Reliable Label-Supervised Pixel Attention Mechanism for Weakly Supervised Building Segmentation in UAV Imagery

Building segmentation for Unmanned Aerial Vehicle (UAV) imagery usually requires pixel-level labels, which are time-consuming and expensive to collect. Weakly supervised semantic methods image-level labeling have recently achieved promising performance in natural scenes, but there been few studies on UAV remote sensing imagery. In this paper, we propose a reliable label-supervised pixel attention mechanism building Our method is based the class activation map. However, classification networks tend capture discriminative parts of object insensitive over-activation; therefore, maps cannot directly guide network training. To overcome these challenges, first design Pixel Attention Module that captures rich contextual relationships, can further mine more regions, order obtain modified Then, use initial seeds generated by synthesize labels. Finally, label loss, defined as sum differences between labels Notably, loss handle over-activation. The preceding steps significantly improve quality pseudo-labels. Experiments our home-made data set indicate achieve 88.8% mIoU test set, outperforming previous state-of-the-art weakly methods.