LBNet: A Lightweight Bilateral Network for Semantic Segmentation of Martian Rock

Rock segmentation on the Martian surface is particularly critical for rover navigation, obstacle avoidance, and scientific target detection. We propose a lightweight bilateral network for semantic segmentation of Martian rock (LBNet). The network consists of a shallow spatial detail branch (SDB) and a deep semantic information branch (SIB). In the shallow spatial detail branch, dense connection channel aggregation convolution (CAConv) is adopted to establish local dependencies for each pixel and preserve detailed information. In the deep semantic information branch, channel split convolution (CSConv) is adopted to extract features by adopting different convolution kernels on different channel, reducing the similarity between different feature maps and increasing feature maps diversity. Finally, a feature fusion module (FFM) is designed to effectively fuse feature maps at different levels. With only 0.37M parameters, the model achieved 93.85% mIoU and 147.8 FPS on the dataset of Perseverance, and 88.62% mIoU and 152.5 FPS on the Curiosity dataset. Experiments show that the model achieves a good balance between segmentation accuracy and inference speed.