DSG-PL: ROI Extraction Based on Dual Saliency Guided Progressive Learning for Weakly Labeled Remote Sensing Images

Region of Interest (ROI) extraction from weakly annotated remote sensing images (RSIs) can save the huge labor cost of labelling accurate pixel-level annotations. However, weakly supervised approaches with sparseness and incompleteness inevitably result in a performance gap compared with fully supervised counterparts. To tackle this issue, a dual saliency guided progressive learning (DSG-PL) framework is developed, which focuses on progressively enhancing the quality of supervision from the image level to the precise pixel level. To begin, a dual saliency constraint mechanism is created to guide the training of a classification network in both an explicit and implicit manner for generating integral pixel-wise pseudo labels (PLs). Then, to gradually refine the initial PLs, an adaptive label self-correction module is presented, in which the updated labels are used to iteratively train a context-enhanced segmentation network, therefore constantly boosting model performance. Finally, a confidence-aware denoising loss is intended to alleviate the impacts of training with noisy PLs by adaptively reweighting the pixel-wise loss with confidence scores. Comprehensive evaluations and ablation studies verify the superiority of the proposed DSG-PL.