Use of Attention Mechanism for Decoder in Deep Learning-based Image Super Resolution

One of major issues in training deep neural network for image super resolution is checkerboard artifact. It degrades the quality and resolution of output images. It usually appears at diagonal edges and curved edges of input images. It can also occur when too much information is compressed and lost during encoding process. To address this issues, we propose U-Net based structure with its decoder reinforced with attention modules for image super resolution task. U-Net structure is used to increase the feature reusability and reduce the loss of information during the encoding process. Attention modules are implemented in decoding layers in order to enhance the capability to produce high resolution output at diagonal edges and curved edges. Our proposed method shows improvements in PSNR and SSIM compared to the existing methods.