QuantSR+: Pushing the Limit of Quantized Image Super-Resolution Networks

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Low-bit quantization compresses super-resolution (SR) models for efficient deployment.
• Ultra-low precision (2-4 bits) in SR models often leads to significant performance degradation.
• Existing methods struggle to balance accuracy and efficiency at extreme low bitwidths.

Purpose of the Study:

• To propose QuantSR+, a unified framework for ultra-low bit quantization in SR models.
• To improve the trade-off between accuracy and efficiency for quantized SR.
• To enable effective SR model deployment on resource-constrained hardware.

Main Methods:

• Redistribution-driven Bit Determination (RBD) reshapes quantization distributions to maintain representational fidelity.
• Quantized Slimmable Architecture (QSA) progressively prunes models to meet efficiency targets.
• Slimming-guided Function-localized Distillation (SFD) aligns features and accelerates convergence.

Main Results:

• QuantSR+ achieves state-of-the-art performance compared to existing quantized SR methods.
• On Urban100 (×4), QuantSR+ improves PSNR by 0.29 dB over the 2-bit SOTA baseline.
• Achieves significant efficiency gains: up to 87.9% reduction in operations and 89.4% in storage at 2-bit.

Conclusions:

• QuantSR+ effectively addresses performance drops in ultra-low bit SR models.
• The framework demonstrates broad applicability across convolutional and transformer-based SR architectures.
• QuantSR+ offers a superior balance of accuracy and efficiency for practical SR deployment.