Semi-supervised image semantic segmentation method based on double-discriminator adversarial learning

Abstract

The invention designs a semi-supervised image semantic segmentation method based on double-discriminator adversarial learning. And a strategy of alternately training by adopting double discriminators and a divider is adopted. And performing linear mixing on different class probability graphs and target values to serve as training data of the double discriminators, and outputting a space confidence graph. When data with labels are used for training, the segmentation network is jointly supervised by cross entropy loss and adversarial loss based on a standard. And after multiple rounds of alternate training, adding label-free data, and continuing training. And carrying out binarization processing on confidence maps output by the two discriminators through a set threshold value to obtain two different high-confidence regions, and taking the intersection of the two regions as a pseudo tag for calculating the cross entropy loss of label-free data. At the moment, the segmentation network adds one item of cross entropy loss based on pseudo labels. According to the method, double discriminator adversarial learning is introduced, so that the stability of model training is improved, the segmentation performance is also improved, and the effectiveness of the method is proved.

Description

technical field [0001] The invention relates to the field of image semantic segmentation, in particular to a semi-supervised image semantic segmentation method based on dual discriminator confrontation learning. Background technique [0002] In recent years, thanks to the birth of large-scale data, the use of high-computing GPUs, and the proposal of various advanced model methods, computer vision technology based on deep learning has rapidly emerged. Image semantic segmentation, as a basic problem in the field of computer vision applications, is to classify images at the pixel level, assign an appropriate category label and boundary location to each pixel in the image, and has a very broad practical application prospect. The traditional semantic segmentation method focuses on the manually extracted features, which has achieved good segmentation results to a certain extent, and has the advantage of low algorithm cost. However, the traditional method also has certain limitati...

AI Technical Summary

Problems solved by technology

However, the segmentation model based on the convolutional neural network relies heavily on a large amount of data with pixel-level annotations for supervised learning, and the work of pixel-level manual annotation is cumbersome and difficult, which often requires a lot of manpower and material costs, and is not suitable for actual engineering projects.

Images