Abstract:

Convolutional Neural Networks are used extensively in Computer Vision. However, they are used seldom in Image Compression. Studies have shown Recurrent Neural Networks(RNNs) can be used as well for Image encoding and reconstruction. Autoencoder architecture is particularly useful for image compression and reconstruction. However, fully convolutional variational autoencoder fails to provide results superior to the architecture which employs RNNs as well.
Image compression Framework consists of three parts - a network to compress an image, encoder/binizer, and a network to reconstruct an image.
The objective of this project is to find an optimal way to achieve image compression using neural networks in terms of computation and/or image compression ratio and quality of an image reconstructed. To achieve this, I have planned to reconstruct these three parts by making use of architecture components already found in the literature.

CS297 Results

• Implemented image compression architecture using fully convolutional neural networks.
• Implemented single image superresolution using Generative Adversarial Networks.
• Wrote a script to compare image quality, performed expriments on Deliverable - 1.
• Implemented image compression using recurrent neural networks.

Proposed Schedule

Key Deliverables:

• Design
  • Propose a design of neural network based image compression framework based on CNNs and / or RNNs with three componants - Compression network, encoder/decoder block, Reconstruction network.
• Software
  • Python based application to find image similarity indices.
  • Proposed image compression framework implemented using PyTorch.
• Report
  • CS 298 report.
  • CS 298 presentation.

Innovations and Challenges

• Use of data augmentation and image rebuilding technique to rebuild original image.
• An encoding technique to achieve optimal compression of compressed representation.
• Provide a new optimal image compression architecture with at least comparable results as that of existing neural network based architectures.

References:

1. O. Rippel and L. Bourdev, "Real-Time Adaptive Image Compression," The 34th Int. Conf. on Mach. Learn., 2017. doi: arXiv:1705.05823v1.
2. G. Toderici et al., "Full Resolution Image Compression with Recurrent Neural Networks," arXiv e-prints.,2016. doi: arXiv:1608.05148.
3. W. Tao et al., "An End-to-End Compression Framework Based on Convolutional Neural Networks," Data Compression Conf. (DCC), Snowbird, UT, 2017, pp. 463-463. doi: 10.1109/DCC.2017.54.
4. C .Ledig, et al., "Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network.," Conf. on Comp. Vision and Pattern Recog. Vol. 2. No. 3, 2017. doi: arXiv:1609.04802.
5. J. Chung et al., "A recurrent latent variable model for sequential data," In Advances in neural inform. process. syst., pp. 2980-2988. 2015.
6. D. Minnen et al., "Joint autoregressive and hierarchical priors for learned image compression". Advances In Neural Inform. Process. Syst., 2018. doi: arXiv:1809.02736.
7. R. Pascanu et al., "How to construct deep recurrent neural networks", In Proc. of the Second Int. Conf. on Learning Representations (ICLR 2014)., 2014. doi:arXiv:1312.6026.
8. P. Cheng, "A New Single Image Super-Resolution Method Based on the Infinite Mixture Model", Access IEEE, vol. 5, pp. 2228-2240, 2017. doi: 10.1109.
9. Thanh Tran, Dat & Iosifidis, Alexandros & Gabbouj, Moncef. (2017). "Improving Efficiency in Convolutional Neural Network with Multilinear Filters. Neural Networks." 105. 10.1016/j.neunet.2018.05.017.