A target detection performance optimization method, comprising: in the training process for a detection model, using metric learning to adjust the distribution of samples in a feature space for generating features having a higher degree of differentiation; in the iterative training for a deep neural network corresponding to the metric learning, a candidate box used in each iteration is a candidate box determined by intersection over union (IoU) information and has a positional relation in which distances of identical target objects meet a certain constraint condition and distances of different targets meet a certain constraint condition; checking whether the features of a candidate box target generated in each iteration of the iterative training meets a similarity constraint condition; if the features of a candidate box target generated in an iteration of the iterative training meets the similarity constraint condition, the detection model does not generate loss in the current iteration, and does not need to reversely propagate output errors corresponding to all layers in a network; and during a test, inputting a picture to be detected and a candidate box set of the picture into the trained detection model to obtain target object coordinates and class information output by the detection model. The method can improve detection capability and optimize detection performance.
展开▼