DWTSISR: a discrete wavelet-based approach to lightweight super-resolution reconstruction.

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Traditional image super-resolution methods struggle with low contrast, blurred details, and computational bottlenecks.
• Existing networks often fail to preserve crucial image information and reduce processing complexity effectively.

Purpose of the Study:

• To introduce DWTSISR, a novel super-resolution reconstruction network integrating discrete wavelet transform (DWT).
• To enhance image quality by improving contrast, details, and edge clarity while reducing computational load.
• To overcome limitations of traditional methods by combining DWT with physical optics priors.

Main Methods:

• The DWTSISR network decomposes input images using DWT for multi-scale analysis.
• Features are extracted and enhanced in the wavelet domain via a lightweight convolutional network.
• Inverse DWT reconstructs high-quality images in the spatial domain.
• A hybrid loss function combining reconstruction and perceptual loss is employed for optimization.

Main Results:

• DWTSISR demonstrates superior performance on benchmark datasets, achieving high PSNR and SSIM values.
• The network effectively preserves key image information and reduces data volume.
• On the CUMID mine dataset, DWTSISR achieved a 0.3 higher PSNR than RCAN with only 0.5 parameter count.

Conclusions:

• DWTSISR offers an efficient and effective solution for super-resolution reconstruction.
• The integration of DWT and a hybrid loss function significantly improves image detail and visual quality.
• This approach presents a promising alternative to traditional methods, offering better performance with reduced complexity.