Smilies: A Soft-Multi-Label-Guided Weakly Supervised Semantic Segmentation Framework for Remote Sensing Images

Abstract

Weakly supervised semantic segmentation (WSSS) can well solve the problem of insufficient samples for semantic segmentation of remote sensing images. In order to better solve the problem of insufficient training samples, we introduce a Soft-MultI-Label-guIded wEakly Supervised semantic segmentation framework (SMILIES). It can generate pixel-level labels relatively under the supervision of few-shot image tag levels. We test it on the GID dataset (150 images) after training on only five images, and obtain 58.7% mIoU, which is higher than other WSSS methods when using few shot training samples. When we adopt image tag as supervision apply inferring on test data, our method has a better performance than fully-supervised DeepLabV3 with the same training samples. It can be inferred from the experiment that the SMILIES has better generalization performance and manual pixel-level labeling can benefit from it.