YOLO11-SBS: leveraging ACE and ORCM to gauge robust and accurate apple counting in complex orchards

Precise apple counting in complex orchards is vital for efficient management, yield estimation, and automated picking. Traditional YOLO-based models suffer from fixed convolutional weights, ineffective feature screening, and interference from branch/leaf shadows and film bagging, leading to poor robustness. They also can't fully utilize multi-scale feature complementarity, causing detail loss and inaccuracy in counting dense or small apples. To address these, we propose the YOLO11-SBS model. It introduces the C3k2_SAC module for intelligent feature extraction, the C2PSA_BIF module to optimize feature interactions, and adds SSFF and TFE modules to boost feature representation. Furthermore, to comprehensively evaluate model robustness and counting quality, this study applies the ACE metric and constructs a task-oriented Overall Robustness and Counting Metric (ORCM). The ACE metric is used to assess robustness under illumination perturbation, whereas ORCM is designed to jointly characterize false-detection and missed-detection errors under different occlusion levels. The improved model shows marked advantages, reducing missed and false detections. It maintains high accuracy and recall, especially for green apples similar to the background, providing an efficient and accurate solution for apple counting in complex orchards with high practical value.