Efficient and accurate apple detection is crucial for the operation of apple-picking robots. To improve detection accuracy and speed, we propose a lightweight apple-detection model based on the YOLOv8n framework. The proposed model introduces a novel Self-Calibrated Coordinate (SCC) attention module, which enhances feature extraction, especially for partially occluded apples, by effectively capturing spatial and channel information. Additionally, we replace the C2f module within the YOLOv8n neck with a Partial Convolution Module improved with Reparameterization (PCMR), which accelerates detection, reduces redundant computations, and minimizes both parameter count and memory access during inference. To further optimize the model, we fuse multi-scale features from the second and third pyramid levels of the backbone architecture, achieving a lightweight design suitable for real-time detection. To address missed detections and misclassifications, Polynomial Loss (PolyLoss) is integrated, enhancing class discrimination for different apple subcategories. Compared to the original YOLOv8n, the improved model increases the mAP by 2.90% to 88.90% and improves the detection speed to 220 FPS, which is 30.55% faster. Additionally, it reduces the parameter count by 89.36% and the FLOPs by 2.47%. Experimental results demonstrate that the proposed model outperforms mainstream object-detection algorithms, including Faster R-CNN, RetinaNet, SSD, RT-DETR-R18, RT-DETR-R34, YOLOv5n, YOLOv6-N, YOLOv7-tiny, YOLOv8n, YOLOv9-T and YOLOv11n, in both mAP and detection speed. Notably, the improved model has been used to develop an Android application deployed on the iQOO Neo6 SE smartphone, achieving a 40 FPS detection speed, a 26.93% improvement over the corresponding deployment of YOLOv8n, enabling real-time apple detection. This study provides a valuable reference for designing efficient and lightweight detection models for resource-constrained apple-picking robots.