Remote sensing image Super-resolution reconstruction by fusing multi-scale receptive fields and hybrid transformer

Area of Science:

• Remote Sensing
• Computer Vision
• Image Processing

Background:

• Super-resolution reconstruction algorithms struggle with missing details in remote sensing images during training.
• Enhancing high-frequency perceptual information and texture details is crucial for remote sensing applications.

Purpose of the Study:

• To propose an improved remote sensing image super-resolution reconstruction model.
• To address challenges of missing details and improve training stability in super-resolution algorithms.

Main Methods:

• Generator network utilizes multi-scale convolutional kernels and a multi-head self-attention mechanism for feature extraction and fusion.
• Multi-stage Hybrid Transformer structure processes features progressively across resolutions for enhanced reconstruction.
• Discriminator incorporates multi-scale convolution, global Transformer, and hierarchical features for refined image quality evaluation.
• Charbonnier and total variation (TV) loss functions are employed to improve training stability and convergence speed.

Main Results:

• The proposed model achieves significant performance gains compared to the SRGAN algorithm.
• Average improvements include 3.61 dB in Peak Signal-to-Noise Ratio (PSNR), 0.070 (8.2%) in Structural Similarity Index (SSIM), and 0.030 (3.1%) in Feature Similarity Index (FSIM).
• Demonstrated enhanced ability to capture fine details and global information in remote sensing images.

Conclusions:

• The improved super-resolution model effectively enhances perceptual information and texture details in remote sensing images.
• The novel architecture and loss functions contribute to superior reconstruction quality, detail recovery, and training efficiency.
• The method shows substantial performance improvements across multiple datasets, offering a promising solution for remote sensing image super-resolution.