High Performance Framework for Image Scanning Techniques

Abstract

Computer vision is considered as one of the most computationally intensive problems, in specific face detection due to the wide variations. Many algorithms have been introduced ranging from trying to solve face detection and tracking problem until trying to reach better performance in terms of accuracy and processing speed. Since each algorithm has its own nature and way of operation; then each algorithm will have its average processing time. However, this processing time is also affected by the environment it is hosted on. Each GPU architecture has its own specifications, register file size, memory speed and occupancy ratio. The researchers choose Haar-like and linear binary pattern (LBP) algorithms for face detection due to the similarity between both algorithms. The study proposed a framework that hosts different algorithms, and different approaches for each algorithm. The results of deployment of those algorithms are studied for different GPU architectures. The researchers also focused on the obstacles and the techniques to resolve them based on each GPU architecture.