1 Bioinspired Compression Schemas 16/07/2009 Khaled MASMOUDI Pierre KORNPROBST INRIA Marc ANTONINII3S
2 Neuromath kickoff meeting How to use a retina model for efficient Video coding-decoding ?
3 The Thorpe retinal model Implement a simplistic model to generate spike trains. What data structures to use to represent them. Estimate the « quality » and « cost » of such a signal. Finally: Are those spike trains suitable for static image compression?
4 What we see Vs what we get
5 Quality metrics Is What we get at the end of the decoding similar to what we see before coding Different possibilities experimented for similarity measures: Peak SNR Weightened SNR Mean SSIM
6 Quality as a functional of spikes number
7 Cost metrics Use Information theory metrics as Shannon Entropy : Get a theorical Threshold Real encoded image file size What we do really get after Image decomposition Representation transform Lossless Coding
8 From bit per pixel to bit per spike
9 Towards video Classical video coding :Video is a series of frames Code first frame than use differential coding Difference is the Schalwijk distance between two possible rank ordered series
10 Next step Still some things to do with Thorpe Some technical improvement : Use parallelism as in the actual neural circuitry Consider continuity in Spike train generation: Use 2D+t filters Use « Virtual Retina » model to integrate more capabilities in the coder (Gain Control)
11 Epilogue New retinal model : With video coding efficiency as a design principle Get all of that on GPU
12 For the « Journal club » Brief summary of W. Bialek Work: ‘’Efficient representation as a design principle for neural coding and computation’’ (2007) Presentation of E. Simoncelli’s : ‘’Spike-triggered neural characterization’’(2006)