| As deep space exploration technology continues to improve,Mars has become a new research focus.Martian rocks contain a wealth of geological information,making them ideal targets for Mars exploration.The operators on Earth can select suitable targets by the rover and then measure them for in-depth analysis using the onboard instruments.Due to the long communication delay between Mars and Earth,it is very important to improve the autonomous operation capability of the rover.The autonomy of the rover includes two aspects:one is to choose the right rock targets for exploration independently,and the other is to choose the desired data to send back to Earth.In this thesis,the deep learning technology is used to improve the autonomy of the rover in the following aspects:(1)Conducting petrological investigation and collecting data,collecting four types of rock images based on Martian petrological research,and making Martian rock data sets.The convolutional neural network method based on VGG-16 and parameter transfer was used to build a model for automatic recognition.For the problem of small scale of Martian rock data set,the flip method and random image fusion method are used to enhance the data,and the experiment shows that this method is helpful to improve the accuracy of the model.Compared with other convolutional neural network methods and traditional feature extraction combined with SVM(Support Vector Machine),the combination of VGG-16 and transfer learning achieved the best effect among the seven methods,and achieved high recognition accuracy.(2)Aiming at the task of automatically acquiring the rock object in the picture,the Martian rock image segmentation data set was made by ourselves,and the method based on the full convolutional neural network architecture was used for automatic image segmentation.Traditional methods,such as the ROCKSTER rock image segmentation method used by Curiosity,are prone to misjudge shadows as rocks and miss judgments.The method in this thesis overcomes these problems and can basically identify rocks suitable for research and exclude shadows from the range of rocks.(3)In practical application,neural network is easily limited by factors such as delay,energy consumption and model size,which are even more severe in the Martian environment.To solve this problem,we adopt the convolutional kernel pruning method,and use the polynomial expansion method to evaluate the importance of the convolutional kernel,then the convolutional kernels that has a low impact on the loss value are prunned,so as to achieve the purpose of improving the model running speed and reducing the model size.The CPU configured with Jetson TX2 can provide a total of 375 MIPS(Million Instructions Per Second)of computing power,which is close to the 400 MIPS of computing power of Curiosity’s processor.Therefore,Jetson TX2 is used as the reference device of Curiosity computer to test the pruning effect and roughly discuss the application feasibility.Through experiments,the recognition time of single sheet was reduced from 0.779s to 0.062s while the pruning model maintained a high recognition accuracy. |
