Localizing Features with Masking for Satellite and Debris Classification

Abstract

In this work, we propose a localization and masking-based satellite and debris classification technique. SPAce-craft Recognition leveraging Knowledge of space environment (SPARK) dataset consists of 120K images where both RGB and corresponding Depth images are available. However, the depth images are noisy and inaccurate and significantly affect the classification task performance. To address this issue, we first create mask images of the RGB images which are used as input to the Convolutional Neural Network (CNN) for efficient classification of different satellites and debris. The depth images are first de-noised and hole filled using a simple morphological opening operation. Then masked images are calculated using both RGB and processed depth images. This masking operation provides two advantages: 1. it removes noise and fills the holes in the depth images and 2. it highlights satellites and debris while suppressing other information which does not contribute towards the classification task. We use the pre-trained EfficientNet B4 architecture and fine-tuned it with an edition of Global average pooling (GAP) and three dense layers. Our results show that the inclusion of the masking operation significantly improves the overall classification performance, achieving 97.76% accuracy on the validation data.