基于增强型多尺度残差生成对抗网络的图像压缩

Image compression based on enhanced multi-scale residual generation adversarial network

为解决低码率下更符合人类视觉感知的图像压缩,提出一种基于增强型多尺度残差生成对抗网络的有损压缩方法。在网络框架的自动编码器中,使用一种结构上改进的增强型多尺度残差块,其可以扩大感受野,更容易获得图像的全局信息。引入简易注意力模块,帮助网络更加关注图像复杂的部分,减少简单部分的比特。判别器部分采用全新的相对平均判别器,在网络框架中使用LPIPS(learned perceptual image patch similarity)感知损失减轻图像伪影问题。采用两阶段训练的方式解决引入生成对抗网络导致训练不稳定的问题。实验结果表明了在低码率下所提模型的有效性,与之前的工作相比,所提方法在感知失真指标上表现更优,性能提升了65%左右,重建图像更符合人类视觉感知。

To solve the problem of image compression that is more suitable for human visual perception at low bit rate,a lossy compression method based on enhanced multi-scale residual generation adversarial network was proposed.In the automatic encoder of the network framework,an enhanced multi-scale residual block with improved structure was proposed,which expanded the receptive field and was more easily to obtain the global information of the image.A simple attention module was introduced to help the network pay more attention to the complex part of the image and reduce the bit of the simple part.A new relativistic average discriminator was used in the discriminator part and LPIPS(learned perceptual image patch similarity)perception loss was used in the network framework to reduce the image artifact problem.The two-stage training method was adopted to solve the problem of unstable training caused by the introduction of the generated countermeasure network.Experimental results show that the proposed model is effective at low bit rate.Compared with previous work,the proposed method performs better on the perceptual distortion index,with a performance improvement of about 65%,and the reconstructed image is more consistent with human visual perception.