YOLOv8-Based Visual Detection System for Taro Peeling State Assessment

摘要

Taro peeling quality affects the consistency of downstream washing, cutting, sorting, and prepared-food processing. Manual inspection is simple but difficult to scale because taro roots have irregular shapes, mottled brown skin, moist surfaces, and variable illumination during handling. This paper presents a YOLOv8-based visual detection system for assessing taro peeling state. A custom dataset of 1246 images was collected from web images and local photography, annotated into three states: completely unpeeled, partial peeling, and full peel. The dataset was divided into 873 training images, 187 validation images, and 186 test images. A YOLOv8n detector was trained for 100 epochs on a workstation equipped with an NVIDIA GeForce RTX 4050 Laptop GPU. Experimental results from the project records show strong convergence and high precision-recall performance, with validation mAP@0.5 of 0.992 and combined test/validation mAP@0.5 of 0.988. The partial-peeling category is the most challenging class, which is consistent with its boundary ambiguity between remaining skin and exposed flesh. A PySide6/PyQt6 graphical interface and a conceptual mechanical inspection line are also described to connect the detector with practical image and video inspection workflows. The study demonstrates the feasibility of lightweight deep-learning detection for taro peeling assessment while identifying dataset diversity, timing measurement, and production-line validation as necessary next steps.