Multi-Granularity Canonical Appearance Pooling for Remote Sensing Scene Classification.

Area of Science:

• Computer Vision
• Remote Sensing
• Machine Learning

Background:

• Recognizing remote sensing scenes is difficult due to visual-semantic discrepancies.
• Lack of detailed annotations hinders pixel-level to semantic label alignment.
• Manual tagging is labor-intensive and subjective.

Purpose of the Study:

• To propose a novel Multi-Granularity Canonical Appearance Pooling (MG-CAP) method.
• To automatically capture the latent ontological structure of remote sensing datasets.
• To improve the accuracy of remote sensing scene recognition.

Main Methods:

• A granular framework progressively crops images to learn multi-grained features.
• Canonical appearance is discovered from transformations for each granularity.
• CNN features are learned using a maxout-based Siamese architecture.
• Features are replaced with Gaussian covariance matrices and normalized.
• Stable GPU training for eigenvalue decomposition and back-propagation using matrix calculus are provided.

Main Results:

• The proposed MG-CAP framework achieves promising results on public remote sensing scene datasets.
• The method effectively addresses visual-semantic discrepancies in remote sensing imagery.
• Enhanced feature discriminative power is achieved through matrix normalization.

Conclusions:

• The MG-CAP method offers a robust solution for remote sensing scene recognition.
• Automating the capture of dataset structure significantly improves performance.
• The framework demonstrates the potential for more accurate and efficient analysis of remote sensing data.