Unsupervised domain adaptive semantic segmentation method based on regional feature alignment

The invention relates to an unsupervised domain adaptive semantic segmentation method based on regional feature alignment, and the method comprises the steps: training a student-teacher model: inputting a source domain picture into a student model to obtain a segmentation prediction result, carrying out the cross entropy loss calculation of the segmentation prediction result of the source domain picture and a source domain label, carrying out category-based fusion on pixel points of the source domain picture and the target domain picture, carrying out consistency loss calculation on an obtained first fusion picture through a segmentation prediction result of a student model and a pseudo label of the first fusion picture, carrying out region-based fusion on the pixel points of the source domain picture and the target domain picture, carrying out regional feature alignment on the second fusion picture and the target domain picture to obtain comparison loss, calculating gradients of three types of loss, carrying