# Image Super-Resolution Using VDSR-ResNeXt and SRCGAN

**Authors:** Saifuddin Hitawala, Yao Li, Xian Wang, Dongyang Yang  
**arXiv:** [1810.05731](https://arxiv.org/abs/1810.05731)  
**Date:** 2018-10-10  
**Summary page:** https://www.summarizepaper.com/en/paper/1810.05731v1/

## Key points

- Deep learning has advanced the quality and speed of high-resolution image generation in Super Resolution (SR) techniques.
- Generative adversarial networks (GAN) and very deep convolutional networks (VDSR) are promising algorithms in this field.
- VDSR-ResNeXt is a multi-branch convolutional network that combines VDSR and ResNeXt architectures to enhance HR image quality while maintaining computational efficiency.
- SRCGAN is a conditional GAN that incorporates class labels as input, allowing for more targeted and context-aware image super-resolution.
- Extensive experiments using common SR benchmark datasets were conducted to evaluate the performance of VDSR-ResNeXt and SRCGAN.
- Both quantitative and qualitative assessments showed significant improvements in image quality and computational speed compared to existing techniques.
- The combination of VDSR-ResNeXt and SRCGAN showcases the potential for achieving superior HR image generation with improved contextual understanding.
- These advancements have implications for applications such as medical imaging, remote sensing, and video processing.

## Summary

In the field of Super Resolution (SR) techniques, deep learning has played a significant role in advancing the quality and speed of high-resolution (HR) image generation. Two particularly promising algorithms are generative adversarial networks (GAN) and very deep convolutional networks (VDSR). Building on these advancements, this paper introduces two novel approaches: VDSR-ResNeXt and SRCGAN. <br>
VDSR-ResNeXt is a deep multi-branch convolutional network inspired by both VDSR and ResNeXt architectures. By combining the strengths of these models, VDSR-ResNeXt aims to further enhance HR image quality while maintaining computational efficiency. On the other hand, SRCGAN is a conditional GAN that incorporates class labels as input to the GAN model. This explicit inclusion of class labels allows for more targeted and context-aware image super-resolution. <br>
To evaluate the performance of these proposed methods, extensive experiments were conducted using common SR benchmark datasets. Both quantitative and qualitative assessments were performed to measure the effectiveness of VDSR-ResNeXt and SRCGAN in generating high-quality HR images. The results demonstrated significant improvements over existing techniques in terms of image quality and computational speed. Overall, this paper presents two innovative approaches for image super-resolution using deep learning techniques. The combination of VDSR-ResNeXt and SRCGAN showcases the potential for achieving superior HR image generation with improved contextual understanding. These advancements have implications for various applications such as medical imaging, remote sensing, and video processing where high-resolution imagery plays a crucial role.