Abstract: The invention relates to a method (200) for training a dynamic-architecture convolutional neural network for semantic image segmentation, comprising a phase (102) of auto-annotation of images, to constitute a database (B1), called the first database, of auto-annotated images, comprising the following steps: applying (104) at least one transformation to at least one source image, and adding (106), to each transformed image thus obtained, a tag, called the first tag, corresponding to said transformation; said method (200) comprising at least one iteration of a phase (110) of training said neural network with the first image database (B1), and a second image database (B2) annotated with a tag, called a second tag, relating to the content of said image. It also relates to a computer program and a device implementing such a method, and a deep-learning dynamic-architecture convolutional neural network obtained by such a method.