Abstract
Fruits are considered among the most nutrient-dense cash crops around the globe. Since fruits come in different types, sizes, shapes, colors, and textures, the manual classification and disease identification of a large quantity of fruit is time-consuming and sluggish, requiring massive human intervention. We propose a multilevel fusion method for fruit disease identification and fruit classification that includes intensive fruit image pre-processing, customized image kernels for feature extraction with state-of-the-art (SOTA) deep methods, Gini-index-based controlled feature selection, and a hybrid ensemble method for identification and classification. We noticed certain limitations in the existing literature of adopting a single data source, in terms of limited data sizes, variability in fruit types, variability in quality, and variability in disease type. Therefore, we extensively aggregated and pre-processed multi-fruit data to simulate our proposed ensemble model on comprehensive datasets to cover both fruit classification and disease identification aspects. The multi-fruit imagery data contained regular and augmented images of fruits including apple, apricot, avocado, banana, cherry, fig, grape, guava, kiwi, mango, orange, peach, pear, pineapple, and strawberry. Similarly, we considered normal and augmented images of rotten fruits including beans (two categories), strawberries (seven categories), and tomatoes (three categories). For consistency, we normalized the images and designed an auto-labeling mechanism based on the existing image clusters to label inconsistent data to appropriate classes. Finally, we verified the auto-labeled data with a complete inspection to correctly assign it to the relevant classes. The proposed ensemble classifier outperforms all other classification methods, achieving 100% and 99% accuracy for fruit classification and disease identification. Further, we performed the analysis of variance (ANOVA) test to validate the statistical significance of the classifiers' outcomes at alpha = 0.05. We achieved F-values of 32.41 and 11.42 against F-critical values of 2.62 and 2.86, resulting in p-values of 0.00 (<0.05) for fruit classification and disease identification.