The present invention proposes an image semantic segmentation method based on an object-level Gauss-Markov random field. The method comprises a step of carrying out initialization and over-segmentation on a pixel-level image to obtain an object-level image and a region adjacency graph and defining a neighborhood system, an observation feature field and a segmentation label field in the region adjacency graph, a step of carrying out Gauss-Markov modeling on the features of each region and the features of an adjacent neighborhood of the observation feature field according to the object-level segmentation label field and the neighborhood system and constructing an object-level linear regression equation of each region, and a step of carrying out probability modeling on the feature field and the label field, obtaining the posterior distribution of the segmented label field according to the Bayesian criterion and obtaining a final segmentation result based on the maximum posterior probability criterion. The method can be used in a system of semantically segmenting images in batches in the backgrounds of complex semantics and high spatial resolution, and the working efficiency is greatlyimproved compared with manual detection.