MFDN: Multiception Feature Distillation Network

Image super-resolution refers to obtaining high-resolution images with refined details and enhanced visual quality from low-resolution inputs with coarse details. Because of its wide range of applications, image super-resolution has attracted the attention of researchers in the image processing and computer vision community. Even though CNN and GAN-based approaches are quite successful, it is unfeasible to use them in edge devices due to their heavy computational complexity. This work is our attempt in this direction to develop a lightweight image super-resolution model which uses multiception convolution layers for feature distillation. Our inspiration in this work was Residual Feature Distillation Network (RFDN), where we substantially decreased the parameters while maintaining the PSNR metric. The proposed approach was able to achieve image super-resolution with enhanced visual quality as compared to the baseline approach.

I. Introduction

Image super-resolution (SR) refers to the task of upscaling a given low-resolution (LR) image with coarse details to a high-resolution (HR) image with refined details and enhanced visual quality. Image super-resolution is often associated to other terminologies such as upsampling, interpolation, image scaling, zooming and enlargement. It has been proved that upsampling an image via super-resolution methods can largely refine the amount of available information and thus lead to accurate and robust vision-based machine learning systems [1]. As a result, super-resolution methods have achieved cross domain acceptability and enjoy a wide range of applications such as medical imaging, surveillance and security, aerial imaging, compressed image/video enhancement, action recog-nition, remote sensing, astronomical images, forensics, pose estimation, fingerprint and gait recognition and many more. Apart from improving the perceptual quality, it also helps in other deep learning based computer vision tasks such as object detection, image segmentation.