Face Detection Based On AdaBoost Algorithm And Its Transplantation On DSP Platform

Face detection is a classical problem in the field of computer vision and pattern recognition, face detection technology has been a research focus in the field. With the rapid development of computer vision technology, the research and application in face detection technology has also made considerable progress. At the same time, computer technology and semiconductor manufacturing technology have been developed rapidly, a variety of high-performance data-processing chips emerge continuously. The effective combination of the two aspects lays the foundation for face detection in embedded systems. The face detection on embedded platform only by meeting the real-time targets can have a strong practicality. Therefore, the real-time face detection algorithm on embedded platform is the prerequisites that push the human face detection into utility.This thesis begins with describing the development and the current research situation of face detection algorithm, and then starting from the basic AdaBoost face detection algorithm, improves and optimizes the process of training and face detection respectively, thereby reduces the algorithm complexity and volume of computation effectively, getting ready for the transplant of algorithm on the embedded platform. In the process of algorithm training, using the method by restrictions on Haar-like features of height, width and area, the number of features is reduced effectively, thereby reducing the training time; in the detection process, namely in the implementation of different sizes face detection, the thesis analyzes and compares the advantages and disadvantages of weak classifier magnifying and image pyramid sampling. Finally, it proposes a method that combined with the two mechanisms implementing in multi-scale face detection.Next, it transplants the algorithm implemented on a PC platform to DM642 platform, and optimizes the algorithm's performance on the DSP platform. This part mainly uses a lot of optimization methods such as floating-point fixed-point-based, memory optimization and optimization of a linear assembly etc. The efficiency of algorithm processing CIF (352×288) images before optimization is 6 frmaes per second, and the efficiency of the algorithm optimized is 30 frmaes per second. The efficiency after the algorithm being optimized is much greater than 25 frames per second that is real-time targets, ultimately realizing the real-time detection of human face detection on the DSP platform. Thus, it achieves a real-time face detection algorithm on embedded platform.