YOLOv10-PTA: enhancing pre-thinning apples detection in orchards

Abstract To address the challenge of low detection accuracy for pre-thinning apples owing to illumination variations, multi-scale fruit distribution, and branch occlusion in complex orchard environments, this study proposes a novel model named YOLOv10-PTA, based on the YOLOv10 architecture. Specifically, starting from YOLOv10n, the model incorporates three key architectural enhancements. First, to improve the model’s ability to distinguish fine-grained features between apples and the background under strong or backlit illumination, a pre-thinning apple attention module is embedded in the second layer of the backbone, improving feature robustness in complex lighting environments. Second, to address the significant variations in apple scale and posture within the dataset, the original Path Aggregation Feature Pyramid Network (PAFPN) is replaced by an Enhanced PAFPN, which strengthens multi-scale and multi-angle target representation. Finally, to mitigate issues of fruit occlusion and difficulties in small apple detection, the hierarchical receptive field fusion module is introduced in the neck stage, substituting the original C2f structure. By integrating multi-scale receptive fields, the model enhances its contextual understanding, resulting in improved detection performance for occluded targets. The experimental results demonstrate that YOLOv10-PTA outperforms the baseline models in the task of pre-thinning apple detection. It achieves mAP@0.5 of 96.5% and mAP@0.5:0.95 of 63.3%, representing improvements of 1.9% and 6.1%, respectively, over the YOLOv10n baseline. More importantly, the model exhibits strong resilience to varying lighting conditions, significant scale changes, and branch occlusions, providing reliable detection performance that is well-suited for intelligent orchard management systems.