Efficient-tomato: a high-precision and lightweight method for cherry tomato ripeness and disease detection

Accurate detection of cherry tomatoes at different ripeness and disease stages is important for automated harvesting and crop-quality monitoring, but illumination variation, fruit overlap, leaf occlusion, complex backgrounds, and small target sizes remain challenging. This study proposes Efficient-Tomato, a lightweight detector developed from YOLOv8s through the coordinated use of EfficientViT, BiFormer, CARAFE, and WIoUv1. EfficientViT serves as the backbone for lightweight multiscale feature extraction. BiFormer is inserted into the neck to enhance informative spatial regions, while CARAFE and lightweight neck-channel adaptation improve content-aware feature reconstruction. WIoUv1 replaces the CIoU component of the bounding-box regression loss, with DFL retained. The model was evaluated on a consolidated dataset of 1,583 original images containing six categories: disease, half-ripe, overripe, ripe, rotten, and unripe. The original images were divided into training, validation, and test subsets before augmentation, and only the training subset was augmented. Each model configuration was trained independently three times, and arithmetic means were reported. Compared with YOLOv8s, Efficient-Tomato increased precision from 89.5% to 97.5%, recall from 87.3% to 96.9%, mAP@0.5 from 88.5% to 96.7%, and mAP@0.5:0.95 from 81.6% to 87.9%. The F1 score calculated from the tabulated mean precision and recall increased from 88.4% to 97.2%. FPS increased from 150.2 to 172.5, while the parameter count decreased from 22.5 M to 14.9 M. These results indicate that the coordinated design improves detection performance while reducing model complexity. Because inference speed was measured on an NVIDIA V100 GPU, further validation on embedded devices and harvesting robots is required before field deployment.