Rapid Detection and Identification of Lightweight String Tomato Based on the Modified YOLOv8

Due to the challenges posed by factors such as light intensity, leaf occlusion, and fruit overlap in greenhouse environments, the accuracy of cherry tomato detection is often compromised, and computational resource consumption is high. To address these issues, this study proposes a lightweight and efficient cluster tomato detection model, YOLOv8s-CGBI, based on an improved version of YOLOv8.version of YOLOv8. To reduce the computational burden, the PConv module is used to replace the Bottleneck in the C2f module. To enhance the model’s ability to extract key features, a Global Attention Module (GAM) is incorporated with the three C2f modules that connect the Neck network to the Detect module. To improve detection accuracy in complex environments, a BiFPN (Bidirectional Feature Pyramid Network) replaces the Concat module in the Neck network. Additionally, the Inner-EIoU loss function is introduced to accelerate convergence and improve detection performance.Experimental results demonstrate that the YOLOv8s-CGBI model achieves a recognition accuracy of 98%, an inference speed of 140.85 FPS, and a computational load of 23.8 GFLOPS. Compared with YOLOv8s, the accuracy, recall, mAP@.5, mAP@.5:.95, and F1-score have been improved by 2.0, 2.9, 0.1, 1.7, and 2.0 percentage points, respectively, the volume has been reduced by 17.8%, and the amount of parameters has been reduced by 18.1%; Compared with FasterR-CNN, SSD, YOLOv5s, YOLOv7, and YOLOv8s, mAP@.5:.95 improves 12.8, 8.6, 0.3, 3.5, and 1.7 percentage points, respectively. This method provides a model reference for the lightweight recognition research of cluster tomatoes.