Learn Image Compression for Classification Task

Machine learning has been used in various areas with tremendous success in tasks such as classification or regression. Recent work suggests, that the key technique leading to this success, deep neural networks, can also be used for data compression. Google even initialized a challenge on image compression at this years CVPR 2018 [2].

In order to measure the quality of the compression learned, one has to define a metric. Typically this relies on a human visual system inspired measure. The question we ask here:

Can we train a neural network for encoding and decoding images, which does optimize for a given subsequent machine learning task?

We would start reproducing the result from the challenge [2]. Then this method can be adopted for classical image classification. The idea of learning compression can be extended by trying to reduce the number of bit per pixels (bbp) to a minimum, whilst allowing no, or only a fixed portion of loss in accuracy.

Supervised by Cedric Renggli

cedric.renggli@inf.ethz.ch

References

  • [1] Toderici, George, et al. "Full resolution image compression with recurrent neural networks." arXiv preprint (2016).
  • [2] URL: www.compression.cc