Image Super-Resolution by Augmentation of Region Information by Rapid Segmentation

Spatial Super-Resolution (SR) of images finds practical applications in multiple domains like resolution enhancement of old images, improvement of resolution where imaging systems are limited, like microscopy, surveillance and remote sensing. Out of the many techniques available, deep learning-based models have shown improvement performance when augmentation methods like geometric transformations and information-based fusion augmentation techniques are used. In the current work, the technique chosen is Single Image Super-Resolution (SISR) using segmentation information augmentation, a form of fusion technique on the Caltech University Birds 200 (CUB) dataset, which contains images of birds. The deep learning model, which was used for performing SR is based on EDSR technique. It was found that the best performing hyper-parameters were 64 filters with 8 residual blocks. Also, the performance of region augmented EDSR was around 0.5 dB better.