TB-YOLOv8n: an improved YOLOv8n-based model for tip-burn detection in plant factory-grown pakchoi under LED lighting

Résumé

Introduction Plant factory-grown pakchoi is prone to tip-burn under high-density cultivation, limited airflow, and rapid growth conditions, which reduces leaf integrity and commercial quality. Image-based detection of small or mild visible tip-burn symptoms remains challenging because the affected regions are usually small, irregularly shaped, low in color contrast, and easily disturbed by red-blue LED lighting, leaf overlap, curled leaf margins, and heart-leaf occlusion. Methods To address these challenges, this study constructed a pakchoi tip-burn image dataset under plant factory conditions and proposed TB-YOLOv8n, a lightweight detection model based on YOLOv8n. AFGCAttention was introduced at the end of the backbone to enhance informative channel responses related to subtle tip-burn regions and suppress redundant background information. GLSA was embedded between the backbone and neck to integrate local texture details with global spatial context under leaf overlap and low-contrast symptom conditions. In addition, a P2-enhanced BiFPN neck was constructed to strengthen cross-scale feature fusion and preserve fine spatial information for small visible tip-burn regions. Results Experimental results showed that TB-YOLOv8n achieved a precision of 89.4%, a recall of 87.6%, an mAP@0.5 of 93.2% ± 0.3%, and an mAP@0.5–0.95 of 59.0% ± 0.3%, which were 4.2, 3.9, 3.4, and 10.0 percentage points higher than those of the original YOLOv8n, respectively. The model contained only 2.2 M parameters, required 7.6 G FLOPs, had a model size of 4.8 MB, and achieved a model-forward inference speed of 138 FPS on the RTX 4090 and 33.0 FPS on the Jetson AGX Orin. Ablation experiments, heatmap visualization, and comparisons with mainstream detection models further confirmed the effectiveness of the proposed improvements. Discussion The results indicate that TB-YOLOv8n achieves a favorable balance among detection accuracy, lightweight structure, and real-time inference capability. The proposed model provides a basis for visible tip-burn monitoring and subsequent production decision support in plant factories.

Citer ce document

Accès au document

Texte intégral en lecture en ligne, réservé aux abonnés SPHAERO et aux membres de l'institution. Se connecter

Auteur(s)

Statistiques

Consultations : 1

Téléchargements : 0