BoSR: A CNN-based aurora image retrieval method.

Area of Science:

• Computer Vision
• Remote Sensing
• Astronomy

Background:

• Deep learning, particularly Convolutional Neural Networks (CNNs), excels in natural image retrieval.
• Remote sensing images, especially those from all-sky-imagers (ASI) capturing aurora, often require manual or traditional retrieval methods.
• Leveraging advanced CNNs for remote sensing image retrieval, specifically aurora imagery, presents a significant research challenge.

Purpose of the Study:

• To develop an advanced CNN-based method for efficient and accurate retrieval of aurora images.
• To address the limitations of existing retrieval methods for anamorphic remote sensing imagery.

Main Methods:

• A salient region determination (SRD) scheme was developed and integrated into the Mask R-CNN framework.
• Images were treated as a "bag" of salient regions (BoSR), with each region represented by CNN features.
• Extracted CNN features were clustered to form a visual vocabulary, and features were quantized for indexing.
• Online retrieval involved calculating similarity scores and ranking images based on these scores.

Main Results:

• The proposed method demonstrated improved retrieval accuracy compared to traditional approaches.
• Significant gains in retrieval efficiency were observed on a large aurora dataset.
• The integration of SRD within Mask R-CNN effectively handled the characteristics of ASI aurora images.

Conclusions:

• The developed SRD-embedded Mask R-CNN framework offers a robust solution for remote sensing image retrieval.
• This approach effectively leverages deep learning for specialized imagery like aurora, outperforming existing methods.
• The findings pave the way for more advanced applications in automated analysis of remote sensing data.