Single Image Super-Resolution Using Deep Ensemble Learning

Abstract

The advancement in the digital world has made it possible to take high-quality images with even cellphones, but the size of the captured photos is a challenging issue since transferring and storing them are costly. Compression algorithms are widely used in industry and aerial imagery since some details are lost while compressing the images. These lost details are needed in some cases and applications. To address this issue, this paper proposes an ensemble deep learning method that shares information between different deep learning models to recover the original image. These models are employed to recover high-quality versions of images, known as super-resolution. We benefited from the Single Image technique and used one image as the input of the deep networks. To perform the ensemble learning, a CNN was used to integrate the outputs of different GAN models (LapSRN, SRResNet, RESNeXt, SRCNN, and ESCPN) to generate the high-resolution image. The proposed method could achieve intriguing results to increase the recovery quality compared to the rival methods in this realm.