To improve the object detect precision of Convolutional Neural Network (CNN), we present a YOLOv3 network which based on improved loss function. This network model uses a new loss function Tan-Squared Error (TSE) which transferred from primary Sum Squared Error(SSE), and works better on continuous variable error computing. Meanwhile, the properties of TSE could decrease the impact of vanishing gradient problem in sigmoid function, and speed up model converging. The experiment results in Pascal VOC dataset show that TSE improves the detect precision effectively compared with the performance of primary network model, and the convergence is accelerated.