MFTs-Net: A Deep Learning Approach for High Similarity Date Fruit Recognition

Abstract

—Artificial Intelligence and Deep Learning applications are well-developed as a part of human life. In the field of object recognition, Convolutional Neural Network (CNN) based methods are getting more and more important and challenging. However, existing CNN methods do not perform well on datasets that exhibit high similarities, resulting in confusion between different classes. In this study, we propose a new Deep Learning approach for recognizing date fruit categories based on the Deep Convolutional Neural Network (DCNN). The modified fine-tuning (MFTs-Net) approach can recognize with high accuracy the different date fruit categories. In order to train and to test the robustness of our proposed method, we have collected a dataset that takes into account different date fruit categories. The presented dataset is challenging as it contains classes of a unique object and presents high similarities concerning the shape, color and texture of date fruit. We show that the MFTs-Net CNN we implemented, trained and tested using the collected dataset can recognize with high accuracy the different date categories compared with state-of-the-arts works. The presented methodology works perfectly with very small datasets, which is one of the main strengths of the proposed method. Our MFTs-Net architecture performs perfectly on test data with an accuracy of 98%. 1