Innovative Date Fruit Classifier Based on Scatter Wavelet and Stacking Ensemble

Abstract

Dates are essential fruits loaded with vital nutrients that keep bones healthy and prevent bone-related disorders. Approximately 8.46 million tons of different types of dates are cultivated and produced annually around the globe. There are more than 400 types of dates that are time-consuming and expensive to produce. Classifying them using conventional methods is labor-intensive, and this is one of the biggest problems for the date industry. Dataset fruit classification plays a vital role in the food industry. Dates can be classified from a luxury class to a less quality class. Accordingly, the food industry needs an automotive date fruit classifier that can work in food factories. This study proposes a pioneering method to classify date fruit that relies on extracting features from the texture of dates using Scattering Wavelet Transformation (SWT). The SWT yields in numeric coefficients were found to be immune to the deformation of invariants. This feature set trains an ensemble classifier that combines a voting mechanism to eliminate overfitting. The ensemble classifier consists of a random forest, a support vector machine classifier, and a logistic regression hyper-learner. Our novel approach was tested on two benchmarked datasets. The first data set scored F1 between 0.95 and 1.0 at the same time. The second dataset registered F1 between 0.96 and 1.0 in each of the 20 date classes. Some dates are close to each other in texture, resulting in high false positives or recall, causing a lower F1 score accuracy degree. The novelty of this approach comes from the featured representative of each date class, relying on the texture of the fruit as a discriminative feature, not on the fruit shape or color, which may not be robust enough as distinguishable features, especially in date classes that are close to each other in shape. Doi: 10.28991/HIJ-2024-05-02-010 Full Text: PDF

Keywords

• dates
• scatter wavelet transform
• stacking ensemble learning
• random forest classifier
• linear support vector machine
• performance metrics.