Want to create interactive content? It’s easy in Genially!
ICONS PRESENTATION
Adriano Lopez de onate
Created on November 28, 2023
Start designing with a free template
Discover more than 1500 professional designs like these:
Transcript
Generative modelling for image compression and reconstruction of X-RAY images
Adriano Lopez de Onate - BSc in Computer ScienceSupervised by Dr. Amir Atapour December the 5th, 2023
START
JUSTIFICATION AND BACKGROUND: SILICON VALLEY
THE PIED PIPER ALGORITHM
- The Pied Piper algorithm from the show Silicon Valley is a video compression software program written in C that achieves a Weissman score in the fives.
- This type of video compression technology is said to be able to actually shrink the internet by as much as 10 percent with widespread adoption.
jUSTIFICATION AND bACKGROUND: COMPRESSION - THE MAGIC KEY TO EVOLUTION
Napster 1999
HHDs 1980 - 1990
DVDs 1995
USBs 1996
Mobile device 2010
jUSTIFICATION AND bACKGROUND: whAT IS dATA COMPRESSION
DATA COMPRESSION
Data compression is a process of reducing the size of data files or streams by encoding information using fewer bits than the original representation.
LOSSLESS
TWO MAIN TYPES:
Originial
COMPRESSED
Originial
Originial
LOSSY
PERSONAL PROJECT: AIMS AND SCOPE
Introduction
ORIGINAL
INTERMEDIATE
FINAL OUTPUT
SIZE: 798KB
SIZE: 412KB
SIZE: 1.4Mb
- This project aims to merge the fields of image enhancement and data compression by developing an innovative approach using Generative Networks to achieve adaptive image compression.
PERSONAL PROJECT APPROACH: GAN VS DIFFUSION
VS
DIffusion-Based Models
Generative Adversarial Models
Diffusion-based models aim to strike a balance between compression and reconstruction fidelity through a diffusion process.
Generative Adversarial Network (GAN) involves training an encoder and decoder to compress and reconstruct the original images.
PERSONAL PROJECT: FINAL SCOPE
- Models apart, my project is therefore entirely based on finding the "perfect" matching point between compression and quality using a lossy apporach
COMPRESSION
PERCEPTUAL QUALITY
SUPERB!
PERSONAL PROJECT: work and further work!
1.
2.
3.
4.
Getting the best out of Diffusion and Adversarial models
Evaluate the current research papers
Train the model on medicals datatset
Try to beat the current state-of-art! (Perhaps a tad too ambitious)
"Entropy is the silent DJ at the data compression party. It spins the records of disorder, but our algorithms are here to dance with chaos and make sure the beats are still groovy."
- Richard Endricks, Silicon Valley [HBO, 2014]
Got an idea?
Use this space to add awesome interactivity. Include text, images, videos, tables, PDFs... even interactive questions! Premium tip: Get information on how your audience interacts with your creation:
- Visit the Analytics settings;
- Activate user tracking;
- Let the communication flow!
Got an idea?
Use this space to add awesome interactivity. Include text, images, videos, tables, PDFs... even interactive questions! Premium tip: Get information on how your audience interacts with your creation:
- Visit the Analytics settings;
- Activate user tracking;
- Let the communication flow!
Got an idea?
Use this space to add awesome interactivity. Include text, images, videos, tables, PDFs... even interactive questions! Premium tip: Get information on how your audience interacts with your creation:
- Visit the Analytics settings;
- Activate user tracking;
- Let the communication flow!
Got an idea?
Use this space to add awesome interactivity. Include text, images, videos, tables, PDFs... even interactive questions! Premium tip: Get information on how your audience interacts with your creation:
- Visit the Analytics settings;
- Activate user tracking;
- Let the communication flow!
Got an idea?
Use this space to add awesome interactivity. Include text, images, videos, tables, PDFs... even interactive questions! Premium tip: Get information on how your audience interacts with your creation:
- Visit the Analytics settings;
- Activate user tracking;
- Let the communication flow!
Example of Diffusion-based compression
Lorem ipsum dolor sit
Overview of the architecture. Given an input image x and target rate factor λrate, we obtain a base codec reconstruction x ̃. the DDPM is conditioned on x ̃ and learns to model a reverse diffusion process that generates residuals r0 from sampled gaussian noise latents rT . The enhanced reconstruction xˆ is then obtained by adding the predicted residual to x ̃
Example of GAN-based compression
The entire model is trained in the first stage, and the second stage trains only the decoder. The interpolated decoder reconstructs an input image from quantized latent code. Q, AE, and AD are a quantizer, an arithmetic encoder and an arithmetic decoder, respectively.