Performance analysis of Satellite Image Super Resolution using Deep Learning Techniques

Super-resolution has gained significant importance recently owing to its finer sampling image details. Deep learning algorithms have remarked as an entity for developing single image high-quality reconstruction. Super-resolution with Deep Learning algorithms has demonstrated state of the art approaches for reconstructing sharper and more accurate images. Satellite images are highly prone to lose minute details of the image when subjected to algorithmic modeling. Thus, it is necessary to preserve the details of image. In this paper, an attempt is made to incorporate the state of the art approaches for reconstructing the satellite images. This requires careful conditioning of validating parameters like bias value, weights, appropriate usage of filters and scaling factors. The existing super-resolution algorithms such as Bicubic interpolation, Super resolution convolutional neural network (SRCNN), fast Super resolution convolutional neural network (FSRCNN) and Deep Laplacian Pyramid (LapSRN) are simulated to reconstruct the satellite images obtained from benchmark data sets of Indian and International satellite sensors. An extensive quantitative and qualitative evaluation of the super-resolution algorithms shows that the Deep Laplacian Pyramid networks perform favorably against the other state-of-the-art methods exclusively for satellite images.