Kuntoro Adi Nugroho, Yudi Eko Windarto, Cries Avian
Traditional denoising methods often struggle with computational intensity and model size, limiting their practical application. This paper introduces a novel approach to ternary quantization aimed at improving the efficiency and effectiveness of deep learning models for image denoising. Effectively ternarizing neural network weights while maintaining performance remains a challenge. Current methods present varied designs, often resulting in compatibility issues with different tasks. To address these issues, we propose a simplified approach that sets symmetric ternary quantization values equal to a threshold optimized by linear search, maintaining this value throughout training and inference. Our experimental results demonstrate the efficacy of the proposed method compared to established benchmarks and datasets, achieving PSNR values of 27.14 dB, 29.81 dB, and 26.36 dB in synthetic noise experiment and 37.88 dB in real-world noise experiment, which are closest to the full-precision model among the benchmarks. We ensure a fair comparison by using consistent training procedures and hyperparameters across all quantization techniques. The results indicate that our approach not only achieves competitive performance in image denoising tasks but also produces balanced quantized value distributions. These findings suggest the potential of the proposed method for ternary weight quantization in image denoising. © The Author(s), under exclusive licence to Springer-Verlag London Ltd., part of Springer Nature 2025.
Department of Computer Engineering, Diponegoro University, Semarang, Indonesia; Department of Electrical Engineering, Brawijaya University, Malang, Indonesia
