While deep learning-based image compression methods have shown impressive coding performance, most existing methods are still in the mire of two limitations: (1) unpredictable compression efficiency gain when adopting convolutional neural networks with different depths, and (2) lack of an accurate model to estimate the entropy during the training process. To address these two problems, in this paper, a deep multi-stage representation based image compression (MSRIC) method is proposed. Owing to this architecture, the detail information of shallow stages and the compact information of deep stages can be utilized for image reconstruction. Furthermore, a data-dependent channel-wised factorized probability model (DCFPM) is adopted to increase the accuracy of entropy estimation. Experimental results indicate that the proposed method guarantees better perceptual performance at a wide range of bit-rates. Also, ablation studies are carried out to validate the above mentioned technologies.

虽然基于深度学习的图像压缩方法已经显示出令人印象深刻的编码性能，但大多数现有方法仍处于两个限制的泥潭中：（1）采用不同深度的卷积神经网络时无法预测的压缩效率增益，以及（2）缺乏准确的模型在训练过程中估计熵。为了解决这两个问题，本文提出了一种基于深度多级表示的图像压缩（MSRIC）方法。由于这种架构，浅阶段的细节信息和深阶段的紧凑信息可以用于图像重建。此外，采用数据相关的信道分解概率模型（DCFPM）来提高熵估计的准确性。实验结果表明，所提出的方法在广泛的比特率范围内保证了更好的感知性能。此外，还进行了消融研究以验证上述技术。