xingling0 / image-quality-assessment

Commen Image Quality Assessment indexes: (FID, SIFID, CleanFID, LPIPS) and (Scoot, PSNR, SSIM, FSIM & MAE), implemented by Pytorch and Matlab.

Image-Quality-Assessment

FID

Pytorch and Python3 version of

• FID（Fréchet inception distance）[1]: smaller is better.
• SIFID（Single Image Fréchet Inception Distance）[2]: smaller is better.
• C-FID（Clean FID）[3]: smaller is better.
• LPIPS（Learned Perceptual Image Patch Similarity）[4]: smaller is better.

IQA

Matlab version of

• SCOOT（Structure Co-Occurrence Texture）[5]: bigger is better.
• PSNR（Peak Signal-to-Noise Ratio）: bigger is better.
• SSIM（Structural Similarity Index Measure）[6]: bigger is better.
• FSIM（Feature Similarity Index Measure）[7]: bigger is better.
• MAE（Mean Absolute Error）: bigger is better.

NLDA

Matlab version of

• NLDA（Null-space Linear Discriminant Analysis）[8]

---

[1] Heusel M, Ramsauer H, Unterthiner T, et al. Gans trained by a two time-scale update rule converge to a local nash equilibrium[J]. Advances in neural information processing systems, 2017, 30.

[2] Shaham T R, Dekel T, Michaeli T. Singan: Learning a generative model from a single natural image[C]. Proceedings of the IEEE/CVF International Conference on Computer Vision. 2019: 4570-4580.

[3] Parmar G, Zhang R, Zhu J Y. On Aliased Resizing and Surprising Subtleties in GAN Evaluation[J].

[4] Zhang R, Isola P, Efros A A, et al. The unreasonable effectiveness of deep features as a perceptual metric[C]. Proceedings of the IEEE conference on computer vision and pattern recognition. 2018: 586-595.

[5] Fan D P, Zhang S C, Wu Y H, et al. Scoot: A perceptual metric for facial sketches[C]. Proceedings of the IEEE/CVF International Conference on Computer Vision. 2019: 5612-5622.

[6] Wang Z, Bovik A C, Sheikh H R, et al. Image quality assessment: from error visibility to structural similarity[J]. IEEE transactions on image processing, 2004, 13(4): 600-612.

[7] Zhang L, Zhang L, Mou X, et al. FSIM: A feature similarity index for image quality assessment[J]. IEEE transactions on Image Processing, 2011, 20(8): 2378-2386.

[8] Chen L F, Liao H Y M, Ko M T, et al. A new LDA-based face recognition system which can solve the small sample size problem[J]. Pattern recognition, 2000, 33(10): 1713-1726.