Research On Rock Instance Segmentation And Classification Method For Mars Exploration

Mars exploration can enable human beings to go out of the earth,develop and utilize Martian resources to serve human beings.Rock sampling and data return are important tasks in Mars exploration.Through automatic segmentation and classification to identify rocks on the surface of Mars,and selectively drill and sample rocks,it can efficiently survey the geological profile of Mars.In this thesis,combined with the mission background of Mars rock sampling,the problem of rock instance segmentation and classification recognition for Mars exploration is systematically studied.The main research content of the thesis includes the following parts:(1)Rock image segmentation techniques are studied.For the rover navigation camera,motion blur and defocus blur appear when capturing the target object,and a simple convolutional neural network(CNN)algorithm cannot better distinguish the foreground and background.This thesis proposes the Mars Rover Instance Segmentation Network(MRISNet).First,the generative adversarial network Deblur GAN is used to solve the motion blur and defocus blur that appear in Mars images.Second,the residual network is improved by using the attention mechanism and feature pyramid structure,so that the extracted feature maps have rich semantic information.Finally,Simple Linear Iterative Clustering(SLIC)is used to group similar pixels in the feature map,replacing a large number of pixels with a small number of superpixels to make the segmentation finer.Experimental results show that the proposed MRISNet algorithm achieves an average precision of 76.8% on Martian rock images,and the segmentation effect is better.(2)Lightweight rock image classification techniques are studied.In view of the limitations of the on-board computer data processing capability and memory resource capacity.This thesis proposes a Mars image classification method based on Iterative Pruning Visual Geometry Group Network(IPVGGNet).First,transfer learning is used to train the connectivity of the network in order to evaluate the importance of neurons;Secondly,unimportant neurons are pruned by the iterative pruning method,so as to reduce the parameter amount and memory usage of the fully connected layer;Finally,Huffman coding is used to compress the VGGNet weight parameters after iterative pruning and quantization,so as to reduce the amount of storage and the amount of floating-point calculations.The experimental results show that the memory,Flops and accuracy of the compressed VGGNet model are better than those of the lightweight image classification algorithm,and the proposed model performs better.