Throssed object anomaly detection method based on Vit network heuristic self-supervised training

The invention discloses a thrown object anomaly detection method based on Vit network heuristic self-supervised training. The method comprises the following steps: obtaining a highway thrown object data set, carrying out self-supervised training of a Vit network, building a dichotomy network by using a teacher network, carrying out classification network training and abnormal region positioning in the dichotomy network, and detecting thrown object anomaly. According to the invention, based on a Vit teacher-student network, self-supervised training is carried out on a super-large-scale natural image data set, so that the Vit network learns structural features of a natural normal image in a high-dimensional space, local spatial irregularity of an abnormal image can be found in a complex scene environment of an expressway, and the accuracy of the abnormal image is improved. And then on this basis, a binary classification network based on the pre-trained Vit skeleton network is constructed to carry out highway th