<p>Object detection technology is critical for intelligent fruit sorting. To address issues such as the missed detection of small targets, like cherries, and susceptibility to background interference in complex scenarios, this paper proposes Cherry-YOLO, a real-time detection algorithm based on the improved YOLOv8, and constructs a dedicated cherry dataset to support the task. The contributions of this work are twofold. First, a high-quality cherry dataset is built, comprising 7,904 annotated images (expanded via data augmentation) that cover orchard and indoor environments, three maturity stages (immature, semi-mature, mature), and multiple surface defect types. This dataset targets the characteristics of small cherries (e.g., subtle features, dense distribution) and fills the gap in dedicated small-fruit detection datasets. Second, Cherry-YOLO is proposed with four key improvements: (1) A FasterNet-RepMixer backbone to enhance small-target feature extraction; (2) A DySlimNeck dynamic neck with cross-scale fusion to optimize multi-scale feature integration; (3) A Detect-FRM detection head to improve localization accuracy; (4) An AGF-CIoU loss function to enhance robustness under occlusion. Experimental results show that Cherry-YOLO achieves an mAP50 of 0.915, mAP50-95 of 0.641, and an F1 score of 0.85 on the constructed dataset, outperforming YOLOv8n and other mainstream models. The model and dataset together provide a reliable solution for intelligent cherry sorting in real-world environments.</p>