Anycost Network Quantization for Image Super-Resolution

Area of Science:

• Computer Vision
• Deep Learning
• Network Quantization

Background:

• Image super-resolution (SR) networks are often deployed on mobile devices with fluctuating resource availability.
• Conventional quantization methods require extensive retraining for each new resource constraint, leading to high computational costs.
• Existing approaches struggle to balance efficiency and accuracy across diverse hardware limitations.

Purpose of the Study:

• To develop an efficient network quantization method for image super-resolution that accommodates variable resource budgets.
• To enable image SR networks to be deployed on devices with dynamic hardware constraints without significant retraining.
• To achieve an optimal trade-off between efficiency and accuracy for image SR under varying resource limitations.

Main Methods:

• Propose an anycost network quantization method utilizing a hyper-network.
• Dynamically search for optimal bitwidths per convolution patch based on feature maps and resource constraints.
• Employ active patch-wise bitwidth sampling and adaptive gradient ensembling during training for improved convergence and generalization.

Main Results:

• The proposed hyper-network efficiently adapts to different resource budgets with minimal finetuning.
• Achieved comparable efficiency-accuracy trade-offs compared to existing quantization methods.
• Significantly reduced the cost of adapting models to new resource budgets.

Conclusions:

• The anycost quantization method offers a flexible and cost-effective solution for deploying image SR networks on devices with variable resource constraints.
• This approach enhances the adaptability and generalization of quantized networks.
• It provides a practical pathway for efficient image super-resolution in diverse mobile computing environments.