LR Aerial Photo Categorization by Cross-Resolution Perceptual Knowledge Propagation

Area of Science:

• Computer Vision
• Remote Sensing
• Artificial Intelligence

Background:

• Satellite imagery provides vast data, with high-altitude (low-resolution) and low-altitude (high-resolution) sources.
• Analyzing low-resolution (LR) aerial imagery semantics is crucial but challenging due to perceptual complexity and data labeling costs.

Purpose of the Study:

• To develop a framework for adapting visual perception knowledge from high-resolution (HR) to low-resolution (LR) aerial photos.
• To improve the semantic understanding and categorization of LR aerial images.

Main Methods:

• A cross-resolution perceptual knowledge propagation (CPKP) framework is proposed.
• A low-rank model decomposes LR images into salient foreground and background regions, generating gaze-shifting paths (GSP) and deep features.
• A kernel-induced feature selection algorithm identifies discriminative GSP features for classifier training.

Main Results:

• The CPKP framework effectively adapts perceptual experiences from HR to LR imagery.
• Collaborative training using LR and HR labels optimizes classifier performance.
• The approach demonstrates superiority in categorizing LR aerial photos compared to existing methods.

Conclusions:

• The proposed CPKP framework offers a robust solution for semantic analysis of LR aerial imagery.
• Mimicking human visual perception and leveraging HR data significantly enhances LR image understanding.
• This method addresses challenges in perceptual complexity and data annotation for remote sensing applications.