This article proposed an improved YOLOv5s model to improve the accuracy of detecting plums （Prunus salicina Lindl.） with high occlusion and density in orchards and the lightweight. Firstly， a new Focus-Maxpool module was used to replace the down-sampling convolution in the backbone network， enabling the model to retain more feature information of small and highly occluded targets during down-sampling. Secondly， the weighted loss of focal loss and cross-entropy function was used as the classification loss of the model to improve its recognition ability for adhesive targets. Finally， several sets of detection experiments were designed to evaluate the performance of the model. The results showed that the average accuracy of the improved YOLOv5s model was better than that of YOLOv5s，YOLOv4，Faster RCNN，SSD，and Centernet. Compared with the detection results of the YOLOv5s model， the average accuracy， recall rate， and accuracy of the improved model increased by 2.84，9.53，and 1.66 percentages， respectively. The detection speed of the improved model reached 91.37 frames per second， meeting the requirements of real-time detection. It is indicated that the model improved has higher accuracy of detection and robustness in real orchard environments. It will provide data reference for studying fruit-picking robots and monitoring orchard environments.