Objective To address the problems of low detection accuracy high model parameter count and high computational cost in existing apple detection in orchard environments an improved orchard apple detection algorithm based on YOLOv5s is proposed. Methods First Soft-NMS was used in the YOLOv5s algorithm to replace the original NMS of the model and the way the original model processed prediction boxes was changed so as to reduce false detections and missed detections caused by the overlapping and occlusion of apples in the orchard environment and improve the detection accuracy. Then OTA was used to optimize the label assignment of the original model. The label assignment was regarded as an optimal transport problem and the context information was fully utilized to reduce the number of ambiguous boxes so as to better handle the problem of dense apple occlusion and further improve the model?? s detection performance for orchard apples. Finally the convolution module in the backbone network was replaced with FasterNet blocks which served as a new feature extraction network. This replacement reduced the parameter count and the computational cost of the model thereby achieving a lightweight model. Results The experimental results showed that on the orchard apple dataset compared with the original model the mAP of the improved algorithm model increased by 2. 8% and the parameter count and the computational cost decreased by 21% and 29% respectively. Compared with other mainstream detection models in the same series the improved model had higher detection accuracy fewer parameters and lower computational cost. Conclusion The improved model is more suitable for apple detection in orchard scenarios and can be used as a solution for reference and application in related fields.