An empirical evaluation of feature extraction for Vietnamese fruit classification

Abstract

In recent years, Vietnamese fruit production has marked significant progress in terms of scale and product structure. Viet Nam enjoys suitable climates for tropic, subtropical fruits, and some temperate fruits. Thanks to the diversified ecology, implementing an automatic classification system has received a significant concern. In this paper, we address the problem of Vietnamese fruit classification. However, the first requirement to explore this problem is the qualified dataset. To this end, we first introduced the UIT-VinaFruit20 dataset, a novel Vietnamese fruit image dataset that includes 63,541 images from 20 types of fruits from three regions of Viet Nam. The diversity in different fruits in distinct areas poses many new challenges. In addition, we further leverage the feature extraction from hand-crafted and deep learning features along with the SVM classification model to effectively classify Vietnamese fruits. The extensive experiments conducted on the UIT-VinaFruit20dataset provide a comprehensive evaluation and insightful analysis. An encouraging empirical success was obtained as the EfficientNetB0 feature achieved the best results of 85.465 % and 86.919 % in terms of Accuracy and Macro F1-score, respectively