SCSA-APFN-YOLO: A lightweight YOLOv8n-based network for robust fragrant pear detection in complex orchard environments

To address missed and false detections caused by large-scale variation, frequent branch and leaf occlusion, fruit overlap, obvious illumination changes, and similar colors between fragrant pears and leaves in natural orchard environments, an improved recognition algorithm for fragrant pear picking robots based on YOLOv8n was proposed. The proposed model uses YOLOv8n as the baseline framework. First, a spatial-channel synergistic attention mechanism (SCSA) is introduced at the end of the backbone network to enhance the response of fragrant pear target regions and suppress complex background interference. Second, the original PAN-FPN neck of YOLOv8n is replaced with an APFN progressive feature fusion structure, and an ASFF-k adaptive spatial fusion mechanism is combined to achieve dynamic weighting of multi-scale features, thereby improving the representation capability of the model for small-scale, occluded, and overlapping fragrant pear targets. Finally, the WIoU v3 loss function is introduced to optimize the bounding box regression process and improve the matching accuracy between predicted boxes and ground-truth boxes in complex scenes. Experimental results show that the improved model achieved a precision of 86.9% and a recall of 91.2% in complex orchard environments, with mAP@0.5 and mAP@[0.5:0.95] reaching 93.7% and 86.1%, respectively, which were 3.4, 10.8, 4.7, and 7.5 percentage points higher than those of YOLOv8n. Ablation experiments demonstrate that SCSA, APFN, and WIoU v3 all effectively enhance detection performance, and their synergistic optimization achieves the best comprehensive detection performance. Compared with mainstream algorithms such as YOLOv5n, YOLOv7-tiny, and YOLOv11n, the improved model shows superior detection accuracy and robustness, providing technical support for visual recognition in fragrant pear-picking robots.