Performance Analysis of MobileNet: Effects of Dataset Size, Class Balancing, and Data Split Ratios

MobileNet’s Analysis using Dataset Size, Class Balancing, and Data Split Ratios

Abstract

Image classification plays an important role in computer vision and has a variety of real-world applications. In order to perform an image classification task efficiently, among the most popular lightweight Convolutional Neural Network models is the MobileNet, which is optimized to be utilized in resource-constrained devices. In this paper, we assess the binary image classification task using the MobileNet architecture, with consideration of the three experimental factors, including the effects of dataset size, dataset balance and the train-test split ratio. We tested the MobileNet model with different size ratios of the Cats and Dogs dataset, such as 25%, 50%, 75%, and 100% with balanced and unbalanced data partitions. The findings indicate that the size and balance of the dataset have a significant impact on the classification accuracy, and results on the balanced datasets consistently outperform the unbalanced datasets. In addition, the MobileNet model performed well in the experiment with multiple train-test splits, including ratios of 50%:50%, 60%:40%, 70%:30%, 80%:20%, and 90%:10%, especially when using a 70% training and 30% testing split.