Self-Improving SAM With Specialist Knowledge via Adaptive Direct Preference Optimization for Crater Segmentation

Martian crater segmentation is essential for large-scale data analysis in planetary science. Current methods often face a hard trade-off between accuracy and efficiency. Fast generalist models lack precision on complex craters with ambiguous boundaries, while accurate specialist models are too computationally expensive for domain-specific deployment. To address this issue, we propose CraterSAM+, a self-improving framework that leverages Dynamic Expert Invocation (DEI) driven by Direct Preference Optimization (DPO). CraterSAM+ first handles craters using a YOLOv8 detector for localization and a SAM-based generator for initial segmentation. For challenging craters with ambiguous boundaries, we dynamically invoke a specialized expert model to refine the initial segmentation output. DPO leverages the expert’s outputs as preferred references to continuously fine-tune the generalist model, reducing long-term reliance on the computationally expensive expert. We validate CraterSAM+ generalization ability on several diverse public datasets. The framework achieves a mean IoU of 89.5%, compared to 65.8% of the basic SAM. CraterSAM+ surpasses the SOTA expert model on the ASU dataset by 1.4%, achieving specialist-level accuracy with the computational efficiency of a generalist model.