The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
Publications Copernicus
Download
Citation
Articles | Volume XLIII-B2-2020
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLIII-B2-2020, 1553–1559, 2020
https://doi.org/10.5194/isprs-archives-XLIII-B2-2020-1553-2020
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLIII-B2-2020, 1553–1559, 2020
https://doi.org/10.5194/isprs-archives-XLIII-B2-2020-1553-2020

  14 Aug 2020

14 Aug 2020

TOWARDS DISTILLATION OF DEEP NEURAL NETWORKS FOR SATELLITE ON-BOARD IMAGE SEGMENTATION

F. de Vieilleville1, A. Lagrange1, R. Ruiloba1, and S. May2 F. de Vieilleville et al.
  • 1AGENIUM Space, Toulouse, France
  • 2CNES, Centre National d’Etudes Spatiales, France

Keywords: deep learning, parameters reduction, ablation study, low rank approximation, distillation

Abstract. Cubesats platforms expansion increases the need to simplify payloads and to optimize downlink data capabilities. A promising solution is to enhance on-board software, in order to take early decisions, automatically. However, the most efficient methods for data analysis are generally large deep neural networks (DNN) oversized to be loaded and processed on limited hardware capacities of cubesats. To use them, we must reduce the size of DNN while accommodating efficiency in terms of both accuracy and inference cost. In this paper, we propose a distillation method which reduces image segmentation deep neural network’s size to fit into on board processors. This method is presented through a ship detection example comparing accuracy and inference costs for several networks.