Enhanced aggregation mask and diffusion model-based instance segmentation method

Abstract

The invention relates to an instance segmentation method based on an enhanced aggregation mask and a diffusion model. The method comprises the following implementation steps: firstly, extracting an image feature map of an image by using a backbone network; performing up-sampling fusion by using a pixel decoder to obtain multi-scale pixel-level mask features of the image; performing noise addition on the labeled data by using a diffusion model, constructing a noise filter to suppress and remove noise in the input data, and constructing a dynamic mask head to remodel an instance mask; constructing a local decoder, generating local mask branches of different tasks, and respectively processing classification, regression and segmentation tasks; performing enhanced aggregation on the dynamic prediction instance mask and a local prediction mask branch; defining a loss function for the model to train the model; the trained model can obtain a category, a bounding box and a segmentation mask of an object in an image according to the input image; according to the method, image details can be better reserved in a complex instance segmentation task, and the boundary and range of each instance can be more accurately determined.

Images