Active learning method based on positioning stability and attention loss prediction

The invention discloses an active learning method based on positioning stability and attention loss prediction. A target detection model YOLOv5 is combined to realize landing application in a railway train fault detection related data set. The method comprises the following steps: preparing a railway train fault image as a target detection data set, and selecting an image target detection model as a task model; constructing an attention module and a loss prediction module; selecting a feature map as input of a loss prediction module according to the task model; calculating a positioning stability score according to the prediction positioning frame change of the sample before and after noise disturbance by the task model; and finally, comprehensively investigating prediction loss and positioning stability, and realizing active learning sample selection. According to the method, the loss prediction module based on the attention mechanism is used for carrying out loss prediction, and multi-dimensional investigat