Facial Expression Recognition System Using Raspberry Pi

초록

Convolutional neural network (CNN) is one of the most popular choices for image classification tasks. Due to its high computational cost, it requires careful considerations in order to make it run in real-time on mobile devices with low computational power. In this paper, we address the trade-offs between the accuracy and the network size for verifying the applicability of a CNN model on mobile devices for the task of facial expression recognition (FER) and propose two simple CNN models. Compared to AlexNet-based architectures [1], the proposed models achieve higher accuracy of 68.33 % and 67.40 % with the trained model sizes of 651MB and 25MB respectively. To verify the proposed approach, we designed a demo of a music control system that manipulates music being played according to the recognized user’s facial expression on a Raspberry Pi. Through this work, we confirm the applicability of an FER system on a lightweight device in real-time. We believe that our approach demonstrates further applicability of a variety of CNN- based computer vision algorithms to lightweight mobile devices without any dedicated processors.