Comparison of different Convolutional Neural Network models on Fruit 360 Dataset

Abstract

Numerous Convolutional Neural Networks emerged in the past decade, each varies in accuracy, speed, and architecture. From AlexNet to ResNet, CNN models have been developing rapidly, and the architecture of the models become more complicated. These models are known for their accuracy on ImageNet, so the topic of this research is to explore how CNN models can perform differently on the Fruit 360 dataset. A model constructed specifically in this research and three significant models developed in the past decade are applied to the Fruit 360 dataset for result comparison: VGG-16, ResNet-50, MobileNet, and SC-3. The models are selected to be different in architecture for comparison. The models are modified in similar ways and are trained on the same device under the same circumstances. The research found that ResNet-50 obtains the best prediction accuracy. MobileNet takes the least time to train and predict. The ResNet-50 and MobileNet models can be considered for real-world application, whereas VGG-16 and SC-3 are either inefficient or fails to generalize to complicated situations.