Mars-TRP: Classification of Mars imagery using dynamic polling between transferred features
Arpan Nandi,Arjun Mallick,Arkadeep De,Asif Iqbal Middya,Sarbani Roy
DOI: https://doi.org/10.1016/j.engappai.2022.105014
IF: 8
2022-06-12
Engineering Applications of Artificial Intelligence
Abstract:Expeditions on Mars and interest in research orienting around these exploration missions have been accelerating more than ever, recently. Due to lack of active human interference in Mars missions, processing and accurate classification of images taken by the rovers is a very essential part of the system. Proper identification of landforms governs the accessibility of the mobile rovers on Mars' surface. Moreover, NASA has already collected over two million images from the planet, and more volumes are yet to arrive as these photographs serve as major documents for photogrammetry and studies based on remote sensing. Automatic labeling of incoming images and also making searching of the image database easier in the public interest requires highly accurate image classifiers. Deriving motivation from the above causes, this study intends to implement an efficient supervised multi-class image classifier for identifying Mars imagery. However, this objective is confronted by a major bottleneck. Most datasets that are accurately labeled, portray a highly skewed nature and insufficient data to train a deep model from scratch. The MSL surface imagery labeled dataset captured by the Curiosity rover, that has been considered for this study, is one such dataset with only 6691 images distributed unevenly into 25 classes. These obstacles are less signified in the existing literature and hence this paper addresses these challenges, outperforming the state-of-the-art metrics. Due to the absence of large data volume, a transfer learning based methodology was considered, using very deep convolutional networks pre-trained on ImageNet dataset. But images from Mars often involve a difference in hue, contrast and clarity when compared with images taken on Earth. Hence, the deep model was fine-tuned with our dataset and the extracted feature from the tuned neural network was used for the final classification. It was found that the results obtained from a single pre-trained model were not optimum and that ensemble approaches could unify many such results into a better result. Similar feature vectors were extracted from a few other pre-trained models. The whole setup converges into a dynamic routing module, a novel polling algorithm, which for each image, comes to an agreement about the best set of features while generating the output probability vector. The proposed approach is evaluated by several numeric metrics like accuracy, precision and recall, confusion matrices and roc curves, against the chosen individual pre-trained models and most prominent ensemble methods. Mars-TRP produces a test accuracy of around 88% in the standard test set of MSL surface dataset and an accuracy of 96% in the HiRise dataset, outperforming the individual pre-trained models, all the ensemble baselines and other existing approaches.