LMDC: Learning a multiple description codec for deep learning-based image compression

Lijun Zhao

doi:10.1007/s11042-022-12216-4

Return

LMDC: Learning a multiple description codec for deep learning-based image compression

2022-02-25

PRE

DOI：10.1007/s11042-022-12216-4

Original

Save

Abstract

En 中文

Although deep learning technique has been widely leveraged to compress image, few attentions are paid to multiple description (MD) coding based on this technique. Meanwhile, packet loss and bit error may occur inevitably during transmission over unreliable networks, which seriously corrupts image quality. Thus, we introduce a deep Multiple Description Coding (MDC) framework for low bit-rate image compression, which is optimized according to deep MD rate-distortion minimization. Our framework includes MD multi-scale encoder network, MD cascaded-Resblock decoder networks, and arithmetic codec as well as a pair of learnable scalar quantizers and conditional probability models. In encoder, a pair of MD tensors are firstly generated by MD multi-scale encoder network, after which these tensors are discretized by a pair of learnable scalar quantizers respectively. Then, these quantized tensors are fed into conditional probability models for soft entropy estimation to regularize the learning of our encoder and decoder network during training, while these quantized tensors are lossly coded by arithmetic coding during testing. In decoder, these quantized tensors are decoded from bit-streams and then they are decompressed by MD cascaded-Resblock decoder networks for image decoding. To greatly reduce the total amount of network parameters, we design an auto-encoder network composed of these encoder and decoder networks with a symmetrical parameter sharing structure. Additionally, considering image spatial variation, each scalar quantizer is accompanied by an importance-indicator map to generate MD tensors, rather than using direct quantization. Experimental results have demonstrated that our MDC framework performs better than several existing MDC approaches in terms of several image quality metrics.

Keywords:

Multiple description coding

Scalar quantizer

Auto-encoder

Symmetrical sharing