E2FIF: Push the Limit of Binarized Deep Imagery Super-Resolution Using End-to-End Full-Precision Information Flow

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Binary Neural Networks (BNNs) offer efficient deployment of deep learning models on resource-constrained devices.
• Single Image Super-Resolution (SISR) models are computationally intensive, limiting their real-world application.
• Existing BNNs for SISR focus on approximating binarized convolutions to mitigate performance loss.

Purpose of the Study:

• To improve the generalization performance of BNNs for SISR.
• To identify the critical factors for effective BNN performance in SISR tasks.
• To propose a novel, computationally inexpensive enhancement for BNNs in SISR.

Main Methods:

• Introduced full-precision skip connections over binarized convolution layers in BNNs.
• Ensured preservation of full-precision information and accurate gradient flow.
• Applied the proposed scheme to various existing BNN backbones for SISR.

Main Results:

• Demonstrated enhanced forward expressive capability and back-propagated gradient accuracy.
• Achieved superior performance compared to existing BNNs for SISR.
• Outperformed even some 4-bit SISR competitors on benchmark datasets.

Conclusions:

• Full-precision skip connections are crucial for effective BNNs in SISR.
• The proposed method enhances BNN performance without increasing computational cost.
• This approach offers a significant advancement for deploying SISR on edge devices.