BiMars: bilateral heterogeneous network with Cross-Branch Alignment for Martian terrain semantic segmentation

Source abstract

Abstract Autonomous navigation of Mars rovers relies on high-performance semantic segmentation to realize path planning and target recognition in complex and harsh Martian terrain. However, Martian surface image parsing still faces three core bottlenecks: extreme class imbalance, low discriminability between geologically similar materials, and severe interference from illumination variations. To address the above limitations, this paper proposes BiMars, a bilateral heterogeneous network for Martian terrain semantic segmentation. The network decouples local detail preservation and global context modeling through parallel Convolutional encoder and State-Space (Mamba) encoder, and designs a Cross-Branch Alignment (CBA) mechanism to bridge the semantic gap between heterogeneous feature streams, realizing collaborative optimization of local boundary features and global semantic features. On this basis, we construct a hierarchical feature enhancement framework with Shallow Detail Enhancement (SDE) and Deep Semantic Aggregation (DSA) modules to improve feature representation for multi-scale and anisotropic Martian terrain structures. Meanwhile, a Multi-Strategy Fusion (MSF) module is designed in the decoding stage, which integrates multi-level features through three task-specific pathways to explicitly alleviate the adverse effects of class imbalance and illumination artifacts. Extensive experiments on three public Mars datasets (SynMars-TW, S5Mars and AI4Mars) show that BiMars achieves competitive performance (mIoU of 84.15%, 73.31% and 88.06% respectively), with significant accuracy improvement especially on rare and small target categories. Furthermore, we build a science-driven path planning application framework based on the segmentation results, which realizes the conversion from pixel-level terrain classification to scientific value navigation, providing technical support for autonomous geological exploration of Mars rovers.