基于深度学习的图像压缩是当前研究热点。在分级超先验结构的图像编码框架中，通过超先验层捕捉潜层部分的空间依赖性，实现了较好的编码效果。现有方案深度网络的训练损失采用函数率失真模型进行信源码率优化。在实际通信环境中，为减少数据传输过程错误，还需要信道编码。本文就分级超先验的图像框架在信道传输环境中的率失真优化问题展开研究。

    本文的工作：（1）分析了分级先验结构的层间异质信息的重要性差异。通过分析卷积网络层间的扩展过程，得到超先验层信道信息的错误对图像重建的影响程度更大，重要性更高。（2）提出了异质信息的不等错误（Unequal Error Protection，UEP）信道传输条件下的信道模型，设计了基于UEP的信道码率和失真的损失函数，从而实现不等错误保护信道中的率失真模型。（3）验证了此率失真模型在实际信道中图像压缩效果的有效性。本文在基于Pytorch的图像压缩研究库CompressAI的基础上，利用Kodak数据集进行测试。实验表明：本研究在相同的码率下能实现更好的图像压缩性能，其PSNR和MS-SSIM的提升率最高可达到13.3%和28.3%。

Deep learning-based image compression is a current research topic. The conventional rate-distortion model uses source code rates for optimization, while the source information will be encoded before being sent to different channel systems for transmission. Since there are channel errors in the process of data transmission, and the parameter errors generated by different layers have a differential impact on the information distortion, the optimization needs to be combined with the channel code rate. The hyperprior in the model is designed to capture spatial dependencies in the latent representation, and this paper investigates the optimization problem of rate-distortion in channel transmission for this image coding framework.

In this paper, we (1) analyzes the difference in the importance of heterogeneous information between layers of hierarchical prior structure. By analyzing the expansion process between layers of convolutional networks, it is obtained that errors in the channel information of the hyperprior have a greater degree of influence and importance on image reconstruction. (2) A channel model for Unequal Error Protection (UEP) channel transmission conditions with heterogeneous information is proposed, and UEP-based channel code rate-distortion loss functions are designed to realize the rate-distortion model. (3) The effectiveness of this rate-distortion model for image compression in real channels is verified. In this paper, tests are conducted using the Kodak dataset based on the Pytorch-based image compression research library CompressAI. Experiments show that this study can achieve better image compression performance at the same code rate, and the improvement of its PSNR and MS-SSIM can reach up to 27.2% and 57.0%.