YOLO-LSBA: A high-precision model for detecting stems of small-sized cherry tomatoes.

The diversity in fruit posture has become the key factor that limits improvements in stem recognition precision during cherry tomato harvesting. To effectively enhance the recognition of small target features in cherry tomato stems, data augmentation strategies are employed to expand the dataset selectively, improving the model’s adaptability to complex scenarios. First, based on the YOLO11n model, the Large Separable Kernel Attention (LSKA) mechanism is integrated into the Spatial Pyramid Pooling-Fast (SPPF) to construct the SPPL module. This design effectively improves detection accuracy and model robustness while expanding the receptive field and enhancing feature extraction capabilities. This reduces computational complexity and enhances model robustness. Second, the Spatial and Channel Reconstruction Convolution (ScConv) module is embedded into the Bottleneck architecture to replace the original C3K2 module, thereby reducing feature redundancy and improving the extraction of fine-grained features. Finally, the BAFPN module was designed, which integrates the Asymptotic Feature Pyramid Network (AFPN) module to enhance the perception capability for small objects. Experimental results indicate that the YOLO-LSBA model achieves a precision of 97.1% and a recall of 78.3% for fruit stem recognition, with an AP of 92.4%. These metrics show improvements of 3.9%, 0.6%, and 2.2%, respectively, compared to the baseline model. Field trials further demonstrate that this model outperforms baseline models in detecting fruit stems under real agricultural conditions. This method offers new insights for intelligent harvesting.