Politecnico di Torino (logo)

Implicit Neural Representations for Image Compression

Francesco Maria Chiarlo

Implicit Neural Representations for Image Compression.

Rel. Enrico Magli, Diego Valsesia. Politecnico di Torino, Corso di laurea magistrale in Ingegneria Informatica (Computer Engineering), 2021

PDF (Tesi_di_laurea) - Tesi
Licenza: Creative Commons Attribution Non-commercial No Derivatives.

Download (1MB) | Preview

The main focuse of the current thesis proposal is related to performing image compression via Implicit Neural Representations based deep neural network architectures, specifically employing SIREN architectures, adopting pruning and quantization techniques for constraining in the number of weights parameters and their numerical representation, respectively, while measuring and evaluating induced performance image quality metrics to be related to achieved image compression bit rate calculated from deep neural network overall footprint. The results obtained while running and collecting data related to the involved image quality metrices, such as Pnsr score and bit rate, have suggested us as well as provided us evidence of how critical and still heavy task represents the attempt of reaching image compression throughout common Neural Network Compression Techniques applied directly to Neural Network Models. Finally, due to the vast number of potential suitable hyper-parameter configurations, we have noticed that there are chance to reduce the gap we can measure, in terms of performance, between well-established image compression methods such as Jpeg and SIREN compressed models.

Relators: Enrico Magli, Diego Valsesia
Academic year: 2020/21
Publication type: Electronic
Number of Pages: 83
Corso di laurea: Corso di laurea magistrale in Ingegneria Informatica (Computer Engineering)
Classe di laurea: New organization > Master science > LM-32 - COMPUTER SYSTEMS ENGINEERING
Aziende collaboratrici: UNSPECIFIED
URI: http://webthesis.biblio.polito.it/id/eprint/19102
Modify record (reserved for operators) Modify record (reserved for operators)