Presentation Information
[A-15-07]Perceptual Image Denoising with a Lightweight Gated CNN
〇Takamichi Miyata1 (1. Chiba Institute of Technology)
Keywords:
Image denoising,perceptual quality,LPIPS,DISTS
Mean squared error is widely used as a training loss for CNN-based image denoising; however, it often leads to over-smoothed restored images and loss of fine textures. To address this issue, perceptual image denoising methods using a weighted sum of perceptual image quality metrics, such as LPIPS and DISTS, have been proposed. These methods, however, typically employ U-Net as the denoising network, resulting in a large model size and high computational cost. In this paper, we propose a perceptual image denoising method that applies the same perceptual loss function to GTCNN, a lightweight gated convolutional neural network. By exploiting the GTCNN architecture, which separates denoising and contextual feature extraction through a gating mechanism, the proposed method aims to generate perceptually natural restored images with a small number of parameters and low computational cost. Experiments using DIV2K for training and CBSD6 for evaluation demonstrate that, under a high noise level, the proposed method substantially reduces the number of parameters and computational complexity compared with the U-Net-based method, while achieving the best DISTS score. Qualitative results also show that the proposed method produces restored images with more natural textures.