Image Resolution Enhancement Using Convolutional Autoencoders with Skip Connections

Abstract

Improving image resolution, restoring images, denoising images has been a topic of wide study in deep learning domain. Due to the lack of ground truth images in practical scenarios, the enhanced images help tremendously in understanding and studying the phenomenon in an effective and efficient way. The paper presented here uses autoencoders which in turn comprise of encoder and decoder parts. In order to improve performance of autoencoder, skip connections from initial layers of encoder to the final layers of decoder have also been used. The deconvolutional part can be understood as combination of upsampling layers and convolutional layers. The proposed technique achieves impressive performance on a dataset (WHU-RS19) that has images of different geographies and are highly unrelated. The method proposed in this paper that uses symmetric convolutional and deconvolution layers, is able to achieve an accuracy of 89%; showing the merit of proposed network.