Date of Award

Spring 2023

Document Type

Open Access Thesis

Department

Computer Science and Engineering

First Advisor

Ramtin Zand

Abstract

Facial expression recognition is a popular and challenging area of research in machine learning applications. Facial expressions are critical to human communication and allow us to convey complex thoughts and emotions beyond spoken language. The complexity of facial expressions creates a difficult problem for computer vision systems, especially edge computing systems. Current Deep Learning (DL) methods rely on large-scale Convolutional Neural Networks (CNN) which require millions of floating point operations (FLOPS) to accomplish similar image classification tasks. However, on edge and IoT devices, large-scale convolutional models can cause problems due to memory and power limitations. The intent of this work is to propose a neural network architecture inspired by deep CNNs which is tuned for deployment on edge devices and small-form-factor edge AI accelerators. This will be carried out by strategically reducing the size of the network while still achieving good discrimination between classes. Additionally, performance metrics such as latency, accuracy, throughput, and power consumption will be captured and compared with several popular deep CNN models. It is expected that there will be trade-offs between network size and performance when the model is deployed and running model inference on edge AI accelerators such as the Intel Movidius Neural Compute Stick II and the NVIDIA Jetson Nano GPU accelerator. An additional benefit of smaller-scale convolutional models is that they are better suited to be converted into spiking neural networks and deployed on neuromorphic hardware such as the Intel Loihi neuromorphic chip. Furthermore, this work will also examine various image processing techniques across multiple datasets in an effort to increase the performance of the edge-efficient model.

Rights

© 2023, Mark Heath Smith

Download

Share

COinS