Automatic Differentiating of Postharvest Banana Fruits with High Traits Using Imagery Data

Abstract

Visually differentiating banana cultivar with high similarity in shape, color and peel texture requires skill and experience during harvesting to reduce mistake on identifying cultivar. This study aims to identify automatically some similar banana cultivars using banana finger imagery and computer vision. The identification process was carried out to distinguish two groups of bananas with high similarities, namely group 1 (Ambon, Hijau, Goroho) and group 2 (Barlin, Mas). The test was conducted on the pair of datasets of unripe Ambon-Hijau-Goroho, ripe Hijau-Goroho, ripe and unripe Barlin-Mas. Testing was done to determine the performance of identification and to find out the most effective characteristics that could be used as cultivar identification. Results of classification using extreme learning machine (ELM) showed that texture features extracted from local binary pattern (LBP) could accurately distinguish unripe Ambon-Goroho, unripe Goroho-Hijau, ripe Goroho-Hijau with 100% accuracy. While unripe Ambon-Hijau, unripe Barlin-Mas and ripe Barlin-Mas could be optimally distinguished using a combination of shape and peel texture features with accuracy of 93.39%, 89.68%, 99.31% respectively. This result indicated that the proposed method could be used as an alternative of automatic banana sortation during post-harvest. The use of shape and peel texture features had shown effectively differentiating these high similarity banana cultivars.