遥感图像的实例分割能够同时实现感兴趣物体的目标级定位与像素级分类，是一项重要且极具挑战性的任务。当前大多遥感图像实例分割方法依赖精细的像素级注释，其制作成本高昂。此外，遥感图像混杂的前背景和复杂的目标轮廓同样增加分割的难度。为了迎接这些挑战，本文构建出一套适用于遥感图像弱监督实例分割任务的先验信息驱动体系，并提出一种基于多重先验驱动的遥感图像弱监督实例分割网络。具体地，将遥感图像弱监督实例分割任务中的先验信息按照其来源分为任务先验和图像先验，其中任务先验源自与实例分割紧密相关的边界框检测任务，图像先验来源于对图像本身信息的归纳与挖掘。进一步地，设计框-掩码投影一致性约束、像素区分难度表征函数和中心位置先验约束三个具体组件实例化任务先验信息，驱动网络确定掩码的尺寸并使其充分关注图像中的重点像素和区域；设计邻域视觉一致性约束和梯度一致性约束两个组件构建图像先验信息，使网络高效地进行前背景区分且适应遥感图像中复杂的目标轮廓。在光学和SAR遥感图像数据集上的实验结果表明，所提方法无需任何像素级注释即可实现52.5和54.1的AP值，优于现有的弱监督分割方法，且达到全监督Mask R-CNN的89.3%和84.3%。该方法能够为遥感图像的细粒度解译提供一种高性能且低成本的解决方案。

Abstract: Objective Remote sensing image interpretation has essential application values in various fields, such as urban management, maritime monitoring, and resource planning. As an important and challenging task in remote sensing image interpretation, instance segmentation of remote sensing images can achieve target-level localization and pixel-level classification of objects of interest with fine granularity, making it a current research hotspot. However, most existing remote sensing image instance segmentation methods adopt the fully supervised paradigm and require expensive pixel-level labels. Moreover, remote sensing images often have issues such as mixed foreground and background and complex target contours, making segmentation challenging. Method To overcome these challenges, we propose a prior information-driven system suitable for weakly supervised instance segmentation tasks in remote sensing images and a multi-prior driven weakly supervised instance segmentation network (MPD-WSIS-Net) to address these challenges. The prior information of weakly supervised instance segmentation can be divided into task prior and image prior. The task prior mainly comes from the bounding box detection task, which is closely related to instance segmentation. Specifically, the paper obtains prior information from three components: box-mask projection consistency constraint, pixel discrimination difficulty representation function, and center position prior constraint. The image prior comes from summarizing or excavating information about the image itself. In this study, we focus on the relationships between neighboring pixels in the image and the gradient information of targets. By integrating these constraints and the pixel discrimination difficulty representation function, we establish a complete prior information driving system to effectively enable MPD-WSIS-Net to perform instance segmentation tasks under weakly supervised conditions in remote sensing images. Result MPD-WSIS-Net was compared with weakly supervised methods, hybrid supervised methods, and fully supervised methods on optical and SAR remote sensing image datasets. Compared to weakly supervised methods, MPD-WSIS-Net achieved better segmentation results. Compared to hybrid supervised methods, the segmentation performance of MPD-WSIS-Net on both optical and SAR remote sensing image datasets significantly surpasses that of Mask R-CNN and CondInst of 50% pixel-level annotations. It is also competitive compared to Mask R-CNN and CondInst under 75% pixel-level annotation conditions. Compared to fully supervised methods trained with pixel-level labels, MPD-WSIS-Net can achieve 89.3% of fully supervised Mask R-CNN"s AP value on optical and 84.3% on SAR remote sensing image datasets. Furthermore, we have demonstrated the positive impact of each prior information component in MPD-WSIS-Net on instance segmentation performance in optical remote sensing images through ablation experiments. Conclusion This study constructs a prior information-driven system consisting of task and image priors through a detailed analysis of the prior information in weakly supervised instance segmentation tasks. The specification of prior information is achieved through the box-mask projection consistency constraint, pixel discrimination difficulty representation function, center position prior constraint, neighborhood visual consistency constraint, and gradient consistency constraint. This research can enable MPD-WSIS-Net to complete instance segmentation tasks without pixel-level annotations and provide high-performance and low-cost prescription for fine-grained interpretation of optical and SAR remote sensing images.