CNN Transfer Learning for Automatic Fruit Recognition for Future Class of Fruit

Abstract

Deep fruit recognition model learned on big dataset outperform fruit recognition task on difficult unconstrained fruit dataset. But in practice, we are often lack of resources to learn such a complex model, or we only have very limited training samples for a specific fruit recognition task. In this study we address the problem of adding new classes to an existing deep convolutional neural network framework. We extended our prior work for automatic fruit recognition by applying transfer learning techniques to adding new classes to existing model which was trained for 15 different kind of fruits. Pre-trained model was previously trained on a large-scale dataset of 44406 images. To add new class of fruit in our pre-trained model, we need to train a new classifier which will be trained for scratch, on the top of pre-trained model so, that we can re- purpose the feature learned previously for the dataset. Transfer learning using our pre-trained model has been demonstrated to give the best classification accuracy of 95.00%. The experimental results demonstrate that our proposed CNN framework is superior to the previous state-of-the- art networks.