Mars Surface Image Segmentation Based On Atrous Convolutional Networks And Few-shot Learning

From ancient times to the present,out of curiosity and search for the unknown universe,human beings have never stopped the pace of the exploration of exoplanets.Since the distance between the Mars and the Earth is far,real-time communication cannot be realized,so how to achieve the autonomous and intelligent exploration of the probe has become an important problem that need to be solved urgently.The image segmentation is the key to help the machine to understand the surroundings,and it aims to assign a corresponding semantic label to each pixel in the image.This task is widely used in autonomous driving,video surveillance and other areas.The Mars surface exploration mission has some difficult problems such as harsh unknown environment,lack of prior knowledge and severely limited resources.Meanwhile,targets on the surface are characterized by too similar texture,huge difference in scale and fuzzy boundaries.Therefore,compared with the natural scene,how to effectively perceive the Mars scene and accurately segment the surface targets are more challenging.Moreover,in the real scenario of Mars,samples that can be labeled is very limited,and the dataset cannot be constructed for some rare targets.Based on atrous convolutional networks and few-shot learning,this paper is systematically specialized in the Mars surface image segmentation task and the proposed segmentation model can provide rich global and local information which are suitable for various scale targets.With the help of multi-task learning,the accurate representation of features is learned,so as to effectively improve the consistency within the segmentation target boundary.Meanwhile,few-shot learning method is used to solve the problem of insufficient training samples of some targets.In the non-structural complex Mars scene,this paper aims to establish the accurate segmentation and characterization of multi-scale targets.The main research achievements are as follows:(1)An image segmentation method based on Cycle Atrous Spatial Pyramid Pooling.It is used to solve the problem of excessive difference between the scales of various targets on the Mars surface.Firstly,a deep atrous convolutional network is used to extract the features of images.Secondly,the generated features are passed through the Cycle Atrous Spatial Pyramid Pooling module composed of a series of bidirectionally connected atrous convolutional layers.The forward densely connected atrous convolutions can continuously expand the receptive fields and provide rich contextual information.Then,the produced high-level features are backward to the bottom atrous convolutions to refine the kernels and suppress noises.Finally,multi-scale features are fusion by Alignment with Deformable Convolution module to compensate for the loss of detailed information at the encoding stage.Great performance improvements have been made on the Simulation Mars dataset and Mars Curiosity dataset.It shows the effectiveness of Cycle Atrous Spatial Pyramid Pooling and Alignment with Deformable Convolution module in semantic feature learning.(2)A multi-task image segmentation method based on Edge Detection.It proposed to solve the problem of inaccurate boundary segmentation and low semantic consistency within the target due to the similar target texture.Firstly,the features of the image are extracted by a general deep atrous convolutional network,and the multi-scale features generated by it are regard as the input of the Edge Detection module.Secondly,the features selection and integration are performed on the high-resolution scale,which makes it possible to fully retain the detailed information.Finally,the Feature Fusion module is used to fuse semantic features and edge features.Besides,the prediction boundary from Edge Detection module is introduced into Feature Fusion model to enhance the boundary position.At optimization stage,the generally used cross-entropy losses are used to supervise the Segmentation module and Edge Detection module respectively.In addition,the duality of the two tasks that means the consistency between the edge of the predicted segmentation result and boundary ground truth is also used for joint learning.Experiments on the Simulation Mars dataset and the Mars Curiosity rover dataset prove the good performance of the multi-task learning method on the image segmentation problem.(3)A few-shot segmentation method based on Class Feature Assignment model,which is used to solve the difficult problem of the lack of training samples for rare targets such as meteorites,minerals and so on.Few-shot learning draws support from the features’similarity of targets in a same class.But the model trained on the natural image dataset cannot provide accurate representations for the Mars targets.Therefore,the Features Updated module is designed to update the original features by recombining prototypes according to the labels.In this way,the similarity of features is improved.Then,on the premise of predicting by support prototypes,a self-prediction branch is added to segment parts without appearing in the support image.Experiments on Mars datasetMARS-1~i and the natural image dataset PASCAL-5~i demonstrate that the use of Features Updated module and Dual-path Prediction module have significantly improved the performance of few-shot segmentation.