In recent years, neural networks have been widely employed for data embedding across various fields. In this work, I present a novel approach for image deep representation where the key innovation is represented by the use of a GAN inversion network to embed and decode images. This method engages two main networks: a generator that serves as the decoder, and an inverse generator that acts as the encoder. A crucial aspect of this approach is the introduction of pseudo-randomness, specifically through a Linear Congruential Generator (LCG), to provide the latent codes to feed the generator. The key aspect of these sequences is that they can be reduced to a single value providing an extremely low bit-rate technique to embed images. This thesis wants to prove why LCG provides optimal pseudo-randomness results for this context through a comparative evaluation of randomness and sequences correlation against three other functions: Hénon map, Logistic map and Tent Map. This work will further detail the network architecture, training process, and optimization, alongside a comparative performance analysis on MNIST and Anime faces datasets, against one of the leading approaches in the field, Autoencoders.
Negli ultimi anni, le reti neurali sono state largamente impiegate per il data embedding in diversi campi. In questo lavoro presento un nuovo approccio all' image embedding basato sulla GAN inversion Questo approccio coinvolge l'utilizzo di due reti: una GAN e un inverse GAN rispettivamente con la funzione di decoder ed encoder. Un altro aspetto cruciale di questo approccio è l'utilizzo della pseudorandomicità, più nello specifico Linear Congruential Generator(LCG), il quale funge da generatore di un rumore invertibile da fornire in input alla GAN e il quale può essere rappresentato da un singolo valore. Nella mia tesi, andrò a esporre come questo network è stato pensato e implementato e a fornire un analisi comparativa, in termini di performance, utilizzando due dataset(MNIST e Anime faces) con uno una delle principali architetture nel medesimo contesto, gli Autoencoders.
Rappresentazione di immagine in spazi latenti pseudocasuali
MONCHIERI, LEONARDO
2023/2024
Abstract
In recent years, neural networks have been widely employed for data embedding across various fields. In this work, I present a novel approach for image deep representation where the key innovation is represented by the use of a GAN inversion network to embed and decode images. This method engages two main networks: a generator that serves as the decoder, and an inverse generator that acts as the encoder. A crucial aspect of this approach is the introduction of pseudo-randomness, specifically through a Linear Congruential Generator (LCG), to provide the latent codes to feed the generator. The key aspect of these sequences is that they can be reduced to a single value providing an extremely low bit-rate technique to embed images. This thesis wants to prove why LCG provides optimal pseudo-randomness results for this context through a comparative evaluation of randomness and sequences correlation against three other functions: Hénon map, Logistic map and Tent Map. This work will further detail the network architecture, training process, and optimization, alongside a comparative performance analysis on MNIST and Anime faces datasets, against one of the leading approaches in the field, Autoencoders.File | Dimensione | Formato | |
---|---|---|---|
Monchieri_Leonardo.pdf
accesso aperto
Dimensione
20.35 MB
Formato
Adobe PDF
|
20.35 MB | Adobe PDF | Visualizza/Apri |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/80207