An efficient and data-driven algorithm to evaluate facial recognition

Is this project an undergraduate, graduate, or faculty project?

Undergraduate

Project Type

group

Campus

Daytona Beach

Authors' Class Standing

Adam Kuzmicki, Senior Kaitlyn Cavanaugh, Senior Brady Heddon Kristiyan Stefanov

Lead Presenter's Name

Adam Kuzmicki

Lead Presenter's College

DB College of Arts and Sciences

Faculty Mentor Name

Mututhanthrige Perera, Sirani K.

Abstract

Recognition and classification tasks have become increasingly popular for automation in several fields. These tasks are commonly carried out using convolutional neural networks (CNNs) and feedforward neural networks (FFNNs). Their adaptability and feature extraction lead to high-accuracy image recognition results; despite being computationally expensive. However, high computational demands, large volumes of labeled data, and storage requirements all hinder CNN efficiency. In this paper, we first present a computationally efficient image recognition algorithm using a neural network based on the low-complexity Discrete Cosine Transform algorithm. Once the network is designed, we compare the accuracy and precision of the network architecture using a conventional CNN and Eigenfaces. At the end, we share image classification results utilizing the Forward-forward (FF) algorithm, which helps the network to learn as it propagates. We sum up by presenting the testing and training errors of the FF and comparing them with the FFNN results for image classification.

Did this research project receive funding support (Spark, SURF, Research Abroad, Student Internal Grants, Collaborative, Climbing, or Ignite Grants) from the Office of Undergraduate Research?

Yes, Spark Grant

Share

COinS
 

An efficient and data-driven algorithm to evaluate facial recognition

Recognition and classification tasks have become increasingly popular for automation in several fields. These tasks are commonly carried out using convolutional neural networks (CNNs) and feedforward neural networks (FFNNs). Their adaptability and feature extraction lead to high-accuracy image recognition results; despite being computationally expensive. However, high computational demands, large volumes of labeled data, and storage requirements all hinder CNN efficiency. In this paper, we first present a computationally efficient image recognition algorithm using a neural network based on the low-complexity Discrete Cosine Transform algorithm. Once the network is designed, we compare the accuracy and precision of the network architecture using a conventional CNN and Eigenfaces. At the end, we share image classification results utilizing the Forward-forward (FF) algorithm, which helps the network to learn as it propagates. We sum up by presenting the testing and training errors of the FF and comparing them with the FFNN results for image classification.