Acceleration and higher precision by discrete wavelet transform for single image super-resolution using convolutional neural networks
Manh Tuan Nguyen, Keisuke Iwai, Takashi Matsubara, Takakazu Kurokawa · 2021
Recently, convolutional neural networks (CNNs) are good at providing hierarchical features and have succeeded for single image super-resolution (SISR). Most of them focus on designing a deeper and wider network to learn more discriminative high-level features; thereby, they take much time to process. There has been attention on using discrete wavelet transform (DWT) in SISR to solve this problem. For higher speed and higher performance, this paper proposes a new DWTL (DWT Layer) and IDWTL (Inverse DWTL), which generalized the combining DWT with existing CNNs of SISR. By putting existing CNNs between DWTL and IDWTL, new methods with new CNNs will be generated. This paper applied the proposed methods to four well-known CNNs of SISR, which are SRCNN, VDSR, EDSR, and RDN, and generated new SRWCNN, VDWSR, EDWSR, and RDWNSR. Evaluation results of these four new CNNs in generated image execution time and peak signal-to-noise ratio for the performance show that SRWCNN, VDWSR achieved about three times faster and higher performance. EDWSR, RDWNSR increased its speed about 3.6 times with equivalent performance.