Tobacco-visorts: an intelligent vision-based sorting framework for automated tobacco leaf classification using yolov8 and starfish optimization

Abstract

Accurate tobacco leaf grading is critical for product quality and commercial value in the tobacco industry, yet traditional manual and semi-automated sorting methods suffer from inconsistency, subjectivity, and low efficiency. This paper presents Tobacco-VISORTS, an intelligent vision-based sorting framework that integrates advanced image enhancement, deep learning detection, metaheuristic optimization, and simulated airflow-driven actuation for automated tobacco leaf classification. The system employs Contrast Limited Adaptive Histogram Equalization (CLAHE) to enhance local contrast and preserve fine textural and color features under varying illumination conditions. YOLOv8 serves as the core detection and segmentation model, leveraging its C2f modules, PAN-FPN neck structure, and CIoU loss for precise real-time grading based on shape, texture, and color attributes. To eliminate manual hyperparameter tuning, the Starfish Optimization Algorithm (SFO) automatically optimizes critical parameters including learning rate, IoU threshold, and confidence score using a multi-objective fitness criterion. A Python-controlled virtual air-jet actuator simulates non-contact sorting with a theoretical response delay of 0.257 seconds. Experimental results on a benchmark tobacco dataset demonstrate that Tobacco-VISORTS achieves superior performance with an F1-score of 0.93, mAP@50 of 0.91, and AUC exceeding 0.93 across all grades, significantly outperforming conventional PLC-based systems and unoptimized deep learning models. The framework offers a scalable, real-time solution for intelligent tobacco processing.

Bangladesh J. Bot. 55(3): 441-449, 2026 (September)