CS297 Proposal
Image compression using neural networks
Kunal Deshmukh (kunal.deshmukh@sjsu.edu)
Advisor: Dr. Chris Pollett
Description:
Image compression is one of the most studied problems in computer science. Traditionally, lossless and lossy image compression algorithms such as
Run-length encoding or Transform coding is used for this purpose. Neural nets are used widely for image classification and detection problems.
When a neural network is trained on images, some information about those images gets stored in the neural network in the form of network parameters such as weights
while some information is passed on to next layers in a convoluted representation. This project aims to reconstruct an image using this representation and model parameters.
A new architecture for image compression needs to be developed for this purpose. Further, the quality of an image thus obtained, can be enhanced by using Generative Adversarial Network (GANs).
Schedule:
Week 1:
Aug 28 - Sept4 | Kickoff meeting with Dr. Pollett |
Week 2:
Sept 5 - Sept 11 | Read end to end image compression using neural network [3] . |
Week 3:
Sept 12 - Sept 18 | Deliverable 1 : Implement end to end image compression[3] architecture in tensorflow |
Week 4:
Sept 19 - Sept 25 | Explore convoluted representation and perform experiments on Deliverable 1 |
Week 5:
Sept 26 - Oct 2 | Work on Deliverable 2 |
Week 6:
Oct 3 - Oct 9 | Deliverable 2 : Implement Single Image Super-Resolution(SISR) using GAN[4]. |
Week 7:
Oct 10 - Oct 16 | Explore denoising and sparse autoencoders |
Week 7:
Oct 10 - Oct 16 | Study popular image file formats |
Week 8:
Oct 17 - Oct 23 | Read paper: Real-Time Adaptive Image Compression |
Week 9:
Oct 24 - Oct 30 | Explore Generative adversarial network |
Week 10:
Oct 31 - Nov 6 | Deliverable 3 : Modify Deliverable - 1 to run on ImageNet dataset. Find SSIM, PSNR matrices. |
Week 11:
Nov 7 - Nov 13 | Learn how to measure a loss in image. Objective and Subjective methods to evaluate Image quality. |
Week 12:
Nov 14 - Nov 20 | Work on Deliverable 4 |
Week 13:
Nov 21 - Nov 27 | Deliverable 4 : |
Week 14:
Nov 28 - Dec 4 | Implement "Full resolution image compression using Recurrent Neural Networks" [2] |
Week 15:
Dec 5 - Dec 11 | Work on Deliverable 5 |
Week 16:
Dec 12 - Dec 18 | Deliverable 5 : Complete CS 297 final report |
Deliverables:
The full project will be done when CS298 is completed. The following will
be done by the end of CS297:
- Deliverable 1 : Implement end to end image compression[4] architecture in tensorflow.
- Deliverable 2: Implement SISR using GAN.
- Write a code in Keras for GAN to classify zener card images.
- Find suitable pre-trained model from model zoo and train to identify car in imagenet dataset.
- Complete CS 297 final report.
References:
[1] Rippel Oren, Bourdev Lubomir (May 2017). Real-Time Adaptive Image Compression. Published at ICML 2017.Sydney, Austrelia. arxiv e-archive.
[2] Toderici George, et al. (July 2017). Full Resolution Image Compression with Recurrent Neural Networks. Google Inc. Google research
[3] Jiang Feng, et al. (August 2017). An End-to-End Compression Framework Based on Convolutional Neural Networks. IEEE Transactions on Circuits and Systems for Video Technology. Harbin, China. arxiv e-archive.
[4] Ledig Christian, et al. (May 2017) .Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network. Conference on Computer Vision and Pattern Recognition. Honolulu, Hawaii. arxiv e-archive
|