Abstract:
We proposed a lightweight neural network architecture model that focuses on efficient computation when doing face detection
in real-time with limited system resources, in image processing time to be exact. The concept is to implement layers and neurons as
minimum as possible to reduce processing time and computing resource used in either convolutional layer or fully connected layer. This
research is conducted because current neural network technology does not consider real-time detection scenario needs, such as tracking
an object using camera. The result of proposed neural network implementation is an application that captures video from camera and
generates boundary box that contains human face, with 0.117 seconds processing time each data, 96.735% accuracy, and 0.1219 error
rate. The proposed model used 169.3 MB RAM and taking up to 1.186% CPU processing. Although does not have best accuracy and
error rate, the proposed method does have faster processing time and lower usage of system resources in human face detection. This
will allow computers with lower specification to use this model for face detection, especially in human tracking. This research also
provides the concept of convolutional neural network, object tracking, and face detection that would be the base of create the proposed
lightweight neural network.