# Understanding SSIM

**Authors:** Jim Nilsson, Tomas Akenine-Möller  
**arXiv:** [2006.13846](https://arxiv.org/abs/2006.13846)  
**Date:** 2020-06-24  
**Summary page:** https://www.summarizepaper.com/en/paper/2006.13846v2/

## Key points

- The structural similarity index (SSIM) has been widely used for image quality measurement in various research fields.
- Recent research has revealed unexpected and nonintuitive properties of SSIM that require further investigation.
- Relying solely on SSIM to assess image quality can lead to incorrect conclusions.
- Using SSIM as a loss function for deep learning models may misguide neural network training and result in suboptimal outcomes.
- This study highlights the limitations and potential pitfalls of using SSIM as an assessment tool or loss function.
- Researchers and practitioners should exercise caution when relying solely on SSIM for image quality evaluation or incorporating it into deep learning frameworks.
- It is important to critically examine established metrics like SSIM and explore alternative approaches for accurate image quality assessment and effective deep learning model training.

## Summary

The structural similarity index (SSIM) has been widely used for almost two decades as a measure of image quality in various research fields. Its effectiveness and value in the research community have been widely acknowledged. However, despite its widespread use, there has been a lack of in-depth analysis and scrutiny of this index. Contrary to popular belief, recent research has revealed some intriguing properties of SSIM that warrant further investigation. In a paper titled "Understanding SSIM," authors Jim Nilsson and Tomas Akenine-Möller delve into the mathematical factors underlying SSIM and uncover unexpected and nonintuitive results that can arise when using this index in both synthetic and realistic scenarios. The authors highlight that relying solely on SSIM to assess image quality can lead to incorrect conclusions. They argue that the use of SSIM as a loss function for deep learning models may also misguide neural network training, potentially leading to suboptimal results. This study sheds light on the limitations and potential pitfalls associated with using SSIM as an assessment tool or loss function. By bringing attention to these issues, it urges researchers and practitioners to exercise caution when relying solely on SSIM for image quality evaluation or incorporating it into deep learning frameworks. Overall, this paper emphasizes the importance of critically examining established metrics like SSIM and encourages further exploration into alternative approaches for accurate image quality assessment and effective deep learning model training.