A Fusion-Based Framework for Unsupervised Single Image Super-Resolution

Image super-resolution has been a continuously demanding topic in the computer-vision community in recent decades and has witnessed impressive applications in increasing spatial resolution in every field like medicine, agriculture, remote sensing, defense security, and many more applications. Further, deep learning-based image super resolution methods have shown tremendous improvement in reconstruction performance. However, most of the recent state-of-the-art deep learning-based methods for image super-resolution assume an ideal degradation by the bicubic kernel on standard dataset approaches and perform poorly on real-world satellite images in practice, as real degradations are far away and more complex in nature than pre-defined assumed kernels. Motivated by this real-time challenge, our idea is to enhance the 600 m spatial-resolution image, which is extremely low, and implicitly defines image-specific features in an iterative way without defining any fixed explicit degradation for image super-resolution. Besides, we also did a comparative study based on a No-Reference Image Quality Assessment. The evaluation is done both qualitatively (vision based) and quantitatively without recurring to a reference image for quality assessment. The proposed framework outperforms by incorporating domain knowledge from recently implemented unsupervised single-image blind super-resolution techniques.