Recognition of Tomato Fruit Maturity and Fruit Stems Based on Improved YOLOv8

In the complex environment of the tomato greenhouse, many branches may cast a shadow over tomato fruits, and fruit overlap can occur. Traditional models have demonstrated insufficient accuracy in identifying the ripening stage of tomatoes and detecting small targets among fruit stems; to alleviate this issue, in this study, we propose a tomato ripeness identification and fruit stem detection model: CEVM-YOLOv8. This model enhances the capture of small targets among fruit stems by introducing the EfficientVMamba module to replace CSPDarknet53 within YOLOv8's backbone network. The CBAM attention mechanism is then introduced to enhance the model's ability to extract features from tomatoes at different stages of ripeness. Finally, the MPDIoU loss function is used to optimize the prediction framework's localization accuracy and convergence speed. The experimental results show that the CEVM-YOLOv8 model improves the mean accuracy (mAP) (at a 0.5 threshold), precision, recall, and F1 score with which tomatoes of different ripeness are identified in complex environments; these factors improved by 1.7%, 4.7%, 2.3%, and 3.6%, respectively, relative to the original YOLOv8n model. The detection of fruit stems was improved by 5.9%, 4.3%, 12.3%, and 8.4%. The CEVM-YOLOv8 improved the recognition of all labels from a combined 2.8%, 4.6%, 4.9%, and 4.8% to 94.3%, 89.5%, 92.5%, and 91.0%. This study demonstrates the ability of the CEVM-YOLOv8 model to recognize tomatoes of different ripeness levels and to accurately identify fruit stems within complex environments, thereby providing a more effective method for automatic tomato picking.