SGNet: Design of Optimized DCNN for Real-Time Face Detection

Abstract

This paper proposes to optimize the deep convolution neural networks for real time video processing on detecting faces and facial landmarks. For that, we have to reduce the existing weight size and duplication of weight parameters. By utilizing the strengths of the two previous powerful algorithms which have shown the best performance, we overcome the weakness of the existing methods. Instead of using the old-fashioned searching method like sliding window, we propose our grid-based one-shot detection method. Furthermore, instead of forwarding one image frame through a very deep CNN, we divide the process into 3 stages for incremental detection improvements to overcome the existing limitation of grid-based detection. After lots of experiments with different frameworks, deep learning frameworks are chosen as the best for integration of 3-stage DCNN. By using transfer learning, we can remove the unnecessary convolution layers in the existing DCNN and retrain hidden layers repeatedly and finally succeed in obtaining the best speed and accuracy which can run on the embedded platform. The performance to find small sized faces is better than YOLO v2.