MRISNet:Deep-learning-based Martian instance segmentation against blur

In Mars exploration missions, it is crucial for rovers to identify and locate target rocks for autonomous sampling. The rock sampling of rovers relies on the deep learning-based rock detection algorithms of rovers. However, the navigation camera of rover often has problems such as motion blur and defocus blur. The object detection algorithm cannot accurately detect the pixels of the sampling target, which causes the rover manipulator to malfunction and damage the end drill bit. In this work, we try to solve this problems by developing a deep learning network, dubbed the Mars Rover Instance Segmentation Network (MRISNet). First, the generative adversarial networks are used to mitigate image degradation caused by blur and improve the clarity of image texture. Then, we use the attention mechanism and the feature pyramid structure to improve the residual network so that the extracted feature maps have rich semantic information. We use simple linear iterative clustering (SLIC) to group similar pixels in the feature map and replace a large number of pixels with a small number of superpixels to make the segmentation more refined. Experimental results show that compared with the semantic segmentation algorithm, the proposed MRISNet algorithm achieves an Average Precision of 76.8 and a better segmentation on Martian rock images.