# End-to-end Optimized Image Compression

**Authors:** Johannes Ballé, Valero Laparra, Eero P. Simoncelli  
**arXiv:** [1611.01704](https://arxiv.org/abs/1611.01704)  
**Date:** 2016-11-05  
**Summary page:** https://www.summarizepaper.com/en/paper/1611.01704v1/

## Key points

- The paper presents a novel image compression system utilizing nonlinear encoding and decoding transformations, uniform quantizer, and deep neural network architectures.
- The system incorporates local gain control inspired by biological neurons using convolutional linear filters and nonlinear activation functions.
- Authors optimize the system for rate-distortion performance across a training image database using stochastic gradient descent.
- They introduce a continuous proxy to address the discontinuous loss function from the quantizer, similar to variational autoencoders.
- The optimized coder outperforms standard compression systems like JPEG and JPEG 2000 in terms of rate-distortion performance and visual quality for compressed images when tested on an independent image database.
- The paper is currently under review as a conference paper at ICLR 2017, indicating its potential impact on advancing image compression technology.

## Summary

The paper "End-to-end Optimized Image Compression" by Johannes Ballé, Valero Laparra, and Eero P. Simoncelli presents a novel image compression system that utilizes a nonlinear encoding transformation, uniform quantizer, and nonlinear decoding transformation. This system incorporates deep neural network architectures with convolutional linear filters and nonlinear activation functions to implement local gain control inspired by biological neurons. The authors jointly optimize the system for rate-distortion performance across a database of training images using stochastic gradient descent. They also introduce a continuous proxy to address the discontinuous loss function resulting from the quantizer, similar to variational autoencoders. Unlike generative models that focus on minimizing entropy under the model, this system must operate at any point along the rate-distortion curve. Testing on an independent database of images shows that this optimized coder outperforms standard compression systems like JPEG and JPEG 2000 in terms of rate-distortion performance and visual quality for compressed images. The paper is currently under review as a conference paper at ICLR 2017, highlighting its potential impact on advancing image compression technology.