EDR-TransUnet: Integrating Enhanced Dual Relation-Attention With Transformer U-Net for Multiscale Rock Segmentation on Mars

Detecting Mars rocks on celestial bodies is crucial for the obstacle avoidance and path planning of space probes in deep space environments. However, the irregular shapes, background similarity, sparse pixel features, and susceptibility to lighting and dust effects make rock detection methods face significant problems in complex deep space environments. This challenge significantly impacts the accuracy and computational complexity of most rock detection methods. Therefore, this study develops a novel semantic segmentation network called EDR-TransUnet for rock image segmentation on the surface of Mars. The proposed network uses the enhanced dual relation (EDR)-attention block, heterogeneous kernel convolution (HetConv), and TransUnet hybrid framework. By integrating global and local features, image-specific positional and channel features, and using the Transformer and EDR-Block, the performance of rock image segmentation on Mars is improved. The EDR-TransUnet significantly enhances the multiscale rock feature extraction capability and improves the efficiency of the encoder-decoder structure by incorporating EDR blocks in the embedding layer and each skip-connection layer. Additionally, introducing HetConv reduces the model’s computational complexity and improves its efficiency. Experiments on our newly developed ZhurongRock Mars rock dataset, the publicly available MarsData-V2 database and the synthetic TianWen-1 database (SynMars) demonstrated that EDR-TransUnet surpasses the state-of-the-art rock detection algorithms while maintaining lower computational complexity and faster inference speed. For example, the pixel accuracy (PA) and precision on the ZhurongRock dataset reached 97.82% and 96.18%, respectively. The codes and data are available at https://github.com/JYTCV/EDR-TransUnet.

I. Introduction

With the development of artificial intelligence technology, intelligent perception technology for Mars rovers has gradually become important as a key means to ensure the safety of Mars rover exploration [1], [2], [3], [4]. However, in diverse terrain and landforms, efficiently and quickly determining the distribution of Martian rocks is crucial [5], as the segmentation results of Martian rocks are vital for route planning [6], obstacle avoidance [7], and autonomous navigation of Mars rovers [8], [9]. Additionally, these results can provide geological information such as rock shape, size, weathering degree, and dispersion, which contribute to understanding the geological history and evolution of the planet [10], [11], [12]. However, due to the complexity of deep space environments, Martian rocks have various shapes and sizes, with some rocks having edges that are similar in color, texture, and brightness to the surrounding soil or sand, posing significant challenges for intelligent segmentation [13], [14], [15]. Therefore, intelligent rock segmentation on planetary surfaces still faces significant challenges.