相较于传统的图像压缩技术,深度图像压缩可以提供更优的率失真性能,甚至可以超越最新的压缩编码标准多功能视频编码(Versatile Video Coding,VVC).然而,随着网络复杂度的提升,深度图像压缩技术的提升亦有瓶颈.因此,提出了非对称离散高斯分布的深度图像压缩方法.并非优化编解码器或是熵模型,该方法在隐空间借助语义信息和稀疏过程,实现单高斯分布向非对称高斯分布的迁移,以节约码流.相较其他方法,所提方法具有更优的率失真性能,在Kodak数据集上解码的图像更加真实自然.
Compared with conventional image compression techniques,learned image compression can provide better rate-distortion performance,even surpassing the latest compression coding standard VVC.However,with the increase of network complexity,the improvement of learned image compression technology also has a bottleneck.Therefore,a learned image compression method with asymmetric discretized Gaussian likelihood is proposed.Instead of optimizing the codec or entropy model,this method uses semantic information and sparse process to realize the migration of single Gaussian distribution to asymmetric Gaussian distribution in hidden space to save the bit stream.Compared with other methods,the proposed method has better rate-distortion performance and the images decoded on Kodak dataset are more realistic and natural.