Abstract: Real-time and accurate individual identification of dairy cows is a prerequisite for building a perfect technical architecture for precision dairy farming. It is crucial to ensure that the identification model is lightweight while identifying individual cows quickly and accurately. In this research, a fast and accurate identification model of individual cows with low computation and small number of parameters was proposed. YOLOv5s network was selected as the original model. The scale factor in the batch normalization layer was used as the basis for judging the importance of the channel in the model for reducing the network complexity. In order to compress the model effectively, sparse loss term was added to the loss function to sparse model channels. Experimental results demonstrate that the mAP of the pruned model was 99.50%, the floating point operations (FLOPs) was 8.1 G, the number of parameters (Params) was 1.630 M, and the detection speed was 135.14 frames/s. Among all the similar methods which have been compared, the proposed method has the smallest model complexity. Moreover, the proposed model was less dependent on coat patterns and had better performance under low illumination conditions than other models in robustness. The proposed method has the characteristics of fast, accurate, robust, low computational cost and small number of parameters. It is of great potential in advancing the refinement of cow breeding on farm management.