Image super-resolution based on convolution neural networks using multi-channel input

摘要

In this paper, we propose an image super-resolution (SR) method using multi-channel-input convolutional neural networks (MC-SRCNN) where the multi-channel input is comprised of an original low-resolution (LR) input and its edge-enhanced and variously interpolated version. Recently, Super-Resolution Convolutional Neural Network (SRCNN) showed remarkable performance. However, in SRCNN, deep layer structures make it difficult to learn the network parameters effectively due to vanishing gradient and exploding gradient problems. The proposed MC-SRCNN extends the SRCNN structure with multi-channel input that helps extract better features for restoration of high-resolution (HR) images. To constitute multi-channel input, we generate variously sharpened and interpolated versions of LR images. Such interpolated and sharpened LR images have complementary information for reconstructing the HR images. Therefore, MC-SRCNN could reconstruct the HR images without deep hidden layers. The experiment results showed that the MCCNN with 18 input channels outperformed the SRCNN with average 0.21, 0.34 and 0.18 dB PSNR gains for scale factors of 2, 3 and 4, respectively for Set5 dataset, and 0.18dB PSNR gain for a scale factor of 3 at Set14 dataset.