Tomato Ripeness Detection and Harvesting Decision Making in a Greenhouse Using BIIE-YOLOv10n

Abstract

Accurate detection of tomato ripeness is crucial to improving harvesting efficiency and supporting precise picking decisions. However, occlusion and overlap of fruits, along with complex background interference, can lead to loss of local information, unstable color distribution, and increased difficulty in ripeness discrimination, thereby undermining the stability and reliability of detection. To address these challenges, this study proposes a real-time tomato ripeness detection model based on an improved YOLOv10n framework, named BIIE-YOLOv10n. The model employs an improved bidirectional feature pyramid network (IBiFPN) to achieve adaptive multi-scale feature fusion and enhanced contextual information exchange, integrates an improved iterative channel-spatial attentional feature fusion (ICSAFF) mechanism into the C2f module for effective global-local feature aggregation, and introduces the Inner-EIoU loss function to balance positive and negative samples, thereby improving bounding box regression accuracy under complex environments. Experimental results on a self-constructed tomato ripeness dataset show that the proposed model achieves an accuracy of 82.6 %, a recall of 80.5 %, an F1-score of 82.0 %, and an mAP50 of 85.4 %, representing improvements of 1.3 %, 5.2 %, 4.0 %, and 5.0 % over the baseline model, respectively. Based on the detection results, a visual-driven strategy is developed for the assessment of cluster-level ripeness and picking decision-making, providing support for automated harvesting systems in greenhouse environments. In summary, BIIE-YOLOv10n significantly enhances tomato ripeness detection performance and provides reliable decision-making support for automated harvesting and intelligent grading in greenhouse settings.