Very Low Resolution Image Classification

The problem of classifying low resolution images is both highly challenging as low resolution images carry very limited information and incredibly important for developing robust, resolution-agnostic classification systems. In the past researchers have attempted to solve the low resolution image classification problem directly, that is, by training classifiers on a dataset of low-res images (e.g., CIFAR-10), or by employing a simple image resizing algorithm such as bicubic interpolation. In this work, we take a different approach and explore the use of a super-resolution network to transform input LR images into the high resolution image domain before classification. We show that using the above technique, we are able to improve low resolution image classification accuracy.