low power devices such as mo- bile phones have become increasingly impor- tant in our daily lives. Many applications on such low power devices require face detection to be processed in real-time. Processing speed has always been an impor- tant performance criterion for evaluating face detection algorithms. To illustrate, consider that real-time face detection in video sequences typically requires more than 15 face detections to be processed per second. [2] As such, a real- time face detection system capable of running on low-power platforms is necessary. With the recent popularization of general purpose GPU (GPGPU) programming, the par- allelization of algorithms for data-intensive workloads is more accessible than ever. For ex- ample, the face detection technique considered in this project requires on the order of 10 10 sim- ple arithmetic calculations, and lends itself well to parallelization. Therefore, the goal for this project is to increase the face detection process- ing speed by parallelizing the face detection al- gorithm in CUDA. 1.2 Problem Definition The face detection system considered in this project is depicted in Figure 1. Figure 1: Block diagram of face detection sys- tem The input to the face detection system is a gray-scale digital image in 8-bit unsigned inte- ger format, and the output is the position and size of a box containing a detected face, as il- lustrated in Figure 2. 1