I²NQ: Inter and Intra Nonuniform Quantization for Single Image Super-Resolution.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Neural network quantization is crucial for model compression, converting floating-point to integer representations.
• Existing quantization methods are optimized for general visual tasks, not specifically for image super-resolution (SR).
• SR models require preserving fine details and original feature distributions, which standard quantization can disrupt.

Purpose of the Study:

• To develop a quantization technique tailored for image super-resolution (SR) models.
• To address the challenges posed by nonuniform feature distributions and the need for preserving high-frequency details in SR.
• To improve the efficiency and performance of quantized SR models.

Main Methods:

• Proposed a novel Inter and Intra Nonuniform Quantization method.
• Introduced a flex-scale-weight-adjust (FSWA) technique to maintain weight diversity and minimize quantization errors.
• Evaluated the method on SR reconstruction tasks, comparing against existing quantization approaches.

Main Results:

• The proposed method effectively handles the nonuniform feature distribution characteristic of SR models.
• FSWA successfully preserved weight information diversity, leading to reduced quantization errors.
• Experimental results showed superior performance in both reconstruction metrics and visual quality compared to other methods.

Conclusions:

• The developed quantization strategy is well-suited for image super-resolution tasks.
• The proposed method offers significant improvements over standard quantization techniques for SR.
• This work advances model compression for specialized deep learning applications like SR.