基于融合伽马变换全卷积神经网络的火星地貌分割方法

Image Segmentation Algorithm for Mars Terrain Based on Fully Convolutional Networks with Gamma Transformation

为了让火星巡视器在有限的寿命内能获得更多的科学产出,需要提高巡视器在火星表面的自主通过能力,而火星表面地貌类型是评估巡视器可通过性的重要信息。因此,提出了融合伽马变换的全卷积神经网络(FCN)火星表面地貌分割算法。首先,考虑对巡视器通过性的影响,确认地貌分类的类型,基于好奇号拍摄的火星地表图像Mars32K数据库,构建地貌分割数据集;其次,使用自适应伽马变换(AGT)对灰度单一的火星图像进行预处理,减弱了光照等因素干扰;最后,利用数据集训练一个FCN,以实现对火星表面地貌的分割预测。仿真结果表明:网络测试准确率达到83.03%,地形分割预测结果可靠,验证了方法的可行性。

In order to make the Mars rover to obtain more scientific returns within its limited lifespan, it is necessary to improve the autonomous passability of the rover on the surface of Mars. The type of Martian surface topography is important information for evaluating the passability of the rover. Therefore, in this paper, an image segmentation algorithm for Mars terrain based on fully convolutional networks(FCNs) with gamma transformation(GT) is proposed. First, the topography types are confirmed in view of the effects on the passability of the rover.Then, based on the Mars32K dataset of Mars surface images taken by Curiosity, a geomorphological segmentation dataset is constructed. The adaptive gamma transformation(AGT) is used to process these images. Finally, the constructed dataset is used to train a fully convolutional neural network(FCN) to achieve the segmentation of the Martian surface topography. The results show that the accuracy of the network reaches 83.03%, and reliable terrain segmentation predictions are performed on the test graphics, which verifies the feasibility of the method.