Semi-supervised crowd counting method based on mixed disturbance

The invention discloses a semi-supervised crowd counting method based on mixed disturbance. According to the method, a teacher-student architecture is adopted to enable a model to mine label-free data through a consistency regularization strategy, the dependence of a crowd counting model on labeled data is reduced, and the method specifically comprises the steps of obtaining a small amount of labeled crowd image data sets and a large amount of label-free crowd image data sets, performing weak disturbance on the image data by adopting conventional image data enhancement, and obtaining a crowd image data set; semantic disturbance and non-semantic adversarial disturbance based on spatial texture transformation are used, and diversified strong disturbance is applied to the weak disturbance label-free image from the semantic level and the non-semantic level; a prediction difference result of teacher and student networks on a label-free image is calculated on a feature layer and an output layer as unsupervised loss