Embarrassingly Simple Binarization for Deep Single Imagery Super-Resolution Networks.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Deep convolutional neural networks (DCNNs) excel at single image super-resolution (SISR).
• Binarizing DCNNs (quantizing weights/activations to 1 bit) is crucial for deployment on devices with limited resources.
• Existing binarization methods for SISR suffer significant performance degradation.

Purpose of the Study:

• To develop an effective binarization scheme for DCNNs in SISR that mitigates performance loss.
• To improve the flexibility and generalization of binarized networks for SISR tasks.
• To enable efficient deployment of high-performance SISR models on edge devices.

Main Methods:

• Proposed a novel binarization scheme by enforcing a compact uniform prior on network weights.
• Constrained weights to have small absolute values, facilitating gradient-based reversal of binarization.
• Integrated real identity shortcuts and employed a pixel-wise curriculum learning strategy during training.

Main Results:

• The proposed scheme significantly alleviates performance degradation in binarized DCNNs for SISR.
• Achieved performance comparable to a 5-bit quantization baseline, demonstrating effectiveness across different DCNN architectures.
• Showcased improved flexibility and generalization of the binarized network.

Conclusions:

• The novel weight constraint and training strategy effectively improve binarized DCNN performance for SISR.
• The method is broadly applicable to various DCNN architectures, offering a practical solution for resource-constrained SISR.
• This work paves the way for deploying efficient and high-quality SISR models on edge devices.