Semantic segmentation method based on improved DenseNAS

The invention provides a semantic segmentation method based on improved DenseNAS. The semantic segmentation method comprises the following steps: (1) searching a proper deep neural network model by using the DenseNAS; (2) obtaining an optimal sub-network candidate based on a sub-network depth structure selection algorithm, and selecting a shortest network and an optimal network from the candidate; and (3) taking the optimal network as a backbone network to obtain a teacher model FCN-long, taking the shortest network as a backbone network to obtain a student model FCN-short, extracting features of the image through the FCN-short obtained by knowledge distillation training, and obtaining a semantic segmentation image of the image. The invention provides a structure selection method based on sub-network depth, the precision of a selection network on a data set is improved, the advantages of a dense connection search space are exerted, and a semantic segmentation algorithm considering both efficiency and precisio