Intification of geological salient landforms is a primary requirement for spacecraft motion estimation and obstacle avoidance. As a large volume of high-resolution images are acquired by the Mars reconnaissance orbiter, growing number of approaches are proposed to develop automated approaches to detect a particular landform on Mars. However, most existing objective detection models are limited to sliding window-based and morphology-based algorithms, which require complicated preprocessing operations and can hardly be generalized to detecting different types of landforms. In this article, we aimed at developing a multitype landform detection system based on a light-weight deep learning framework, which has a quite small model size but presents excellent performance. This specific deep learning-based framework is named as mini shot multibox detector (SSD), by downsizing and modifying the existing single SSD. In the mini-SSD, some components are further optimized to adapt to this domain specific problem. A pretraining strategy is well-designed and merged into the entire model training process. In the performance evaluation tests, the proposed framework was trained and tested on images with different scales collected from different locations in high resolution imaging science experiment database. Results demonstrate that the introduced Adam optimizer and pretraining strategy can form positive effective to both model training and inference performance. The proposed framework and strategy combination outperforms the original SSD300, faster R-CNN, and YOLO series models as well as all state-of-the-art sliding window-based detectors in the field, namely AdaBoost with LBP features, AdaBoost with Haar features, and support vector machines with histogram of oriented gradient (HOG) features, in two different testing sets. Additionally, it shows high resilience on detecting target landforms in different environments with various sizes and shapes from the qualitative analysis, and can be generalized as a tool for relevant applications.

中文翻译：

---

使用基于深度学习的轻型探测器自动检测火星上的多种地貌


地质显着地貌的强化是航天器运动估计和避障的首要要求。随着火星侦察轨道飞行器获取大量高分辨率图像，人们提出了越来越多的方法来开发自动化方法来检测火星上的特定地形。然而，大多数现有的客观检测模型仅限于基于滑动窗口和基于形态学的算法，需要复杂的预处理操作，并且很难推广到检测不同类型的地貌。在本文中，我们的目标是开发一种基于轻量级深度学习框架的多类型地貌检测系统，该系统具有相当小的模型尺寸但具有出色的性能。这种基于深度学习的特定框架被命名为迷你镜头多盒检测器（SSD），通过缩小和修改现有的单个 SSD 来实现。在迷你SSD中，一些组件被进一步优化以适应该领域的特定问题。预训练策略经过精心设计，并融入到整个模型训练过程中。在性能评估测试中，所提出的框架在高分辨率成像科学实验数据库中从不同位置收集的不同尺度的图像上进行了训练和测试。结果表明，引入的 Adam 优化器和预训练策略对模型训练和推理性能均产生积极的影响。 所提出的框架和策略组合优于原始SSD300、Faster R-CNN和YOLO系列模型以及该领域所有最先进的基于滑动窗口的检测器，即具有LBP特征的AdaBoost、具有Haar特征的AdaBoost ，以及具有定向梯度直方图 (HOG) 特征的支持向量机，在两个不同的测试集中。此外，它在定性分析中显示出在不同环境中检测不同尺寸和形状的目标地貌的高弹性，并且可以推广为相关应用的工具。