Shallow sparse autoencoders versus sparse coding algorithms for image compression

Thierry Dumas (1) , Aline Roumy (1) , Christine Guillemot (1)

1 Sirocco - Analysis representation, compression and communication of visual data

This paper considers the problem of image compression with shallow sparse autoencoders. We use both a T-sparse autoencoder (T-sparse AE) and a winner-take-all autoencoder (WTA AE). A performance analysis in terms of rate-distortion trade-off and complexity is conducted, comparing with LARS-Lasso, Coordinate Descent (CoD) and Orthogonal Matching Pursuit (OMP). We show that, WTA AE achieves the best rate-distortion trade-off, it is robust to quantization noise and it is less complex than LARS-Lasso, CoD and OMP.

Mots clés: image compression; sparse autoencoders; shallow architectures; complexity

Domaines: Traitement des images [eess.IV]; Apprentissage [cs.LG]

Thierry Dumas, Aline Roumy, Christine Guillemot. Shallow sparse autoencoders versus sparse coding algorithms for image compression. 2016 IEEE International Conference on Multimedia and Expo (ICME 2016) , Jul 2016, Seattle, WA, United States. pp.1 - 6, ⟨10.1109/ICMEW.2016.7574708⟩. ⟨hal-01377907⟩