基于上下文聚类变换的端到端图像压缩方法
End-to-End Image Compression Method Based on Context Cluster Transform
罗英国 1陈芬 1韦玮 1张鹏 1彭宗举1
作者信息
- 1. 重庆理工大学电气与电子工程学院重庆 400054
- 折叠
摘要
针对基于卷积神经网络(Convolutional Neural Network,CNN)变换的端到端图像压缩方法存在图像局部相似特征交互不足的问题,本研究提出一种基于上下文聚类变换的端到端图像压缩方法.首先,上下文聚类的变换网络将图像转化为含有坐标的特征点,并将特征点分成几簇;然后,通过对每一簇内特征点进行聚合和再分配的方式学习图像特征;最后,引入量化器、超先验网络和基于空间-通道联合上下文的熵编码,以构建完整的端到端图像压缩模型.实验结果表明,与基于CNN变换的端到端图像压缩方法相比,所提方法在Kodak、CLIC测试集上BD-rate分别节省了2.75％、4.20％,并取得了不错的主观视觉效果.本研究方法实现了局部相似特征的交互,充分考虑了相邻像素间的相关性,从而获得较为满意的率失真性能.
Abstract
Aiming at the problem that the end-to-end image compression method based on convolutional neural network transform has insufficient interaction of local similar features,an end-to-end image compression method based on context cluster transform was proposed in this study.Firstly,the image was transformed into several feature points containing coordinates,and the feature points were divided into several clusters.Then,the image features were extracted by aggregating and redistributing the feature points within each cluster.Finally,quantization,hyerprprior network and entropy coding based on joint spatial-channel context were introduced to construct a complete end-to-end image compression model.The experimental results showed that compared with the end-to-end image compression method based on CNN transform,the BD-rate of the proposed method saved 2.75％ and 4.20％ on the Kodak and CLIC test datasets,respectively,and achieved good subjective visual effect.The method of this study realizes the interaction of local similar features,and fully considers the correlation between adjacent pixels,so as to obtain satisfactory rate-distortion performance.
关键词
深度学习/图像压缩/相关性/上下文聚类/变换网络Key words
Deep learning/Image compression/Correlation/Context cluster/Transform network引用本文复制引用
出版年
2024
NETL
NSTL
万方数据