<![CDATA[Indonesia, as an agrarian country, has a wide variety of horticultural commodities, one of which is mandarin orange (Citrus reticulata). Post-harvest handling, particularly the sorting process based on fruit ripeness and defects, plays an important role in maintaining product quality and market value. However, manual sorting is considered inefficient because it is repetitive, highly dependent on operator subjectivity, and prone to inconsistency. Several studies report those manual methods can result in classification error rates exceeding 20% and longer processing times compared to computer vision-based systems. This study develops an automatic citrus fruit quality classification system using the YOLOv5 algorithm. The dataset consists of 703 citrus fruit images captured directly using a webcam under varying lighting intensities and color conditions, and is divided into 80% training data and 20% testing data. The classification is performed into three quality categories: ripe, unripe (green), and rotten oranges, based on the visual characteristics of the fruit peel. Experimental results show that a training configuration with 300 epochs, a batch size of 40, and warm white bright lighting conditions achieves the best performance. Real-time testing on 15 citrus fruits yields an average accuracy of 78.2%, indicating the potential of the proposed system as an initial sorting aid, despite limitations related to lighting conditions and the amount of test data.]]>
Copyrights © 2026
