Image Quality Metrics: PSNR vs. SSIM

By Alain Horé, Djemel Ziou. 2010

4,996 citations recorded by the source corpus.

Subfields: Computer Vision and Pattern Recognition

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

Citations in the corpus

DISCO: Dynamic and Invariant Sensitive Channel Obfuscation for deep neural networks (2020, methodology): “For attacks on sensitive inputs, we measure privacy using `1 loss, SSIM and PSNR [62] between reconstructed and input image and top-1 accuracy on the private attribute for attacks on sensitive attributes.”
Fast, Accurate and Lightweight Super-Resolution with Neural Architecture Search (2019, methodology): “We choose PSNR and SSIM as metrics by convention [19].”
A Survey on Neural Speech Synthesis (2021, background): “For example, L1 is suitable for Laplace distribution, L2 is suitable for Gaussian distribution, mixture of Gaussian/Laplace/Logistic are suitable for multi-modal distribution, other losses such as SSIM [112] are for high-order statistics, and GAN based losses can be used for any distribution.”
Multi-Objective Reinforced Evolution in Mobile Neural Architecture Search (2019, methodology): “However, most of the studies concentrate on designing deeper and more complicated networks to boost PSNR (peak-signal-noise-ratio) and SSIM (structural similarity index), both of which are commonly used evaluation metrics (Wang et al., 2004; Hore & Ziou, 2010).”
E2Style: Improve the Efficiency and Effectiveness of StyleGAN Inversion (2021, methodology): “For the quantitative evaluation of the inversion, five metrics are adopted: Peak Signal-to-Noise Ratio (PSNR) [57], Structural SIMilarity (SSIM) [58], IDentity Similarity (IDS), runtime (using a single NVIDIA GEFORCE RTX 3090 GPU), and model size.”
StoryGAN: A Sequential Conditional GAN for Story Visualization (2018, methodology): “We also compare the Structural Similarity Index (SSIM) score between the generated images and ground truth [14].”
Deep Learning on Image Denoising: An overview (2019, neutral): “The quantitative evaluation mainly used peak-signal-to-noise-ratio (PSNR) [84] values of different denoisers to test the de-noising effects.”
Frido: Feature Pyramid Diffusion for Complex Scene Image Synthesis (2022, methodology): “• For the analysis of the multi-scale vector quantization process, we consider the classical Peak signal-to-noise ratio (PSNR) and structural index similarity (SSIM) (Wang et al. 2004) as the reconstruction quality assessment, which are common metrics for evaluating the similarity between images.”
Applications of Deep Learning and Reinforcement Learning to Biological Data (2017, methodology): “e-trained CNN with RF outperformed other methods (e.g., RF with handcrafted features and sequential forward selection with LDA) while analyzing MMM from 4SSIM = (PSNR 15 :0865)=20 069, with ˙fg = 102 [205]. 16/33 INBreast database (RF-CNNPT: 95%, CNNPT: 91%, RF-HF: 90%, SFS-LDA: 89%) [155]. In yet a”
CLNeRF: Continual Learning Meets NeRF (2023, methodology): “Specifically, after updating ✓ t , we iterate over all images in M t 1 and S t and keep the K images with the lowest rendering PSNR [11] from ✓ t .”
The Secret Revealer: Generative Model-Inversion Attacks Against Deep Neural Networks (2019, background): “PSNR is the ratio of an image’s maximum squared pixel fluctuation over the mean squared error between the target image and the reconstructed image [12].”
Recurrent Back-Projection Network for Video Super-Resolution (2019, background): “It has been suggested that PSNR is more sensitive to Gaussian noise, while SSIM is more sensitive to compression artifacts [12].”
Flow Models for Arbitrary Conditional Likelihoods (2019, background): “We note that PSNR is a metric that prefers blurry images over sample diversity (Hore and Ziou 2010).”
MagicAnimate: Temporally Consistent Human Image Animation using Diffusion Model (2023, neutral): “On the TED-talks dataset, we follow MRAA and report L1 error, average keypoint distance (AKD), missing keypoint rate (MKR), and average Euclidean distance (AED).”