Research On Semantic Segmentation Algorithm Based On Deep Learning And Application Of Martian Rock Image

With the gradual development of outer space exploration capabilities,the surface environment of Mars is the most similar to the earth,and various rocks on the surface environment of Mars contain a lot of geological information,which has become the main target of research when using Mars probes.The rover can be used for target selection and research using various scientific devices on board.However,the distance between Mars and the earth is far away,and there is a long delay in the communication between the earth and the fire,so improving the autonomous detection capability of the Mars rover is the main goal.The autonomous exploration of the Mars rover includes two aspects:one is to autonomously segment the rocks in the image semantically to obtain the rock distribution information so as to be able to create a semantic map;the other is to be able to select target rocks from different rock distributions for selective transmission back to the ground.This thesis uses deep learning and other methods to study the following aspects of improving the performance of Mars rover image segmentation.Firstly,the semantic segmentation of Martian rock images is studied.Through the investigation of Martian rock data,various Martian rock images are collected according to planetary geology,and a Martian rock image data set is produced.Secondly,a semantic segmentation architecture based on encoding-decoding is used.The encoder consists of the first 13 layers of the basic VGG16 network.The working principle of the decoder is to up-sample the feature map according to the maximum pooling index in the network,while These feature maps come from the encoder.When using pooled indices for simple neural network segmentation,it is possible to ignore smaller objects.By adding two-channel attention mechanisms,namely spatial attention and channel attention,the network will improve the segmentation performance for small objects.However,due to the existence of large segmentation targets under Mars conditions,by adding dilated convolutions,the receptive field of the network can be increased,so that larger segmentation targets can be completely segmented.On the basis of the foregoing content,the segmentation model is constructed.Simulation experiments are carried out on the general data set and the self-made data set,and the results show that the method in this thesis can effectively solve the segmentation problem of small objects and large objects by comparing with other methods.Then,research on the real-time algorithm of semantic segmentation.At present,many networks can achieve a high level of segmentation speed,but there are also problems such as occupying a lot of memory and requiring high computing power.After comparing with various popular real-time segmentation model frameworks and model optimization methods,real-time semantic segmentation research is carried out using a two-channel network model,which incorporates contextual channels and spatial channels.Improve network segmentation performance by adding short-term densely connected networks and spatially sparse methods.Short-term densely connected networks can extract deep features with scalable receptive fields and multi-scale information,while spatial sparse methods can improve real-time segmentation rates with limited degradation in segmentation accuracy.Simulation testing with public datasets to confirm the effectiveness of improved methods.Tested on the aforementioned datasets.The experimental effects indicate that the performance of network segmentation can be increased at a certain resolution.Finally,the deep neural network model will be limited by factors such as delay and computing energy consumption in the application.The performance limitation of computing equipment in the Martian environment is particularly prominent.For practical application scenarios,choose the Jetson TX2 embedded Jetson TX2 embedded with the same computing power as Curiosity.equipment,and deploy the designed network model on the equipment to conduct experiments and discussions on its segmentation performance and the feasibility of its application.After practical application experiments,the model can maintain a certain segmentation accuracy on devices with low computing power,and the real-time segmentation frame rate can still reach 67FPS.