Salient Object Super-resolution

Abstract

Salient object super-resolution refers to enhancing the resolution and details ofsalient objects or regions in an image. It is a sub-field of image super-resolution,which aims to generate high-resolution salient object images from low-resolutioninputs. Several approaches have been used for either salient object detection orimage super-resolution, but no method employs both in a single mechanism.The aim behind salient object super-resolution is to provide a more focused,informative, and visually pleasing representation of images by prioritizing andenhancing the most relevant and eye-catching regions. This can lead to improvedperformance in various applications like surveillance, medical imaging etc. and abetter viewing experience for users.We propose a salient object super-resolution approach that addresses the challengesinherent in this task, like fine details preservation, inconsistent saliencymap quality, computational complexity, ambiguity and uncertainty etc. This approachinvolves salient object detection, salient object segmentation, salient objectsuper-resolution, restacking of salient objects, and guided image smoothening.Each step is designed to improve salient objects� resolution and visual qualitywhile preserving the remaining image content.For the super-resolution task, we employed three different models, namelySRGAN (Super-Resolution Generative Adversarial Network) [26], NLSN (Non-Local Sparse Attention Network) [39], and DRT (Deraining Recursive Transformer)[32]. We used the "Salient Object Detection with Robust Background Detection"method [58] for saliency detection.Further, we explore the potential of a hybrid model that combines the DRT andNon-Local Sparse Attention techniques for the super-resolution task. The DRTmodel, initially designed for deraining tasks, is adapted for super-resolution torestore fine details and textures within the low-resolution image effectively. TheNon-Local Sparse Attention mechanism is incorporated to selectively attend torelevant spatial and channel information, improving the preservation of essentialfeatures while suppressing noise and artefacts.Overall, our work contributes to advancing salient object super-resolution techniques and explores the potential of a hybrid model, TraNLSN, for further improvements.Analyzing the results from the ongoing training phase will provideinsights into the effectiveness of the hybrid model and its potential applicationsin various domains.