Fusing multi-stage clicks with deep feedback aggregation for interactive image segmentation

The objective of interactive image segmentation is to generate a segmentation mask for the target object using minimal user interaction. During the interaction process, segmentation results from previous iterations are typically used as feedback to guide subsequent user input. However, existing approaches often concatenate user interactions, feedback, and low-level image features as direct inputs to the network, overlooking the high-level semantic information contained in the feedback and the issue of information dilution from click signals. To address these limitations, we propose a novel interactive image segmentation model called Multi-stage Click Fusion with deep Feedback Aggregation(MCFA). MCFA introduces a new information fusion strategy. Specifically, for feedback information, it refines previous-round feedback using deep features and integrates the optimized feedback into the feature representation. For user clicks, MCFA performs multi-stage fusion to enhance click propagation while constraining its direction through the refined feedback. Experimental results demonstrate that MCFA consistently outperforms existing methods across five benchmark datasets: GrabCut, Berkeley, SBD, DAVIS and CVC-ClinicDB.