Massive-Scale Aerial Photo Categorization by Cross-Resolution Visual Perception Enhancement.

Area of Science:

• Computer Vision
• Remote Sensing
• Machine Learning

Background:

• Categorizing aerial photographs under diverse conditions is crucial for autonomous navigation and environmental analysis.
• Existing image recognition methods struggle with challenges like weak annotation, multiple attributes, and cross-domain knowledge transfer.
• Aerial image perception is limited by factors such as varying weather, lighting, and geomorphic complexities.

Purpose of the Study:

• To optimize aerial photograph feature learning by integrating low-resolution spatial composition with high-resolution perceptual features.
• To develop a weakly supervised algorithm for identifying salient regions in aerial images, overcoming the need for pixel-level annotation.
• To enhance aerial photograph recognition by designing a cross-domain knowledge transfer module for multiresolution image analysis.

Main Methods:

• Extracting BING-based object patches and employing a weakly supervised ranking algorithm incorporating multiple attributes to select salient regions.
• Constructing a gaze shifting path (GSP) by linking top-ranking patches to derive deep GSP features for interpretable recognition.
• Formulating a cross-domain multilabel Support Vector Machine (SVM) that leverages global features from low-resolution images to optimize high-resolution GSP features.

Main Results:

• The proposed method demonstrates competitiveness on a million-scale aerial photograph dataset.
• The gaze shifting path (GSP) approach achieves over 92% consistency with human gaze sequences in eye-tracking experiments.
• The system effectively categorizes aerial photographs by optimizing deep features using multi-resolution data.

Conclusions:

• The developed approach significantly enhances aerial photograph categorization by effectively leveraging multi-resolution data and weakly supervised learning.
• The gaze shifting path (GSP) provides an interpretable and human-aligned method for aerial image analysis.
• This work advances the capabilities of autonomous systems in interpreting complex aerial imagery across varied conditions.