Synthesis method of curve object image label pair based on diffusion model

Abstract

The invention relates to the technical field of computer vision, and provides a curve object image label pair synthesis method based on a diffusion model. The method comprises the following steps: performing Markov diffusion on an image block according to an image time step to obtain a noise-added image block; markov diffusion is carried out on the annotation block according to the annotation time step, and a noise-added annotation block is obtained; inputting the noise-added image block, the noise-added annotation block, the image modal mask block, the annotation modal mask block, the image time step and the annotation time step into a reference model for training to obtain a synthesis model; and performing noise prediction on the Gaussian noise image by using the synthesis model to obtain image noise and annotation noise, and performing reverse sampling on the Gaussian noise image by using the image noise and the annotation noise to generate a synthesis image and a synthesis annotation. According to the method, the synthetic model obtained by training can perform synchronous processing on the image and the annotation, so that the synthetic image and the synthetic annotation are synchronously generated, and the consistency between the synthetic image and the synthetic annotation is ensured.

Images