A Comparative Analysis of EfficientNet and MobileNet Models’ Performance on Limited Datasets: An Example of American Sign Language Alphabet Detection
Hongwen Pu, Keyi Yi
Abstract
This paper explored the performance of EfficientNet architecture and MobileNet architecture while processing sign language alphabets using small-scale datasets. EfficientNet and MobileNet are the two most popular architectures in the computer vision industry. The outbreak of the COVID-19 pandemic demonstrates the importance of investigating two model’s performance while a product needs to be developed in a short time. Previously, research has been conducted on two models, mainly focusing on the two model’s performance while handling small datasets related to medicine. However, there remains a research gap for sign language. A combined dataset obtained from two datasets from Kaggle was used to train the model. The models’ performance under 5 epochs, 10 epochs, and 20 epochs were deduced and compared. In general, the performance of MobileNetV2 models is outstanding, especially under 5 epochs, while other MobileNet and EfficientNet models show an intense overfit. Moving forward, the models could be tested on more powerful platforms and more models could be compared.