Acceleration of real-time face recognition pipeline on heterogeneous hardware platforms

Convolutional neural network;Deep neural network;Fast Fourier-transform;Winograd;High-level synthesis;Field-programming gate arrays (FPGA)

Zhuge, Chuanhao ; Chen ; Deming

关键词: Convolutional neural network;    Deep neural network;    Fast Fourier-transform;    Winograd;    High-level synthesis;    Field-programming gate arrays (FPGA);   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/98312/ZHUGE-THESIS-2017.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

In recent years, advancements in machine learning techniques, and specifically, deep learning methods, have started to create a great impact in the world. With the advent of deep neural network, we are able to achieve unprecedented results in previously unsolvable computer vision tasks. Face recognition, one of the critical computer vision tasks, also sees breakthrough in terms of accuracy. This thesis presents an accelerated and optimized end-to-end face recognition pipeline. Such a pipeline consists of three stages: face detection, alignment, and face recognition/verification. Algorithms for these jobs are extremely computation intensive and thus real-time application was not attainable. In order to bring about the goal of high definition real-time multi-face recognition, we leverage different types of hardware to accelerate detection and recognition stages, which are the most time-consuming stages of the recognition pipeline. To achieve this goal, we leverage an embedded Graphic Processing Unit (GPU) platform as the front end, to perform video capture and face detection. For the back end, we employ a powerful Field-Programming Gate Arrays (FPGA) equipped server, which runs a state-of-the-art deep neural network to recognize faces streamed from the front end with low latency. With the two acceleration schemes targeting GPUs and FPGAs, respectively, we are able to achieve real-time performance for the overall task, and such face recognition system can be widely adopted for various applications.