# Suppressing Model Overfitting for Image Super-Resolution Networks

arXiv: https://arxiv.org/abs/1906.04809 · HTML: https://mlsanity.com/paper/1906.04809/

## Key facts

- Cited by 9 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 11 Jun 2019.

## Built on by (most cited first, 9 of 9)

- [Masked Image Training for Generalizable Deep Image Denoising](https://mlsanity.com/paper/2303.13132/) (2023)
- [Reflash Dropout in Image Super-Resolution](https://mlsanity.com/paper/2112.12089/) (2021)
- [RankSRGAN: Super Resolution Generative Adversarial Networks with Learning to Rank](https://mlsanity.com/paper/2107.09427/) (2021)
- [Restoring Spatially-Heterogeneous Distortions using Mixture of Experts Network](https://mlsanity.com/paper/2009.14563/) (2020)
- [Performance of a GPU- and Time-Efficient Pseudo 3D Network for Magnetic Resonance Image Super-Resolution and Motion Artifact Reduction](https://mlsanity.com/paper/2111.14259/) (2021)
- [Image Augmentation for Multitask Few-Shot Learning: Agricultural Domain Use-Case](https://mlsanity.com/paper/2102.12295/) (2021)
- [MaskBlur: Spatial and Angular Data Augmentation for Light Field Image Super-Resolution](https://mlsanity.com/paper/2410.06478/) (2024)
- [DeltaGAN: Towards Diverse Few-shot Image Generation with Sample-Specific Delta](https://mlsanity.com/paper/2207.10271/) (2022)
- [SRTGAN: Triplet Loss based Generative Adversarial Network for Real-World Super-Resolution](https://mlsanity.com/paper/2211.12180/) (2022)

## Abstract (arXiv)

Large deep networks have demonstrated competitive performance in single image super-resolution (SISR), with a huge volume of data involved. However, in real-world scenarios, due to the limited accessible training pairs, large models exhibit undesirable behaviors such as overfitting and memorization. To suppress model overfitting and further enjoy the merits of large model capacity, we thoroughly investigate generic approaches for supplying additional training data pairs. In particular, we introduce a simple learning principle MixUp to train networks on interpolations of sample pairs, which encourages networks to support linear behavior in-between training samples. In addition, we propose a data synthesis method with learned degradation, enabling models to use extra high-quality images with higher content diversity. This strategy proves to be successful in reducing biases of data. By combining these components -- MixUp and synthetic training data, large models can be trained without overfitting under very limited data samples and achieve satisfactory generalization performance. Our method won the second place in NTIRE2019 Real SR Challenge.