Design Methodology for Single-Channel CNN-Based FER Systems

Abstract

Facial Expression Recognition (FER) systems classify emotions by using geometrical approaches or Machine Learning (ML) algorithms such as Convolutional Neural Networks (CNNs). Due to their complexity, these FER systems need to be implemented on high-performance hardware, which makes them unsuitable for embedded devices. To address this challenge, we propose a methodology for the design of low-complexity, CNN-based FER systems. Our methodology includes data preprocessing, Local Binary Pattern (LBP) implementation, Data Augmentation (DA), and CNN design. Here, we also introduce the Model M6, a single-channel CNN that reaches an accuracy of 94% in less than 30 epochs. M6 has 306,182 parameters that correspond to 1.17 MB of memory. Therefore, our methodology and M6 model are feasible for implementation onto embedded systems capable of computing floating point operations. We validated our methodology and M6 model using 66 tests with 6 CNN models and 4 training parameters (batch size, learning rate, number of epochs, optimizer). This validation was performed using the Japanese Female Facial Expression (JAFFE) dataset and TensorFlow. In each test, the relationship between parameters, layers, overfitting, and underfitting was studied. Moreover, we present a step-by-step guideline on how to design the single-channel CNN and provide open-source code for readers interested in reproducing our work.