Learning to Improve Image Compression without Changing the Standard Decoder

Abstract

In recent years we have witnessed an increasing interest in applying Deep Neural Networks (DNNs) to improve the rate-distortion performance in image compression. However, the existing approaches either train a post-processing DNN on the decoder side, or propose learning for image compression in an end-to-end manner. This way, the trained DNNs are required in the decoder, leading to the incompatibility to the standard image decoders (e.g., JPEG) in personal computers and mobiles. Therefore, we propose learning to improve the encoding performance with the standard decoder. In this paper, We work on JPEG as an example. Specifically, a frequency-domain pre-editing method is proposed to optimize the distribution of DCT coefficients, aiming at facilitating the JPEG compression. Moreover, we propose learning the JPEG quantization table jointly with the pre-editing network. Most importantly, we do not modify the JPEG decoder and therefore our approach is applicable when viewing images with the widely used standard JPEG decoder. The experiments validate that our approach successfully improves the rate-distortion performance of JPEG in terms of various quality metrics, such as PSNR, MS-SSIM and LPIPS. Visually, this translates to better overall color retention especially when strong compression is applied. The codes are available at https://github.com/ YannickStruempler/LearnedJPEG.

Neural JPEG: End-to-End Image Compression Leveraging a Standard JPEG Encoder-Decoder

“…A more straightforward compression approach considered redundancy at the decoder side of the system and attempted to decompress by designing an iterative hybrid recurrent decoder [6,7,8]. Similarly, a standard encoder can be replaced with another DNN to enhance the model's internal neural representations and decode information while only using a standard decoder both in the pixel [24] and frequency domains [10].…”

“…Editing with Sparse RNNs: Inspired by [26] and [10], we design an approach that uses a neural models to either pre-edit (or iteratively process) an image I before the quantization step of JPEG or post-edit the inverse DCT coefficients before converting back to the reconstructed image Î. Specifically, for pre-editing, our neural encoder E Θ (I) proceeds according to the following steps to produce a set of "edit" weights:…”

“…We use the rate loss formulation proposed by [10] by replace attention map with sparse map obtained from SMRNN that is represented as follows:…”

“…Recent approaches that craft hybrid decoders [6,7,8] have presented a very promising direction, yet they struggle to operate well at the lowest bit rate, given that the quantized signals they must work with to reconstruct the original input signal are extremely sparse. Another promising alternative is to design a hybrid encoder that enhances encoder signals resulting in better compression even at the lowest bit rates [9,10]. However, these methods fail to remove artifacts at the decoder end, thus compromising compression quality in various situations.…”

An Empirical Analysis of Recurrent Learning Algorithms In Neural Lossy Image Compression Systems

“…A simpler approach to the compression approach takes into account redundancy at the decoder side of the system and attempts to iteratively decompress using a hybrid recurrent decoder [1,2]. Similarly, a standard encoder can be replaced with another DNN to enhance model representation and decode information using a standard decoder [20]. Prior efforts have shown the limitations of backprop based approaches on standard computer vision tasks [21,22,23] as well as natural language processing benchmarks [24].…”