Investigating Efficacy of Transfer Learning for Fruit Classification

Abstract

Automated artefact identification and classification are a highly coveted field of study in a wide variety of commercial fields. While humans can easily discern objects with a high degree of multi-granular similarity, computers face a much more difficult challenge. In several deep learning technologies, transfer learning have shown efficacy in multi-level subject classification. Traditionally, current deep learning models train and test on the transformed features created by the rearmost layer. The objective of this research paper is to fabricate an automated and efficient method of fruit classification using deep learning techniques. Since the algorithm is automatic, it does not require human involvement, and the mechanism is more accurate than human-involved processes. For the classification of fruits, a pre-trained deeply trained model is fine-tuned. To distinguish fruits, we used transfer learning-trained architecture such as VGG16, InceptionV3, ResNet50, DenseNet, and InceptionResNetV2 models. The Fruits-360 dataset is used to conduct the evaluation. Extensive testing reveals that the InceptionResNetV2 outperforms in comparison to other deep learning methods.