Training method for image semantic segmentation model and server

Embodiments of this application disclose a method for training an image semantic segmentation model performed at a server, to locate all object regions in a raw image, thereby improving the segmentation quality of image semantic segmentation. The method includes: obtaining a raw image used for model training; performing a full-image classification annotation on the raw image at different dilation magnifications by applying a multi-magnification dilated convolutional neural network model to the raw image, and obtaining global object location maps in the raw image at different degrees of dispersion corresponding to the different dilation magnifications, wherein a degree of dispersion is used for indicating a distribution of a target object on an object region positioned by the multi-magnification dilated convolutional neural network model at a dilation magnification corresponding to the degree of dispersion; and training an image semantic segmentation network model using the global object location maps as supervision information.