ScholarWorks@동국대학교

초록

Reconstructing high-resolution images from lowresolution inputs is one of the most challenging tasks in computer vision. While deep learning-based super-resolution (SR) models achieve high-quality results, their deep architectures and heavy computations hinder real-time deployment on mobile and edge devices. Among various model compression techniques, quantization is a promising solution that reduces computational cost and memory usage of SR models by representing weights and activations with low-bit integers, making it suitable for resourceconstrained SR applications. However, applying a uniform bitwidth across all layers of the target SR model can cause severe performance drop in quantization-sensitive layers, degrading image quality. To address this, we propose a mixed-precision quantization (MPQ) method that allocates variable bit-widths to each layer depending on quantization sensitivity estimated via mean squared error (MSE). Using a pretrained 32-bit floatingpoint (FP32) original model and calibration data, our method measures the MSE between original and quantized activations to assign more bits to sensitive layers and fewer to less sensitive ones, with the goal of minimizing overall quantization error and mitigating performance degradation. Bit-widths are further refined through lightweight fine-tuning. Unlike complex sensitivity estimation methods, our approach is simple, effective and relies only on a small amount of calibration data. Experimental results show that our method achieves superior peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) compared to other methods, demonstrating its effectiveness in enhancing quantization performance for SR tasks. © 2025 IEEE.

키워드