Frequency distribution-aware network based on discrete cosine transformation (DCT) for remote sensing image super...

Area of Science:

• Remote Sensing
• Computer Vision
• Deep Learning

Background:

• Deep learning-based single-image super-resolution is crucial for remote sensing.
• Current methods often overlook frequency domain similarities, limiting performance.
• Non-local features are vital but extracting them efficiently in the frequency domain is challenging.

Purpose of the Study:

• To propose a novel deep learning network for remote sensing image super-resolution.
• To address the limitations of existing methods by incorporating frequency distribution awareness.
• To improve the accuracy and effectiveness of image reconstruction in remote sensing.

Main Methods:

• A frequency-aware module was developed to extract frequency distribution similarities by rearranging feature matrices.
• A global frequency feature fusion module was introduced to capture multi-scale non-local information in the frequency domain efficiently.
• Discrete Cosine Transformation (DCT) was utilized as the core transformation for frequency analysis.

Main Results:

• The proposed Frequency Distribution Aware Network (FDANet) demonstrated superior performance on two remote sensing datasets.
• Experimental results confirmed effective image reconstruction capabilities.
• The algorithm outperformed several advanced super-resolution methods.

Conclusions:

• The FDANet effectively leverages frequency domain information for enhanced remote sensing image super-resolution.
• Integrating frequency distribution awareness significantly improves non-local feature extraction.
• The proposed approach offers a promising direction for high-performance remote sensing image enhancement.