This invention discloses an open-lexical semantic segmentation method based on unlabeled text prototypes. It comprises three core steps: reliability scoring throughout confidence pixel selection, text prototype updating, and prototype interval preservation. Without training a visual
encoder, learning cue parameters, or relying on
manual annotation, it mines reliable visual evidence from high-confidence pixels in the target domain image. Furthermore, instead of applying the same update magnitude to all categories, it calculates a reliability
score based on the number of pixels in each category, the predicted interval strength, and the interval stability. This
score is used to simultaneously control the text prototype update step size and the prototype
orthogonalization priority. In this way, the invention can suppress the cumulative amplification of
noise pseudo-labels while maintaining the prototype interval between similar categories. Thus, while keeping all network weights of the visual-
language model frozen, this invention only utilizes unlabeled images of the target domain to iteratively correct category text prototypes, transforming text prototypes from prior anchor points of natural image language into discriminative anchor points that conform to the visual distribution of the target domain.