AUTOMATED POTATO DEFECT DETECTION AND SORTING USING CONVOLUTIONAL NEURAL NETWORKS

Abstract

This project presents the development of an automated system for potato defect detection and sorting, leveraging real-time deep learning techniques. A YOLOv8 object detection model, deployed on a Raspberry Pi 4, is used to classify potatoes as ”good” or ”bad” from images captured by a Logitech C920 Pro HD webcam positioned above a black conveyor belt. Upon detection, a mechanical sliding plate actuated by a NEMA 17 stepper motor removes defective potatoes from the conveyor. To enhance model robustness, an extensive dataset was collected under real-world conditions, incorporating manual data augmentation strategies. The hardware design integrates a custom-built conveyor, optimized lighting conditions, and a gentle sorting mechanism to minimize produce damage. The system demonstrates a scalable, efficient, and adaptable framework for agricultural automation, offering significant potential for broader applications in produce sorting.