FABNet: Frequency-Aware Binarized Network for Single Image Super-Resolution.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Binary neural networks (BNN) offer efficient real-time single-image super-resolution (SISR).
• Existing BNN methods often overlook spatial frequency's impact on quantization error.

Purpose of the Study:

• To introduce a frequency-aware binarized network (FABNet) for enhanced SISR.
• To minimize quantization error by considering spatial frequency components.

Main Methods:

• Wavelet transformation to decompose features into low and high frequencies.
• A "divide-and-conquer" strategy to process frequency components separately.
• Dynamic binarization with learned thresholds and approximation for diverse spatial frequencies.

Main Results:

• Reduced quantization error compared to existing methods.
• Improved recovery of image textures and details.
• Superior performance in Peak Signal-to-Noise Ratio (PSNR) and visual quality on benchmark datasets.

Conclusions:

• FABNet effectively addresses spatial frequency variations in binarized super-resolution.
• The proposed approach achieves state-of-the-art results with reduced computational cost.
• FABNet demonstrates significant improvements in both quantitative metrics and visual fidelity for SISR.