An Exploratory study of diverse models and datasets for transfer learning based image classification on sparse data

Abstract

Transfer learning is a powerful technique for image classification, especially when dealing with limited data. However, selection of the best transfer learning approach and model remains challenging, since the strategy is highly influenced by the scenario. This work conducted a thorough analysis of four transfer learning approaches for binary classification, and multiclass classification using four pretrained models i.e. MobileNetV2, MobileNetV3Large, EfficientNetB0, and EfficientNetB5. The test accuracy was measured with respect to the sample size considering the effects of data augmentation and fine tuning. The key finding of the study revealed that data augmentation has improved the test accuracy by up to 11.8%. Moreover, the lightweight models marked a breakthrough by outperforming large scale models’ accuracy for sparse data settings when fine tuning (FT) and data augmentation (DA) are applied appropriately. Hence, this work provides a comprehensive overview on selecting different transfer learning approaches for sparse data classification by exploring diverse models and datasets, while discussing the consequences and challenges for future real-world applications.