The invention provides an image semantic segmentation method based on positive sample and label-free sample learning. The invention belongs to the technical field of computer vision, the method comprises a data preparation step, a data preprocessing step, a deep convolutional neural network construction step, a PU-Learning-based loss function design step, a loss function optimization learning step, and an iterative execution training step until a training result of an image semantic segmentation model satisfies a predetermined convergence condition. According to the invention, a deep neural network is adopted to extract to-be-segmented image features; on the basis, the invention further discloses a preparation method, according to the invention, a cross entropy loss function based on PU-Learning is designed; according to the method, the semantic segmentation model can be trained and optimized under the condition that only part of pixel levels are labeled, the semantic segmentation model can be trained and optimized in an end-to-end mode, meanwhile, direct supervision of the pixel levels is reserved to a certain extent, and the data labeling speed is increased while the good semantic segmentation quality is guaranteed.