Pseudolabel Comprehensive Filtering and Edge Refinement for Semisupervised Instance Segmentation in Remote Sensing Imagery

Abstract

Instance segmentation of remote sensing images enables simultaneous target differentiation and pixel-level classification, boasting broad applications in urban planning, military reconnaissance, and marine monitoring. However, existing methods often exhibit strong label dependency, and acquiring large-scale, fine-grained mask labels is costly. To address this issue, we propose a pseudolabel comprehensive filtering and edge-refinement enhanced semisupervised remote sensing instance segmentation network (CFER-Net). It leverages a small number of labeled images and a large amount of unlabeled data to achieve low-cost fine-grained perception of remote sensing targets. First, we establish the pseudolabel comprehensive (PLC) filtering strategy, which fuses mask quality scores and category scores to select high-quality pseudolabels, thereby guiding the effective training of the student module. Second, the edge refinement perception (ERP) module was introduced, which comprehensively extracts detailed and global features, thereby reducing the impact of boundary noise in pseudolabels and enabling the accurate perception of remote sensing targets. In addition, we develop an instance-level data augmentation method to enhance the model’s ability to learn objects of interest in remote sensing images. Experimental results demonstrate that CFER-Net can achieve high-performance, fine-grained mask perception of remote sensing targets with a limited scale of labeled data, and its segmentation performance outperforms state-of-the-art semisupervised instance segmentation methods.