RUnet : U-Net 기반 각도 분해 광전자 방출 분광법 영상 디노이징

초록

This paper proposes an RUnet, specialized for Angle-Resolved Photoemission Spectroscopy (ARPES) data denoising. A U-Netbased denoising network exhibits excellent performance, however it is difficult to consistently maintain denoising performance for awide range of noise levels. An autoencoder network maintains consistent denoising performance for all ranges of noise levels, yetit performs lower denoising performance than the U-Net based method. In order to increase the denoising performance for allranges of noise levels, the proposed RUnet makes some categories such as high-, medium-, and low-range in accordance with thenoise intensity for training. In addition, by applying the squeeze-and-excitation technique at the end of convolution layers in thedecoder parts to efficiently gather the noise features, the proposed RUnet can significantly improve the denoising performance. Asa result, the proposed RUnet can extract noise characteristics of ARPES image data in greater detail, thus it achieves averageimprovements of 4.79 dB in PSNR and 0.0217 in MS-SSIM compared to the state-of-the-art methods.

키워드