A bounding box weakly supervised image segmentation method based on foreground-background matching

Abstract

The present invention provides a bounding box weakly supervised image segmentation method based on foreground-background matching, comprising sending a LAB image to a fully supervised segmentation network to obtain a foreground segmentation result; sending the bounding box and foreground segmentation result to a mask projection loss optimized segmentation network; using domain pixel consistency loss to constrain the foreground segmentation results of pixel pairs to be consistent; using a pixel representation module to generate a pixel feature representation that can represent its semantics for each pixel in the input image, and using pixel feature representation consistency loss to constrain pixel feature representations with high similarity to be close in space; using the feature representation to demarcate the foreground-background area according to the bounding box, using K-Means to establish a foreground-background model for the background and high-probability foreground area respectively, and using the foreground-background search loss to optimize the foreground segmentation result by comparing the pixel feature representation inside the bounding box with the foreground-background model. The present application can reduce the dependence of image segmentation on pixel-level annotation and improve the accuracy of existing weakly supervised segmentation methods.

Images