JIITA, vol.5 no.2, p.447-452, 2021, DOI: 10.22664/ISITA.2021.5.2.447
Myoung-Kyu Sohn*, Sang-Heon Lee, Hyunduk Kim
Division of Automotive Technology, DGIST, Daegu, Republic of Korea
Abstract: Deep learning has been applied in many areas to solve pattern recognition problems. These methods have made many advances and have shown considerable potential. In particular, it exhibits promising performance in the computer vision field such as object detection and recognition with the CNN (Convolutional Neural Network). To achieve higher accuracy, the network has been deeper and complex. Thus, the system has to process the network with the help of the GPU for inference within a reasonable amount of time. In real-world applications, many devices have some limitations such as the inability to use GPUs. In this paper, we build a deep network for face and landmark localization and demonstrate how to convert this network that works well on PC with GPU to work on a mobile platform without GPU. And, in this conversion, we propose an optimization method to enable real-time operation on mobile and show the experiment results.
Keywords: deep learning, real-time face detection; mobile application, convolution network