Attention-Based SRGAN for Super Resolution of Satellite Images

Single image super resolution plays a vital role in satellite image processing as the observed satellite image generally has low resolution due to the bottleneck in imaging sensor equipment and the communication bandwidth. Deep learning provides a better solution to improve its resolution compared to many sophisticated algorithms; hence, a deep attention-based SRGAN network is proposed. The GAN network consists of an attention-based SR generator to hallucinate the missing fine texture detail, a discriminator to guess how realistic is the generated image. The SR generator consists of a feature reconstruction network and attention mechanism. Feature reconstruction network consists of residually connected RDB blocks to reconstruct HR feature. The attention mechanism acts as a feature selector to enhance high-frequency details and suppress undesirable components in uniform region. The reconstructed HR feature and enhanced high-frequency information are fused together for better visual perception. The experiment is conducted on WorldView-2 satellite data using Googles free cloud computing GPU, Google colab. The proposed deep network performs better than the other conventional methods.