To achieve the accurate detection of strawberry in agricultural harvesting, a lightweight network model based on YOLOv5s is proposed considering the limited memory and low computational power of embedded devices, as well as the huge parameters and computational demands of current target detection models. First, the YOLOv5s structure is lightweight processed by replacing ordinary convolutions with depthwise convolutions （DWConv） and substituting the C3 module in the original network with the C3Ghost module to reduce the model complexity. Second, to enhance the ability of the backbone network to extract feature information and improve the interaction between channels in the input feature maps, an efficient channel attention （ECA） structure is integrated into the C3 module of the backbone network. Additionally, a parameter-free attention module （SimAM） is added to the feature fusion network, so that the model can focus on more effective feature information without increasing the number of parameters of the model while improving the recognition accuracy. Finally, transfer learning is combined to accelerate the convergence speed of the model and further improve the detection accuracy. The results indicate that the lightweight model reduces network size by 55.8% and computation by 55.1%. The mAP@0.75 tested on a custom strawberry dataset reaches 74.9%, which is 3.1% higher than that of the original model. The average inference time per image is only 6.4 ms. This enables accurate and fast detection in strawberry picking tasks and provides support for the intelligent production of strawberries.