Local Means Binary Networks for Image Super-Resolution.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Deep convolutional neural networks (CNNs) drive single image super-resolution (SISR) success but are computationally intensive.
• Lightweight models, through methods like quantization, are crucial for real-time SISR on edge devices.
• Existing 1-bit quantization methods cause significant information loss, especially in color details, hindering SISR performance.

Purpose of the Study:

• To address information loss in binary quantization for SISR.
• To develop a lightweight yet effective binarization scheme for SISR.
• To improve the performance of binary neural networks in image super-resolution tasks.

Main Methods:

• Proposed a binarization scheme based on local means with dynamic thresholds for quantization.
• Each activation value uses a threshold determined by its surrounding values to preserve feature map details.
• Introduced a gradient approximator for adaptive optimization of binary weight updates.

Main Results:

• The proposed method significantly reduces information loss compared to uniform thresholding.
• Achieved superior peak signal-to-noise ratio (PSNR) values on standard SISR benchmarks (VDSR, SRResNet).
• Demonstrated improved visual quality in super-resolved images generated by binary networks.

Conclusions:

• The local means-based binarization scheme effectively preserves detailed color information in SISR.
• This approach enables the creation of highly efficient, lightweight binary networks for real-time image super-resolution.
• The method outperforms existing state-of-the-art algorithms for binary SISR, offering better image fidelity.