Multi-Scale Deep Neural Network Based on Dilated Convolution for Spacecraft Image Segmentation.

Area of Science:

• Computer Vision
• Deep Learning
• Space Exploration Technology

Background:

• Deep learning-based image segmentation is crucial for remote sensing, medical imaging, and autonomous driving.
• Accurate segmentation of spacecraft in monocular images is vital for space station assembly and target pose estimation.
• Currently, no specialized segmentation networks exist for spacecraft targets.

Purpose of the Study:

• To propose an end-to-end spacecraft image segmentation network.
• To enhance feature extraction and contextual information processing for improved segmentation accuracy.
• To address the lack of dedicated segmentation models for spacecraft.

Main Methods:

• Utilized DeepLabv3+ as the base framework for a novel semantic segmentation network.
• Incorporated a multi-scale neural network with sparse convolution for improved feature extraction.
• Integrated dilated convolutions, a channel attention mechanism, and a parallel atrous spatial pyramid pooling (ASPP) structure.

Main Results:

• The proposed network, featuring an encoder-attention-decoder structure, effectively segments spacecraft targets with high accuracy.
• The method demonstrated significant improvements over the standard DeepLabv3+.
• Experiments on a custom spacecraft segmentation dataset validated the network's effectiveness in producing clear and complete segmentation masks.

Conclusions:

• The developed spacecraft image segmentation network offers superior performance compared to existing models.
• The integration of attention mechanisms and enhanced feature fusion is key to achieving high segmentation accuracy.
• This work provides a valuable tool for critical space exploration applications requiring precise spacecraft imagery analysis.