DTL-YOLO: a lightweight model for real-time tobacco leaf damage detection and mechanical harvesting quality assessment

Tobacco is an important industrial crop worldwide. Leaf damage caused by mechanical harvesting severely reduces tobacco leaf quality and commercial value, hindering the promotion of mechanized tobacco production. Real-time and accurate detection of mechanical damage in fresh tobacco leaves is essential for harvesting quality assessment. Nevertheless, current industrial evaluation relies largely on post-harvest manual sampling and statistics, which suffers from high labor costs, low efficiency and strong subjective bias. Automated and real-time approaches for damage detection during tobacco harvesting remain insufficiently investigated. To address this challenge, this study proposes a novel lightweight tobacco leaf damage detection model, termed DTL-YOLO (Damaged Tobacco Leaf-YOLO), based on an improved YOLOv8 architecture. Fresh tobacco leaves harvested by a layered tobacco harvester were used as the research object to enable damage detection under realistic harvesting conditions. The proposed model adopts StarNet as a lightweight backbone and integrates an improved Bidirectional Feature Pyramid Network (BiFPN) for feature fusion, while an adaptive downsampling module (ADown) is incorporated to enhance boundary perception of damaged leaf regions. With only 5.4 GFLOPs and 1.17M parameters, the proposed method attains 80.9% precision, 81.0% recall, and mean average precision of 88.4% in the evaluation. Compared with the YOLOv8n baseline, the proposed model improves recall by 1.9 percentage points and mAP₅₀ by 1.8 percentage points while effectively reducing computational complexity and model size. This approach enables real-time detection of damaged fresh tobacco leaves under mechanical harvesting conditions, thereby facilitating harvest-quality assessment and supporting the design of low-damage layered tobacco harvesters.